Daily Papers

The critique capacity of Large Language Models (LLMs) is essential for reasoning abilities, which can provide necessary suggestions (e.g., detailed analysis and constructive feedback). Therefore, how to evaluate the critique capacity of LLMs has drawn great attention and several critique benchmarks have been proposed. However, existing critique benchmarks usually have the following limitations: (1). Focusing on diverse reasoning tasks in general domains and insufficient evaluation on code tasks (e.g., only covering code generation task), where the difficulty of queries is relatively easy (e.g., the code queries of CriticBench are from Humaneval and MBPP). (2). Lacking comprehensive evaluation from different dimensions. To address these limitations, we introduce a holistic code critique benchmark for LLMs called CodeCriticBench. Specifically, our CodeCriticBench includes two mainstream code tasks (i.e., code generation and code QA) with different difficulties. Besides, the evaluation protocols include basic critique evaluation and advanced critique evaluation for different characteristics, where fine-grained evaluation checklists are well-designed for advanced settings. Finally, we conduct extensive experimental results of existing LLMs, which show the effectiveness of CodeCriticBench.

Large Language Models (LLMs) often struggle with complex reasoning scenarios. While preference optimization methods enhance reasoning performance through training, they often lack transparency in why one reasoning outcome is preferred over another. Verbal reflection techniques improve explainability but are limited in LLMs' critique and refinement capacity. To address these challenges, we introduce a contrastive reflection synthesis pipeline that enhances the accuracy and depth of LLM-generated reflections. We further propose a dual-model reasoning framework within a verbal reinforcement learning paradigm, decoupling inference-time self-reflection into specialized, trained models for reasoning critique and refinement. Extensive experiments show that our framework outperforms traditional preference optimization methods across all evaluation metrics. Our findings also show that "two heads are better than one", demonstrating that a collaborative Reasoner-Critic model achieves superior reasoning performance and transparency, compared to single-model approaches.

Critique ability are crucial in the scalable oversight and self-improvement of Large Language Models (LLMs). While many recent studies explore the critique ability of LLMs to judge and refine flaws in generations, how to comprehensively and reliably measure the critique abilities of LLMs is under-explored. This paper introduces \shortname, a novel benchmark designed to comprehensively and reliably evaluate four key critique ability dimensions of LLMs: feedback, comparison, refinement and meta-feedback. \shortname~encompasses nine diverse tasks, each assessing the LLMs' ability to critique responses at varying levels of quality granularity. Our extensive evaluations of open-source and closed-source LLMs reveal intriguing relationships between the critique ability and tasks, response qualities, and model scales. Datasets, resources and evaluation toolkit for \shortname~will be publicly released at https://github.com/gmftbyGMFTBY/CriticBench.

Critique ability, a meta-cognitive capability of humans, presents significant challenges for LLMs to improve. Recent works primarily rely on supervised fine-tuning (SFT) using critiques generated by a single LLM like GPT-4. However, these model-generated critiques often exhibit flaws due to the inherent complexity of the critique. Consequently, fine-tuning LLMs on such flawed critiques typically limits the model's performance and propagates these flaws into the learned model. To overcome these challenges, this paper proposes a novel data generation pipeline, named MultiCritique, that improves the critique ability of LLMs by utilizing multi-agent feedback in both the SFT and reinforcement learning (RL) stages. First, our data generation pipeline aggregates high-quality critiques from multiple agents instead of a single model, with crucial information as input for simplifying the critique. Furthermore, our pipeline improves the preference accuracy of critique quality through multi-agent feedback, facilitating the effectiveness of RL in improving the critique ability of LLMs. Based on our proposed MultiCritique data generation pipeline, we construct the MultiCritiqueDataset for the SFT and RL fine-tuning stages. Extensive experimental results on two benchmarks demonstrate: 1) the superior quality of our constructed SFT dataset compared to existing critique datasets; 2) additional improvements to the critique ability of LLMs brought by the RL stage. Notably, our fine-tuned 7B model significantly surpasses other advanced 7B-13B open-source models, approaching the performance of advanced 70B LLMs and GPT-4. Codes, datasets and model weights will be publicly available.

Critical thinking is essential for rational decision-making and problem-solving. This skill hinges on the ability to provide precise and reasoned critiques and is a hallmark of human intelligence. In the era of large language models (LLMs), this study explores the ability of LLMs to deliver accurate critiques across various tasks. We are interested in this topic as a capable critic model could not only serve as a reliable evaluator, but also as a source of supervised signals for model tuning. Particularly, if a model can self-critique, it has the potential for autonomous self-improvement. To examine this, we introduce a unified evaluation framework for assessing the critique abilities of LLMs. We develop a benchmark called CriticBench, which comprises 3K high-quality natural language queries and corresponding model responses; and annotate the correctness of these responses. The benchmark cover tasks such as math problem-solving, code completion, and question answering. We evaluate multiple LLMs on the collected dataset and our analysis reveals several noteworthy insights: (1) Critique is generally challenging for most LLMs, and this capability often emerges only when models are sufficiently large. (2) In particular, self-critique is especially difficult. Even top-performing LLMs struggle to achieve satisfactory performance. (3) Models tend to have lower critique accuracy on problems where they are most uncertain. To this end, we introduce a simple yet effective baseline named self-check, which leverages self-critique to improve task performance for various models. We hope this study serves as an initial exploration into understanding the critique abilities of LLMs, and aims to inform future research, including the development of more proficient critic models and the application of critiques across diverse tasks.

Training large language models (LLMs) to spend more time thinking and reflection before responding is crucial for effectively solving complex reasoning tasks in fields such as science, coding, and mathematics. However, the effectiveness of mechanisms like self-reflection and self-correction depends on the model's capacity to accurately assess its own performance, which can be limited by factors such as initial accuracy, question difficulty, and the lack of external feedback. In this paper, we delve into a two-player paradigm that separates the roles of reasoning and critique models, where the critique model provides step-level feedback to supervise the reasoning (actor) model during both test-time and train-time. We first propose AutoMathCritique, an automated and scalable framework for collecting critique data, resulting in a dataset of 76,321 responses paired with step-level feedback. Fine-tuning language models with this dataset enables them to generate natural language feedback for mathematical reasoning. We demonstrate that the critique models consistently improve the actor's performance on difficult queries at test-time, especially when scaling up inference-time computation. Motivated by these findings, we introduce the critique-based supervision to the actor's self-training process, and propose a critique-in-the-loop self-improvement method. Experiments show that the method improves the actor's exploration efficiency and solution diversity, especially on challenging queries, leading to a stronger reasoning model. Lastly, we take the preliminary step to explore training self-talk reasoning models via critique supervision and showcase its potential. Our code and datasets are at https://mathcritique.github.io/{https://mathcritique.github.io/}.

Large language models (LLMs) have witnessed rapid advancements, demonstrating remarkable capabilities. However, a notable vulnerability persists: LLMs often uncritically accept flawed or contradictory premises, leading to inefficient reasoning and unreliable outputs. This emphasizes the significance of possessing the Premise Critique Ability for LLMs, defined as the capacity to proactively identify and articulate errors in input premises. Most existing studies assess LLMs' reasoning ability in ideal settings, largely ignoring their vulnerabilities when faced with flawed premises. Thus, we introduce the Premise Critique Bench (PCBench), designed by incorporating four error types across three difficulty levels, paired with multi-faceted evaluation metrics. We conducted systematic evaluations of 15 representative LLMs. Our findings reveal: (1) Most models rely heavily on explicit prompts to detect errors, with limited autonomous critique; (2) Premise critique ability depends on question difficulty and error type, with direct contradictions being easier to detect than complex or procedural errors; (3) Reasoning ability does not consistently correlate with the premise critique ability; (4) Flawed premises trigger overthinking in reasoning models, markedly lengthening responses due to repeated attempts at resolving conflicts. These insights underscore the urgent need to enhance LLMs' proactive evaluation of input validity, positioning premise critique as a foundational capability for developing reliable, human-centric systems. The code is available at https://github.com/MLGroupJLU/Premise_Critique.

Critiques are important for enhancing the performance of Large Language Models (LLMs), enabling both self-improvement and constructive feedback for others by identifying flaws and suggesting improvements. However, evaluating the critique capabilities of LLMs presents a significant challenge due to the open-ended nature of the task. In this work, we introduce a new benchmark designed to assess the critique capabilities of LLMs. Unlike existing benchmarks, which typically function in an open-loop fashion, our approach employs a closed-loop methodology that evaluates the quality of corrections generated from critiques. Moreover, the benchmark incorporates features such as self-critique, cross-critique, and iterative critique, which are crucial for distinguishing the abilities of advanced reasoning models from more classical ones. We implement this benchmark using eight challenging reasoning tasks. We have several interesting findings. First, despite demonstrating comparable performance in direct chain-of-thought generation, classical LLMs significantly lag behind the advanced reasoning-based model o1-mini across all critique scenarios. Second, in self-critique and iterative critique settings, classical LLMs may even underperform relative to their baseline capabilities. We hope that this benchmark will serve as a valuable resource to guide future advancements. The code and data are available at https://github.com/tangzhy/RealCritic.

Critique, as a natural language description for assessing the quality of model-generated content, has been proven to play an essential role in the training, evaluation, and refinement of Large Language Models (LLMs). However, there is a lack of principled understanding in evaluating the quality of the critique itself. In this paper, we pioneer the critique of critique, termed MetaCritique, which is a framework to evaluate the critique from two aspects, i.e., factuality as precision score and comprehensiveness as recall score. We calculate the harmonic mean of precision and recall as the overall rating called F1 score. To obtain a reliable evaluation outcome, we propose Atomic Information Units (AIUs), which describe the critique in a more fine-grained manner. MetaCritique takes each AIU into account and aggregates each AIU's judgment for the overall score. Moreover, given the evaluation process involves intricate reasoning, our MetaCritique provides a natural language rationale to support each judgment. We construct a meta-evaluation dataset containing 300 critiques (2653 AIUs) across four tasks (question answering, reasoning, entailment, and summarization), and we conduct a comparative study to demonstrate the feasibility and effectiveness. Experiments also show superior critique judged by MetaCritique leads to better refinement, indicating generative artificial intelligence indeed has the potential to be significantly advanced with our MetaCritique. We will release relevant code and meta-evaluation datasets at https://github.com/GAIR-NLP/MetaCritique.

Large language models (LLMs) have demonstrated the capacity to improve summary quality by mirroring a human-like iterative process of critique and refinement starting from the initial draft. Two strategies are designed to perform this iterative process: Prompt Chaining and Stepwise Prompt. Prompt chaining orchestrates the drafting, critiquing, and refining phases through a series of three discrete prompts, while Stepwise prompt integrates these phases within a single prompt. However, the relative effectiveness of the two methods has not been extensively studied. This paper is dedicated to examining and comparing these two methods in the context of text summarization to ascertain which method stands out as the most effective. Experimental results show that the prompt chaining method can produce a more favorable outcome. This might be because stepwise prompt might produce a simulated refinement process according to our various experiments. Since refinement is adaptable to diverse tasks, our conclusions have the potential to be extrapolated to other applications, thereby offering insights that may contribute to the broader development of LLMs.

Since the natural language processing (NLP) community started to make large language models (LLMs), such as GPT-4, act as a critic to evaluate the quality of generated texts, most of them only train a critique generation model of a specific scale on specific datasets. We argue that a comprehensive investigation on the key factor of LLM-based evaluation models, such as scaling properties, is lacking, so that it is still inconclusive whether these models have potential to replace GPT-4's evaluation in practical scenarios. In this paper, we propose a new critique generation model called CritiqueLLM, which includes a dialogue-based prompting method for high-quality referenced / reference-free evaluation data. Experimental results show that our model can achieve comparable evaluation performance to GPT-4 especially in system-level correlations, and even outperform GPT-4 in 3 out of 8 tasks in a challenging reference-free setting. We conduct detailed analysis to show promising scaling properties of our model in the quality of generated critiques. We also demonstrate that our generated critiques can act as scalable feedback to directly improve the generation quality of LLMs.

Ensuring the safety of large language models (LLMs) is critical as they are deployed in real-world applications. Existing guardrails rely on rule-based filtering or single-pass classification, limiting their ability to handle nuanced safety violations. To address this, we propose ThinkGuard, a critique-augmented guardrail model that distills knowledge from high-capacity LLMs by generating structured critiques alongside safety labels. Fine-tuned on critique-augmented data, the captured deliberative thinking ability drastically enhances the guardrail's cautiousness and interpretability. Evaluated on multiple safety benchmarks, ThinkGuard achieves the highest average F1 and AUPRC, outperforming all baselines. Compared to LLaMA Guard 3, ThinkGuard improves accuracy by 16.1% and macro F1 by 27.0%. Moreover, it surpasses label-only fine-tuned models, confirming that structured critiques enhance both classification precision and nuanced safety reasoning while maintaining computational efficiency.

Despite their remarkable performance, the development of Large Language Models (LLMs) faces a critical challenge in scalable oversight: providing effective feedback for tasks where human evaluation is difficult or where LLMs outperform humans. While there is growing interest in using LLMs for critique, current approaches still rely on human annotations or more powerful models, leaving the issue of enhancing critique capabilities without external supervision unresolved. We introduce SCRIT (Self-evolving CRITic), a framework that enables genuine self-evolution of critique abilities. Technically, SCRIT self-improves by training on synthetic data, generated by a contrastive-based self-critic that uses reference solutions for step-by-step critique, and a self-validation mechanism that ensures critique quality through correction outcomes. Implemented with Qwen2.5-72B-Instruct, one of the most powerful LLMs, SCRIT achieves up to a 10.3\% improvement on critique-correction and error identification benchmarks. Our analysis reveals that SCRIT's performance scales positively with data and model size, outperforms alternative approaches, and benefits critically from its self-validation component.

The ability of Large Language Models (LLMs) to critique and refine their reasoning is crucial for their application in evaluation, feedback provision, and self-improvement. This paper introduces CriticBench, a comprehensive benchmark designed to assess LLMs' abilities to critique and rectify their reasoning across a variety of tasks. CriticBench encompasses five reasoning domains: mathematical, commonsense, symbolic, coding, and algorithmic. It compiles 15 datasets and incorporates responses from three LLM families. Utilizing CriticBench, we evaluate and dissect the performance of 17 LLMs in generation, critique, and correction reasoning, i.e., GQC reasoning. Our findings reveal: (1) a linear relationship in GQC capabilities, with critique-focused training markedly enhancing performance; (2) a task-dependent variation in correction effectiveness, with logic-oriented tasks being more amenable to correction; (3) GQC knowledge inconsistencies that decrease as model size increases; and (4) an intriguing inter-model critiquing dynamic, where stronger models are better at critiquing weaker ones, while weaker models can surprisingly surpass stronger ones in their self-critique. We hope these insights into the nuanced critique-correct reasoning of LLMs will foster further research in LLM critique and self-improvement.

As AI capabilities increasingly surpass human proficiency in complex tasks, current alignment techniques including SFT and RLHF face fundamental challenges in ensuring reliable oversight. These methods rely on direct human assessment and become untenable when AI outputs exceed human cognitive thresholds. In response to this challenge, we explore two hypotheses: (1) critique of critique can be easier than critique itself, extending the widely-accepted observation that verification is easier than generation to the critique domain, as critique itself is a specialized form of generation; (2) this difficulty relationship is recursively held, suggesting that when direct evaluation is infeasible, performing high-order critiques (e.g., critique of critique of critique) offers a more tractable supervision pathway. To examine these hypotheses, we perform Human-Human, Human-AI, and AI-AI experiments across multiple tasks. Our results demonstrate encouraging evidence supporting these hypotheses and suggest that recursive self-critiquing is a promising direction for scalable oversight.

The ability of large vision-language models (LVLMs) to critique and correct their reasoning is an essential building block towards their self-improvement. However, a systematic analysis of such capabilities in LVLMs is still lacking. We propose VISCO, the first benchmark to extensively analyze the fine-grained critique and correction capabilities of LVLMs. Compared to existing work that uses a single scalar value to critique the entire reasoning [4], VISCO features dense and fine-grained critique, requiring LVLMs to evaluate the correctness of each step in the chain-of-thought and provide natural language explanations to support their judgments. Extensive evaluation of 24 LVLMs demonstrates that human-written critiques significantly enhance the performance after correction, showcasing the potential of the self-improvement strategy. However, the model-generated critiques are less helpful and sometimes detrimental to the performance, suggesting that critique is the crucial bottleneck. We identified three common patterns in critique failures: failure to critique visual perception, reluctance to "say no", and exaggerated assumption of error propagation. To address these issues, we propose an effective LookBack strategy that revisits the image to verify each piece of information in the initial reasoning. LookBack significantly improves critique and correction performance by up to 13.5%.

Reinforcement learning from human feedback (RLHF) is fundamentally limited by the capacity of humans to correctly evaluate model output. To improve human evaluation ability and overcome that limitation this work trains "critic" models that help humans to more accurately evaluate model-written code. These critics are themselves LLMs trained with RLHF to write natural language feedback highlighting problems in code from real-world assistant tasks. On code containing naturally occurring LLM errors model-written critiques are preferred over human critiques in 63% of cases, and human evaluation finds that models catch more bugs than human contractors paid for code review. We further confirm that our fine-tuned LLM critics can successfully identify hundreds of errors in ChatGPT training data rated as "flawless", even though the majority of those tasks are non-code tasks and thus out-of-distribution for the critic model. Critics can have limitations of their own, including hallucinated bugs that could mislead humans into making mistakes they might have otherwise avoided, but human-machine teams of critics and contractors catch similar numbers of bugs to LLM critics while hallucinating less than LLMs alone.

As Large Language Models (LLMs) are rapidly evolving, providing accurate feedback and scalable oversight on their outputs becomes an urgent and critical problem. Leveraging LLMs as critique models to achieve automated supervision is a promising solution. In this work, we focus on studying and enhancing the math critique ability of LLMs. Current LLM critics provide critiques that are too shallow and superficial on each step, leading to low judgment accuracy and struggling to offer sufficient feedback for the LLM generator to correct mistakes. To tackle this issue, we propose a novel and effective two-stage framework to develop LLM critics that are capable of deliberately critiquing on each reasoning step of math solutions. In the first stage, we utilize Qwen2.5-72B-Instruct to generate 4.5K long-form critiques as seed data for supervised fine-tuning. Each seed critique consists of deliberate step-wise critiques that includes multi-perspective verifications as well as in-depth critiques of initial critiques for each reasoning step. Then, we perform reinforcement learning on the fine-tuned model with either existing human-labeled data from PRM800K or our automatically annotated data obtained via Monte Carlo sampling-based correctness estimation, to further incentivize its critique ability. Our developed critique model built on Qwen2.5-7B-Instruct not only significantly outperforms existing LLM critics (including the same-sized DeepSeek-R1-distill models and GPT-4o) on various error identification benchmarks, but also more effectively helps the LLM generator refine erroneous steps through more detailed feedback.

Despite their unprecedented success, even the largest language models make mistakes. Similar to how humans learn and improve using feedback, previous work proposed providing language models with natural language feedback to guide them in repairing their outputs. Because human-generated critiques are expensive to obtain, researchers have devised learned critique generators in lieu of human critics while assuming one can train downstream models to utilize generated feedback. However, this approach does not apply to black-box or limited access models such as ChatGPT, as they cannot be fine-tuned. Moreover, in the era of large general-purpose language agents, fine-tuning is neither computationally nor spatially efficient as it results in multiple copies of the network. In this work, we introduce RL4F (Reinforcement Learning for Feedback), a multi-agent collaborative framework where the critique generator is trained to maximize end-task performance of GPT-3, a fixed model more than 200 times its size. RL4F produces critiques that help GPT-3 revise its outputs. We study three datasets for action planning, summarization and alphabetization and show improvements (~5% on average) in multiple text similarity metrics over strong baselines across all three tasks.

This study investigates the consistency of feedback ratings generated by OpenAI's GPT-4, a state-of-the-art artificial intelligence language model, across multiple iterations, time spans and stylistic variations. The model rated responses to tasks within the Higher Education (HE) subject domain of macroeconomics in terms of their content and style. Statistical analysis was conducted in order to learn more about the interrater reliability, consistency of the ratings across iterations and the correlation between ratings in terms of content and style. The results revealed a high interrater reliability with ICC scores ranging between 0.94 and 0.99 for different timespans, suggesting that GPT-4 is capable of generating consistent ratings across repetitions with a clear prompt. Style and content ratings show a high correlation of 0.87. When applying a non-adequate style the average content ratings remained constant, while style ratings decreased, which indicates that the large language model (LLM) effectively distinguishes between these two criteria during evaluation. The prompt used in this study is furthermore presented and explained. Further research is necessary to assess the robustness and reliability of AI models in various use cases.

Supervised Fine-Tuning (SFT) is commonly used to train language models to imitate annotated responses for given instructions. In this paper, we challenge this paradigm and propose Critique Fine-Tuning (CFT), a strategy where models learn to critique noisy responses rather than simply imitate correct ones. Inspired by human learning processes that emphasize critical thinking, CFT encourages deeper analysis and nuanced understanding-traits often overlooked by standard SFT. To validate the effectiveness of CFT, we construct a 50K-sample dataset from WebInstruct, using GPT-4o as the teacher to generate critiques in the form of (input=[query; noisy response], output=critique). CFT on this dataset yields a consistent 4-10% improvement over SFT on six math benchmarks with different base models like Qwen2.5, Qwen2.5-Math and DeepSeek-Math. We further expand to MetaMath and NuminaMath datasets and observe similar gains over SFT. Notably, our Qwen2.5-Math-CFT model-trained on just 50K samples-matches or outperforms competitive models such as AceMath and Qwen2.5-Math-Instruct on most benchmarks, both of which use over 2M samples. Ablation studies show that CFT is robust to the source of noisy response and teacher critique model. Through these findings, we argue that critique-based training offers a more effective alternative to advance the reasoning of language models.

Teaching large language models (LLMs) to critique and refine their outputs is crucial for building systems that can iteratively improve, yet it is fundamentally limited by the ability to provide accurate judgments and actionable suggestions. In this work, we study LLM critics for code generation and propose CTRL, a framework for Critic Training via Reinforcement Learning, which trains a critic model to generate feedback that maximizes correction performance for a fixed generator model without human supervision. Our results demonstrate that critics trained with CTRL significantly enhance pass rates and mitigate compounding errors across both base and stronger generator models. Furthermore, we show that these critic models act as accurate generative reward models and enable test-time scaling through iterative critique-revision, achieving up to 106.1% relative improvements across challenging code generation benchmarks.

Recent developments in large language models (LLMs) have been impressive. However, these models sometimes show inconsistencies and problematic behavior, such as hallucinating facts, generating flawed code, or creating offensive and toxic content. Unlike these models, humans typically utilize external tools to cross-check and refine their initial content, like using a search engine for fact-checking, or a code interpreter for debugging. Inspired by this observation, we introduce a framework called CRITIC that allows LLMs, which are essentially "black boxes" to validate and progressively amend their own outputs in a manner similar to human interaction with tools. More specifically, starting with an initial output, CRITIC interacts with appropriate tools to evaluate certain aspects of the text, and then revises the output based on the feedback obtained during this validation process. Comprehensive evaluations involving free-form question answering, mathematical program synthesis, and toxicity reduction demonstrate that CRITIC consistently enhances the performance of LLMs. Meanwhile, our research highlights the crucial importance of external feedback in promoting the ongoing self-improvement of LLMs.

What makes a good Large Language Model (LLM)? That it performs well on the relevant benchmarks -- which hopefully measure, with some validity, the presence of capabilities that are also challenged in real application. But what makes the model perform well? What gives a model its abilities? We take a recently introduced type of benchmark that is meant to challenge capabilities in a goal-directed, agentive context through self-play of conversational games, and analyse how performance develops as a function of model characteristics like number of parameters, or type of training. We find that while there is a clear relationship between number of parameters and performance, there is still a wide spread of performance points within a given size bracket, which is to be accounted for by training parameters such as fine-tuning data quality and method. From a more practical angle, we also find a certain degree of unpredictability about performance across access methods, possible due to unexposed sampling parameters, and a, very welcome, performance stability against at least moderate weight quantisation during inference.

Recent work has explored the capability of large language models (LLMs) to identify and correct errors in LLM-generated responses. These refinement approaches frequently evaluate what sizes of models are able to do refinement for what problems, but less attention is paid to what effective feedback for refinement looks like. In this work, we propose looking at refinement with feedback as a composition of three distinct LLM competencies: (1) identification of bad generations; (2) fine-grained natural language feedback generation; (3) refining with fine-grained feedback. The first step can be implemented with a high-performing discriminative model and steps 2 and 3 can be implemented either via prompted or fine-tuned LLMs. A key property of this approach is that the step 2 critique model can give fine-grained feedback about errors, made possible by offloading the discrimination to a separate model in step 1. We show that models of different capabilities benefit from refining with this approach on the task of improving factual consistency of document grounded summaries. Overall, our proposed method consistently outperforms existing end-to-end refinement approaches and current trained models not fine-tuned for factuality critiquing.

The reflection capacity of Large Language Model (LLM) has garnered extensive attention. A post-hoc prompting strategy, e.g., reflexion and self-refine, refines LLM's response based on self-evaluated or external feedback. However, recent research indicates without external feedback, LLM's intrinsic reflection is unstable. Our investigation unveils that the key bottleneck is the quality of the self-evaluated feedback. We find LLMs often exhibit overconfidence or high randomness when self-evaluate, offering stubborn or inconsistent feedback, which causes poor reflection. To remedy this, we advocate Self-Contrast: It adaptively explores diverse solving perspectives tailored to the request, contrasts the differences, and summarizes these discrepancies into a checklist which could be used to re-examine and eliminate discrepancies. Our method endows LLM with diverse perspectives to alleviate stubborn biases. Moreover, their discrepancies indicate potential errors or inherent uncertainties that LLM often overlooks. Reflecting upon these can catalyze more accurate and stable reflection. Experiments conducted on a series of reasoning and translation tasks with different LLMs serve to underscore the effectiveness and generality of our strategy.

In recent research on large language models (LLMs), there has been a growing emphasis on aligning these models with human values to reduce the impact of harmful content. However, current alignment methods often rely solely on singular forms of human feedback, such as preferences, annotated labels, or natural language critiques, overlooking the potential advantages of combining these feedback types. This limitation leads to suboptimal performance, even when ample training data is available. In this paper, we introduce Constructive and Diverse Feedback (CDF) as a novel method to enhance LLM alignment, inspired by constructivist learning theory. Our approach involves collecting three distinct types of feedback tailored to problems of varying difficulty levels within the training dataset. Specifically, we exploit critique feedback for easy problems, refinement feedback for medium problems, and preference feedback for hard problems. By training our model with this diversified feedback, we achieve enhanced alignment performance while using less training data. To assess the effectiveness of CDF, we evaluate it against previous methods in three downstream tasks: question answering, dialog generation, and text summarization. Experimental results demonstrate that CDF achieves superior performance even with a smaller training dataset.

Evaluation benchmarks are the cornerstone of measuring capabilities of large language models (LLMs), as well as driving progress in said capabilities. Originally designed to make claims about capabilities (or lack thereof) in fully pretrained models, evaluation benchmarks are now also extensively used to decide between various training choices. Despite this widespread usage, we rarely quantify the variance in our evaluation benchmarks, which dictates whether differences in performance are meaningful. Here, we define and measure a range of metrics geared towards measuring variance in evaluation benchmarks, including seed variance across initialisations, and monotonicity during training. By studying a large number of models -- both openly available and pretrained from scratch -- we provide empirical estimates for a variety of variance metrics, with considerations and recommendations for practitioners. We also evaluate the utility and tradeoffs of continuous versus discrete performance measures and explore options for better understanding and reducing this variance. We find that simple changes, such as framing choice tasks (like MMLU) as completion tasks, can often reduce variance for smaller scale (sim7B) models, while more involved methods inspired from human testing literature (such as item analysis and item response theory) struggle to meaningfully reduce variance. Overall, our work provides insights into variance in evaluation benchmarks, suggests LM-specific techniques to reduce variance, and more generally encourages practitioners to carefully factor in variance when comparing models.

Recent advancements in reasoning-based Large Language Models (LLMs), particularly their potential through test-time scaling, have created significant opportunities for distillation in code generation and critique. However, progress in both areas fundamentally depends on large-scale, high-quality datasets. In this work, we introduce OpenCodeReasoning-II, a dataset consists of 2.5M question-solution-critique triples (approx. 35K unique programming questions), making it nearly twice the size of the previous largest publicly available code reasoning dataset. In this work, we employ a two-stage supervised fine-tuning strategy. The first stage focuses on fine-tuning for code generation, while the second stage involves the joint training of models for both code generation and critique. Our resulting finetuned Qwen2.5-Instruct models achieve performance in code generation that either exceeds or equals the best prior open-weight distilled models. Notably, the integration of our code generation and critique models leads to significant improvements in competitive coding performance. Furthermore, we present an extension of the LiveCodeBench benchmark to specifically support the C++ programming language, thereby facilitating more comprehensive LLM evaluation using this benchmark.

The growing capabilities of large language models (LLMs) have led to their use as substitutes for human feedback for training and assessing other LLMs. These methods often rely on `constitutions', written guidelines which a critic model uses to provide feedback and improve generations. We investigate how the choice of constitution affects feedback quality by using four different constitutions to improve patient-centered communication in medical interviews. In pairwise comparisons conducted by 215 human raters, we found that detailed constitutions led to better results regarding emotive qualities. However, none of the constitutions outperformed the baseline in learning more practically-oriented skills related to information gathering and provision. Our findings indicate that while detailed constitutions should be prioritised, there are possible limitations to the effectiveness of AI feedback as a reward signal in certain areas.

With the rapid advancement of Large Language Models (LLMs), developing effective critic modules for precise guidance has become crucial yet challenging. In this paper, we initially demonstrate that supervised fine-tuning for building critic modules (which is widely adopted in current solutions) fails to genuinely enhance models' critique abilities, producing superficial critiques with insufficient reflections and verifications. To unlock the unprecedented critique capabilities, we propose RefCritic, a long-chain-of-thought critic module based on reinforcement learning with dual rule-based rewards: (1) instance-level correctness of solution judgments and (2) refinement accuracies of the policy model based on critiques, aiming to generate high-quality evaluations with actionable feedback that effectively guides model refinement. We evaluate RefCritic on Qwen2.5-14B-Instruct and DeepSeek-R1-Distill-Qwen-14B across five benchmarks. On critique and refinement settings, RefCritic demonstrates consistent advantages across all benchmarks, e.g., 6.8\% and 7.2\% gains on AIME25 for the respective base models. Notably, under majority voting, policy models filtered by RefCritic show superior scaling with increased voting numbers. Moreover, despite training on solution-level supervision, RefCritic outperforms step-level supervised approaches on ProcessBench, a benchmark to identify erroneous steps in mathematical reasoning.

Large Language Models (LLMs) can correct their self-generated responses, but a decline in accuracy after self-correction is also witnessed. To have a deeper understanding of self-correction, we endeavor to decompose, evaluate, and analyze the self-correction behaviors of LLMs. By enumerating and analyzing answer correctness before and after self-correction, we decompose the self-correction capability into confidence (being confident to correct answers) and critique (turning wrong answers to correct) capabilities, and propose two metrics from a probabilistic perspective to measure these 2 capabilities, along with another metric for overall self-correction capability evaluation. Based on our decomposition and evaluation metrics, we conduct extensive experiments and draw some empirical conclusions. For example, we find different models can exhibit distinct behaviors: some models are confident while others are more critical. We also find the trade-off between the two capabilities (i.e. improving one can lead to a decline in the other) when manipulating model self-correction behavior by prompts or in-context learning. Further, we find a simple yet efficient strategy to improve self-correction capability by transforming Supervision Fine-Tuning (SFT) data format, and our strategy outperforms vanilla SFT in both capabilities and achieves much higher accuracy after self-correction. Our code will be publicly available on GitHub.

Recent advancements in large language models (LLMs) have significantly enhanced text generation capabilities, yet evaluating their performance in generative writing remains a challenge. Existing benchmarks primarily focus on generic text generation or limited in writing tasks, failing to capture the diverse requirements of high-quality written contents across various domains. To bridge this gap, we present WritingBench, a comprehensive benchmark designed to evaluate LLMs across 6 core writing domains and 100 subdomains, encompassing creative, persuasive, informative, and technical writing. We further propose a query-dependent evaluation framework that empowers LLMs to dynamically generate instance-specific assessment criteria. This framework is complemented by a fine-tuned critic model for criteria-aware scoring, enabling evaluations in style, format and length. The framework's validity is further demonstrated by its data curation capability, which enables 7B-parameter models to approach state-of-the-art (SOTA) performance. We open-source the benchmark, along with evaluation tools and modular framework components, to advance the development of LLMs in writing.

As AI systems are used to answer more difficult questions and potentially help create new knowledge, judging the truthfulness of their outputs becomes more difficult and more important. How can we supervise unreliable experts, which have access to the truth but may not accurately report it, to give answers that are systematically true and don't just superficially seem true, when the supervisor can't tell the difference between the two on their own? In this work, we show that debate between two unreliable experts can help a non-expert judge more reliably identify the truth. We collect a dataset of human-written debates on hard reading comprehension questions where the judge has not read the source passage, only ever seeing expert arguments and short quotes selectively revealed by 'expert' debaters who have access to the passage. In our debates, one expert argues for the correct answer, and the other for an incorrect answer. Comparing debate to a baseline we call consultancy, where a single expert argues for only one answer which is correct half of the time, we find that debate performs significantly better, with 84% judge accuracy compared to consultancy's 74%. Debates are also more efficient, being 68% of the length of consultancies. By comparing human to AI debaters, we find evidence that with more skilled (in this case, human) debaters, the performance of debate goes up but the performance of consultancy goes down. Our error analysis also supports this trend, with 46% of errors in human debate attributable to mistakes by the honest debater (which should go away with increased skill); whereas 52% of errors in human consultancy are due to debaters obfuscating the relevant evidence from the judge (which should become worse with increased skill). Overall, these results show that debate is a promising approach for supervising increasingly capable but potentially unreliable AI systems.

Modern foundation models rely heavily on using scaling laws to guide crucial training decisions. Researchers often extrapolate the optimal architecture and hyper parameters settings from smaller training runs by describing the relationship between, loss, or task performance, and scale. All components of this process vary, from the specific equation being fit, to the training setup, to the optimization method. Each of these factors may affect the fitted law, and therefore, the conclusions of a given study. We discuss discrepancies in the conclusions that several prior works reach, on questions such as the optimal token to parameter ratio. We augment this discussion with our own analysis of the critical impact that changes in specific details may effect in a scaling study, and the resulting altered conclusions. Additionally, we survey over 50 papers that study scaling trends: while 45 of these papers quantify these trends using a power law, most under-report crucial details needed to reproduce their findings. To mitigate this, we we propose a checklist for authors to consider while contributing to scaling law research.

Reward modeling is crucial for aligning large language models (LLMs) with human preferences, especially in reinforcement learning from human feedback (RLHF). However, current reward models mainly produce scalar scores and struggle to incorporate critiques in a natural language format. We hypothesize that predicting both critiques and the scalar reward would improve reward modeling ability. Motivated by this, we propose Critic-RM, a framework that improves reward models using self-generated critiques without extra supervision. Critic-RM employs a two-stage process: generating and filtering high-quality critiques, followed by joint fine-tuning on reward prediction and critique generation. Experiments across benchmarks show that Critic-RM improves reward modeling accuracy by 3.7%-7.3% compared to standard reward models and LLM judges, demonstrating strong performance and data efficiency. Additional studies further validate the effectiveness of generated critiques in rectifying flawed reasoning steps with 2.5%-3.2% gains in improving reasoning accuracy.

Peer review is fundamental to scientific research, but the growing volume of publications has intensified the challenges of this expertise-intensive process. While LLMs show promise in various scientific tasks, their potential to assist with peer review, particularly in identifying paper limitations, remains understudied. We first present a comprehensive taxonomy of limitation types in scientific research, with a focus on AI. Guided by this taxonomy, for studying limitations, we present LimitGen, the first comprehensive benchmark for evaluating LLMs' capability to support early-stage feedback and complement human peer review. Our benchmark consists of two subsets: LimitGen-Syn, a synthetic dataset carefully created through controlled perturbations of high-quality papers, and LimitGen-Human, a collection of real human-written limitations. To improve the ability of LLM systems to identify limitations, we augment them with literature retrieval, which is essential for grounding identifying limitations in prior scientific findings. Our approach enhances the capabilities of LLM systems to generate limitations in research papers, enabling them to provide more concrete and constructive feedback.

Remarkable performance of large language models (LLMs) in a variety of tasks brings forth many opportunities as well as challenges of utilizing them in production settings. Towards practical adoption of LLMs, multi-agent systems hold great promise to augment, integrate, and orchestrate LLMs in the larger context of enterprise platforms that use existing proprietary data and models to tackle complex real-world tasks. Despite the tremendous success of these systems, current approaches rely on narrow, single-focus objectives for optimization and evaluation, often overlooking potential constraints in real-world scenarios, including restricted budgets, resources and time. Furthermore, interpreting, analyzing, and debugging these systems requires different components to be evaluated in relation to one another. This demand is currently not feasible with existing methodologies. In this postion paper, we introduce the concept of reasoning capacity as a unifying criterion to enable integration of constraints during optimization and establish connections among different components within the system, which also enable a more holistic and comprehensive approach to evaluation. We present a formal definition of reasoning capacity and illustrate its utility in identifying limitations within each component of the system. We then argue how these limitations can be addressed with a self-reflective process wherein human-feedback is used to alleviate shortcomings in reasoning and enhance overall consistency of the system.

Large Language Models (LLMs) have emerged as a milestone in artificial intelligence, and their performance can improve as the model size increases. However, this scaling brings great challenges to training and inference efficiency, particularly for deploying LLMs in resource-constrained environments, and the scaling trend is becoming increasingly unsustainable. This paper introduces the concept of ``capacity density'' as a new metric to evaluate the quality of the LLMs across different scales and describes the trend of LLMs in terms of both effectiveness and efficiency. To calculate the capacity density of a given target LLM, we first introduce a set of reference models and develop a scaling law to predict the downstream performance of these reference models based on their parameter sizes. We then define the effective parameter size of the target LLM as the parameter size required by a reference model to achieve equivalent performance, and formalize the capacity density as the ratio of the effective parameter size to the actual parameter size of the target LLM. Capacity density provides a unified framework for assessing both model effectiveness and efficiency. Our further analysis of recent open-source base LLMs reveals an empirical law (the densing law)that the capacity density of LLMs grows exponentially over time. More specifically, using some widely used benchmarks for evaluation, the capacity density of LLMs doubles approximately every three months. The law provides new perspectives to guide future LLM development, emphasizing the importance of improving capacity density to achieve optimal results with minimal computational overhead.

Expert feedback lays the foundation of rigorous research. However, the rapid growth of scholarly production and intricate knowledge specialization challenge the conventional scientific feedback mechanisms. High-quality peer reviews are increasingly difficult to obtain. Researchers who are more junior or from under-resourced settings have especially hard times getting timely feedback. With the breakthrough of large language models (LLM) such as GPT-4, there is growing interest in using LLMs to generate scientific feedback on research manuscripts. However, the utility of LLM-generated feedback has not been systematically studied. To address this gap, we created an automated pipeline using GPT-4 to provide comments on the full PDFs of scientific papers. We evaluated the quality of GPT-4's feedback through two large-scale studies. We first quantitatively compared GPT-4's generated feedback with human peer reviewer feedback in 15 Nature family journals (3,096 papers in total) and the ICLR machine learning conference (1,709 papers). The overlap in the points raised by GPT-4 and by human reviewers (average overlap 30.85% for Nature journals, 39.23% for ICLR) is comparable to the overlap between two human reviewers (average overlap 28.58% for Nature journals, 35.25% for ICLR). The overlap between GPT-4 and human reviewers is larger for the weaker papers. We then conducted a prospective user study with 308 researchers from 110 US institutions in the field of AI and computational biology to understand how researchers perceive feedback generated by our GPT-4 system on their own papers. Overall, more than half (57.4%) of the users found GPT-4 generated feedback helpful/very helpful and 82.4% found it more beneficial than feedback from at least some human reviewers. While our findings show that LLM-generated feedback can help researchers, we also identify several limitations.

Generating a long story of several thousand words with narrative coherence using Large Language Models (LLMs) has been a challenging task. Previous research has addressed this challenge by proposing different frameworks that create a story plan and generate a long story based on that plan. However, these frameworks have been mainly focusing on maintaining narrative coherence in stories, often overlooking creativity in story planning and the expressiveness of the stories generated from those plans, which are desirable properties to captivate readers' interest. In this paper, we propose Collective Critics for Creative Story Generation framework (CritiCS), which is composed of plan refining stage (CrPlan) and story generation stage (CrText), to integrate a collective revision mechanism that promotes those properties into long-form story generation process. Specifically, in each stage, a group of LLM critics and one leader collaborate to incrementally refine drafts of plan and story throughout multiple rounds. Extensive human evaluation shows that the CritiCS can significantly enhance story creativity and reader engagement, while also maintaining narrative coherence. Furthermore, the design of the framework allows active participation from human writers in any role within the critique process, enabling interactive human-machine collaboration in story writing.

The ability of large language models (LLMs) to utilize external tools has enabled them to tackle an increasingly diverse range of tasks. However, as the tasks become more complex and long-horizon, the intricate tool utilization process may trigger various unexpected errors. Therefore, how to effectively handle such errors, including identifying, diagnosing, and recovering from them, has emerged as a key research direction for advancing tool learning. In this work, we first extensively analyze the types of errors encountered during the function-calling process on several competitive tool evaluation benchmarks. Based on it, we introduce CRITICTOOL, a comprehensive critique evaluation benchmark specialized for tool learning. Building upon a novel evolutionary strategy for dataset construction, CRITICTOOL holds diverse tool-use errors with varying complexities, which better reflects real-world scenarios. We conduct extensive experiments on CRITICTOOL, and validate the generalization and effectiveness of our constructed benchmark strategy. We also provide an in-depth analysis of the tool reflection ability on various LLMs, offering a new perspective on the field of tool learning in LLMs. The code is available at https://github.com/Shellorley0513/CriticTool{https://github.com/Shellorley0513/CriticTool}.

Recent developments in LLMs offer new opportunities for assisting authors in improving their work. In this paper, we envision a use case where authors can receive LLM-generated reviews that uncover weak points in the current draft. While initial methods for automated review generation already exist, these methods tend to produce reviews that lack detail, and they do not cover the range of opinions that human reviewers produce. To address this shortcoming, we propose an efficient two-stage review generation framework called Reviewer2. Unlike prior work, this approach explicitly models the distribution of possible aspects that the review may address. We show that this leads to more detailed reviews that better cover the range of aspects that human reviewers identify in the draft. As part of the research, we generate a large-scale review dataset of 27k papers and 99k reviews that we annotate with aspect prompts, which we make available as a resource for future research.

We study whether language models can evaluate the validity of their own claims and predict which questions they will be able to answer correctly. We first show that larger models are well-calibrated on diverse multiple choice and true/false questions when they are provided in the right format. Thus we can approach self-evaluation on open-ended sampling tasks by asking models to first propose answers, and then to evaluate the probability "P(True)" that their answers are correct. We find encouraging performance, calibration, and scaling for P(True) on a diverse array of tasks. Performance at self-evaluation further improves when we allow models to consider many of their own samples before predicting the validity of one specific possibility. Next, we investigate whether models can be trained to predict "P(IK)", the probability that "I know" the answer to a question, without reference to any particular proposed answer. Models perform well at predicting P(IK) and partially generalize across tasks, though they struggle with calibration of P(IK) on new tasks. The predicted P(IK) probabilities also increase appropriately in the presence of relevant source materials in the context, and in the presence of hints towards the solution of mathematical word problems. We hope these observations lay the groundwork for training more honest models, and for investigating how honesty generalizes to cases where models are trained on objectives other than the imitation of human writing.

There have been widespread claims about Large Language Models (LLMs) being able to successfully verify or self-critique their candidate solutions in reasoning problems in an iterative mode. Intrigued by those claims, in this paper we set out to investigate the verification/self-critiquing abilities of large language models in the context of planning. We evaluate a planning system that employs LLMs for both plan generation and verification. We assess the verifier LLM's performance against ground-truth verification, the impact of self-critiquing on plan generation, and the influence of varying feedback levels on system performance. Using GPT-4, a state-of-the-art LLM, for both generation and verification, our findings reveal that self-critiquing appears to diminish plan generation performance, especially when compared to systems with external, sound verifiers and the LLM verifiers in that system produce a notable number of false positives, compromising the system's reliability. Additionally, the nature of feedback, whether binary or detailed, showed minimal impact on plan generation. Collectively, our results cast doubt on the effectiveness of LLMs in a self-critiquing, iterative framework for planning tasks.

To enable building and testing models on long-document comprehension, we introduce QuALITY, a multiple-choice QA dataset with context passages in English that have an average length of about 5,000 tokens, much longer than typical current models can process. Unlike in prior work with passages, our questions are written and validated by contributors who have read the entire passage, rather than relying on summaries or excerpts. In addition, only half of the questions are answerable by annotators working under tight time constraints, indicating that skimming and simple search are not enough to consistently perform well. Our baseline models perform poorly on this task (55.4%) and significantly lag behind human performance (93.5%).

Machine learning methods have proven useful in transcribing historical data. However, results from even highly accurate methods require manual verification and correction. Such manual review can be time-consuming and expensive, therefore the objective of this paper was to make it more efficient. Previously, we used machine learning to transcribe 2.3 million handwritten occupation codes from the Norwegian 1950 census with high accuracy (97%). We manually reviewed the 90,000 (3%) codes with the lowest model confidence. We allocated those 90,000 codes to human reviewers, who used our annotation tool to review the codes. To assess reviewer agreement, some codes were assigned to multiple reviewers. We then analyzed the review results to understand the relationship between accuracy improvements and effort. Additionally, we interviewed the reviewers to improve the workflow. The reviewers corrected 62.8% of the labels and agreed with the model label in 31.9% of cases. About 0.2% of the images could not be assigned a label, while for 5.1% the reviewers were uncertain, or they assigned an invalid label. 9,000 images were independently reviewed by multiple reviewers, resulting in an agreement of 86.43% and disagreement of 8.96%. We learned that our automatic transcription is biased towards the most frequent codes, with a higher degree of misclassification for the lowest frequency codes. Our interview findings show that the reviewers did internal quality control and found our custom tool well-suited. So, only one reviewer is needed, but they should report uncertainty.

It has become routine to report research results where Large Language Models (LLMs) outperform average humans in a wide range of language-related tasks, and creative text writing is no exception. It seems natural, then, to raise the bid: Are LLMs ready to compete in creative writing skills with a top (rather than average) novelist? To provide an initial answer for this question, we have carried out a contest between Patricio Pron (an awarded novelist, considered one of the best of his generation) and GPT-4 (one of the top performing LLMs), in the spirit of AI-human duels such as DeepBlue vs Kasparov and AlphaGo vs Lee Sidol. We asked Pron and GPT-4 to provide thirty titles each, and then to write short stories for both their titles and their opponent's. Then, we prepared an evaluation rubric inspired by Boden's definition of creativity, and we collected 5,400 manual assessments provided by literature critics and scholars. The results of our experimentation indicate that LLMs are still far from challenging a top human creative writer, and that reaching such level of autonomous creative writing skills probably cannot be reached simply with larger language models.

Large language models frequently encounter conflicts between their parametric knowledge and contextual input, often resulting in factual inconsistencies or hallucinations. We propose Self-Reflective Debate for Contextual Reliability (SR-DCR), a lightweight framework that integrates token-level self-confidence with an asymmetric multi-agent debate to adjudicate such conflicts. A critic, deprived of context, challenges a defender who argues from the given passage; a judge model evaluates the debate and determines the context's reliability. The final answer is selected by combining the verdict with model confidence. Experiments on the ClashEval benchmark demonstrate that SR-DCR consistently enhances robustness to misleading context while maintaining accuracy on trustworthy inputs, outperforming both classical debate and confidence-only baselines with minimal computational overhead. The code is available at https://github.com/smiles724/Self-Reflective-Debates.

Rigorous and reproducible evaluation is critical for assessing the state of the art and for guiding scientific advances in Artificial Intelligence. Evaluation is challenging in practice due to several reasons, including benchmark saturation, lack of transparency in methods used for measurement, development challenges in extracting measurements for generative tasks, and, more generally, the extensive number of capabilities required for a well-rounded comparison across models. We make three contributions to alleviate the above challenges. First, we present Eureka, an open-source framework for standardizing evaluations of large foundation models beyond single-score reporting and rankings. Second, we introduce Eureka-Bench as an extensible collection of benchmarks testing capabilities that (i) are still challenging for state-of-the-art models and (ii) represent fundamental but overlooked language and multimodal capabilities. The inherent space for improvement in non-saturated benchmarks enables us to discover meaningful differences between models at a capability level. Third, using Eureka, we conduct an analysis of 12 state-of-the-art models, providing in-depth insights into failure understanding and model comparison, which can be leveraged to plan targeted improvements. In contrast to recent trends in reports and leaderboards showing absolute rankings and claims for one model or another to be the best, our analysis shows that there is no such best model. Different models have different strengths, but there are models that appear more often than others as best performers for some capabilities. Despite the recent improvements, current models still struggle with several fundamental capabilities including detailed image understanding, benefiting from multimodal input when available rather than fully relying on language, factuality and grounding for information retrieval, and over refusals.

Large language models (LLMs) have shown excellent mastering of human language, but still struggle in real-world applications that require mathematical problem-solving. While many strategies and datasets to enhance LLMs' mathematics are developed, it remains a challenge to simultaneously maintain and improve both language and mathematical capabilities in deployed LLM systems.In this work, we tailor the Self-Critique pipeline, which addresses the challenge in the feedback learning stage of LLM alignment. We first train a general Math-Critique model from the LLM itself to provide feedback signals. Then, we sequentially employ rejective fine-tuning and direct preference optimization over the LLM's own generations for data collection. Based on ChatGLM3-32B, we conduct a series of experiments on both academic and our newly created challenging dataset, MathUserEval. Results show that our pipeline significantly enhances the LLM's mathematical problem-solving while still improving its language ability, outperforming LLMs that could be two times larger. Related techniques have been deployed to ChatGLM\url{https://chatglm.cn}, an online serving LLM. Related evaluation dataset and scripts are released at https://github.com/THUDM/ChatGLM-Math.

The impressive performance of large language models (LLMs) has attracted considerable attention from the academic and industrial communities. Besides how to construct and train LLMs, how to effectively evaluate and compare the capacity of LLMs has also been well recognized as an important yet difficult problem. Existing paradigms rely on either human annotators or model-based evaluators to evaluate the performance of LLMs on different tasks. However, these paradigms often suffer from high cost, low generalizability, and inherited biases in practice, which make them incapable of supporting the sustainable development of LLMs in long term. In order to address these issues, inspired by the peer review systems widely used in academic publication process, we propose a novel framework that can automatically evaluate LLMs through a peer-review process. Specifically, for the evaluation of a specific task, we first construct a small qualification exam to select "reviewers" from a couple of powerful LLMs. Then, to actually evaluate the "submissions" written by different candidate LLMs, i.e., the evaluatees, we use the reviewer LLMs to rate or compare the submissions. The final ranking of evaluatee LLMs is generated based on the results provided by all reviewers. We conducted extensive experiments on text summarization tasks with eleven LLMs including GPT-4. The results demonstrate the existence of biasness when evaluating using a single LLM. Also, our PRE model outperforms all the baselines, illustrating the effectiveness of the peer review mechanism.

With LLMs shifting their role from statistical modeling of language to serving as general-purpose AI agents, how should LLM evaluations change? Arguably, a key ability of an AI agent is to flexibly combine, as needed, the basic skills it has learned. The capability to combine skills plays an important role in (human) pedagogy and also in a paper on emergence phenomena (Arora & Goyal, 2023). This work introduces Skill-Mix, a new evaluation to measure ability to combine skills. Using a list of N skills the evaluator repeatedly picks random subsets of k skills and asks the LLM to produce text combining that subset of skills. Since the number of subsets grows like N^k, for even modest k this evaluation will, with high probability, require the LLM to produce text significantly different from any text in the training set. The paper develops a methodology for (a) designing and administering such an evaluation, and (b) automatic grading (plus spot-checking by humans) of the results using GPT-4 as well as the open LLaMA-2 70B model. Administering a version of to popular chatbots gave results that, while generally in line with prior expectations, contained surprises. Sizeable differences exist among model capabilities that are not captured by their ranking on popular LLM leaderboards ("cramming for the leaderboard"). Furthermore, simple probability calculations indicate that GPT-4's reasonable performance on k=5 is suggestive of going beyond "stochastic parrot" behavior (Bender et al., 2021), i.e., it combines skills in ways that it had not seen during training. We sketch how the methodology can lead to a Skill-Mix based eco-system of open evaluations for AI capabilities of future models.

We propose a new method for estimating how much a model ``knows'' about a datapoint and use it to measure the capacity of modern language models. Prior studies of language model memorization have struggled to disentangle memorization from generalization. We formally separate memorization into two components: unintended memorization, the information a model contains about a specific dataset, and generalization, the information a model contains about the true data-generation process. When we completely eliminate generalization, we can compute the total memorization, which provides an estimate of model capacity: our measurements estimate that GPT-style models have a capacity of approximately 3.6 bits per parameter. We train language models on datasets of increasing size and observe that models memorize until their capacity fills, at which point ``grokking'' begins, and unintended memorization decreases as models begin to generalize. We train hundreds of transformer language models ranging from 500K to 1.5B parameters and produce a series of scaling laws relating model capacity and data size to membership inference.

Large Language Models (LLMs) have emerged as powerful tools in mathematical theorem proving, particularly when utilizing formal languages such as LEAN. The major learning paradigm is expert iteration, which necessitates a pre-defined dataset comprising numerous mathematical problems. In this process, LLMs attempt to prove problems within the dataset and iteratively refine their capabilities through self-training on the proofs they discover. We propose to use large scale LEAN problem datasets Lean-workbook for expert iteration with more than 20,000 CPU days. During expert iteration, we found log-linear trends between solved problem amount with proof length and CPU usage. We train a critic model to select relatively easy problems for policy models to make trials and guide the model to search for deeper proofs. InternLM2.5-StepProver achieves open-source state-of-the-art on MiniF2F, Lean-Workbook-Plus, ProofNet, and Putnam benchmarks. Specifically, it achieves a pass of 65.9% on the MiniF2F-test and proves (or disproves) 17.0% of problems in Lean-Workbook-Plus which shows a significant improvement compared to only 9.5% of problems proved when Lean-Workbook-Plus was released. We open-source our models and searched proofs at https://github.com/InternLM/InternLM-Math and https://huggingface.co/datasets/internlm/Lean-Workbook.

In the experts problem, on each of T days, an agent needs to follow the advice of one of n ``experts''. After each day, the loss associated with each expert's advice is revealed. A fundamental result in learning theory says that the agent can achieve vanishing regret, i.e. their cumulative loss is within o(T) of the cumulative loss of the best-in-hindsight expert. Can the agent perform well without sufficient space to remember all the experts? We extend a nascent line of research on this question in two directions: bullet We give a new algorithm against the oblivious adversary, improving over the memory-regret tradeoff obtained by [PZ23], and nearly matching the lower bound of [SWXZ22]. bullet We also consider an adaptive adversary who can observe past experts chosen by the agent. In this setting we give both a new algorithm and a novel lower bound, proving that roughly n memory is both necessary and sufficient for obtaining o(T) regret.

Given the rapid ascent of large language models (LLMs), we study the question: (How) can large language models help in reviewing of scientific papers or proposals? We first conduct some pilot studies where we find that (i) GPT-4 outperforms other LLMs (Bard, Vicuna, Koala, Alpaca, LLaMa, Dolly, OpenAssistant, StableLM), and (ii) prompting with a specific question (e.g., to identify errors) outperforms prompting to simply write a review. With these insights, we study the use of LLMs (specifically, GPT-4) for three tasks: 1. Identifying errors: We construct 13 short computer science papers each with a deliberately inserted error, and ask the LLM to check for the correctness of these papers. We observe that the LLM finds errors in 7 of them, spanning both mathematical and conceptual errors. 2. Verifying checklists: We task the LLM to verify 16 closed-ended checklist questions in the respective sections of 15 NeurIPS 2022 papers. We find that across 119 {checklist question, paper} pairs, the LLM had an 86.6% accuracy. 3. Choosing the "better" paper: We generate 10 pairs of abstracts, deliberately designing each pair in such a way that one abstract was clearly superior than the other. The LLM, however, struggled to discern these relatively straightforward distinctions accurately, committing errors in its evaluations for 6 out of the 10 pairs. Based on these experiments, we think that LLMs have a promising use as reviewing assistants for specific reviewing tasks, but not (yet) for complete evaluations of papers or proposals.

Textual content is often the output of a collaborative writing process: We start with an initial draft, ask for suggestions, and repeatedly make changes. Agnostic of this process, today's language models are trained to generate only the final result. As a consequence, they lack several abilities crucial for collaborative writing: They are unable to update existing texts, difficult to control and incapable of verbally planning or explaining their actions. To address these shortcomings, we introduce PEER, a collaborative language model that is trained to imitate the entire writing process itself: PEER can write drafts, add suggestions, propose edits and provide explanations for its actions. Crucially, we train multiple instances of PEER able to infill various parts of the writing process, enabling the use of self-training techniques for increasing the quality, amount and diversity of training data. This unlocks PEER's full potential by making it applicable in domains for which no edit histories are available and improving its ability to follow instructions, to write useful comments, and to explain its actions. We show that PEER achieves strong performance across various domains and editing tasks.

In the last decade, several organizations have produced documents intended to standardize, in the normative sense, and promote guidance to our recent and rapid AI development. However, the full spectrum of ideas presented in these documents has not yet been analyzed, except for a few meta-analyses and critical reviews of the field. In this work, we seek to expand on the work done by past researchers and create a tool for better data visualization of the contents and nature of these documents, to understand whether there is consensus or similarity between the principles espoused by various institutions, which may inspire debates on future regulations. We also provide some preliminary thoughts and questions that could guide the continuity of the research through a critical analysis of the results acquired by our methodology into a sample size of 200 documents.

We evaluate a range of recent LLMs on English creative writing, a challenging and complex task that requires imagination, coherence, and style. We use a difficult, open-ended scenario chosen to avoid training data reuse: an epic narration of a single combat between Ignatius J. Reilly, the protagonist of the Pulitzer Prize-winning novel A Confederacy of Dunces (1980), and a pterodactyl, a prehistoric flying reptile. We ask several LLMs and humans to write such a story and conduct a human evalution involving various criteria such as fluency, coherence, originality, humor, and style. Our results show that some state-of-the-art commercial LLMs match or slightly outperform our writers in most dimensions; whereas open-source LLMs lag behind. Humans retain an edge in creativity, while humor shows a binary divide between LLMs that can handle it comparably to humans and those that fail at it. We discuss the implications and limitations of our study and suggest directions for future research.

This study investigates GPT-4's assessment of its performance in healthcare applications. A simple prompting technique was used to prompt the LLM with questions taken from the United States Medical Licensing Examination (USMLE) questionnaire and it was tasked to evaluate its confidence score before posing the question and after asking the question. The questionnaire was categorized into two groups-questions with feedback (WF) and questions with no feedback(NF) post-question. The model was asked to provide absolute and relative confidence scores before and after each question. The experimental findings were analyzed using statistical tools to study the variability of confidence in WF and NF groups. Additionally, a sequential analysis was conducted to observe the performance variation for the WF and NF groups. Results indicate that feedback influences relative confidence but doesn't consistently increase or decrease it. Understanding the performance of LLM is paramount in exploring its utility in sensitive areas like healthcare. This study contributes to the ongoing discourse on the reliability of AI, particularly of LLMs like GPT-4, within healthcare, offering insights into how feedback mechanisms might be optimized to enhance AI-assisted medical education and decision support.

We survey the large language model (LLM) serving area to understand the intricate dynamics between cost-efficiency and accuracy, which is magnified by the growing need for longer contextual understanding when deploying models at a massive scale. Our findings reveal that works in this space optimize along three distinct but conflicting goals: improving serving context length (C), improving serving accuracy (A), and improving serving performance (P). Drawing inspiration from the CAP theorem in databases, we propose a CAP principle for LLM serving, which suggests that any optimization can improve at most two of these three goals simultaneously. Our survey categorizes existing works within this framework. We find the definition and continuity of user-perceived measurement metrics are crucial in determining whether a goal has been met, akin to prior CAP databases in the wild. We recognize the CAP principle for LLM serving as a guiding principle, rather than a formal theorem, to inform designers of the inherent and dynamic trade-offs in serving models. As serving accuracy and performance have been extensively studied, this survey focuses on works that extend serving context length and address the resulting challenges.

In current NLP research, large-scale language models and their abilities are widely being discussed. Some recent works have also found notable failures of these models. Often these failure examples involve complex reasoning abilities. This work focuses on a simple commonsense ability, reasoning about when an action (or its effect) is feasible. To this end, we introduce FeasibilityQA, a question-answering dataset involving binary classification (BCQ) and multi-choice multi-correct questions (MCQ) that test understanding of feasibility. We show that even state-of-the-art models such as GPT-3, GPT-2, and T5 struggle to answer the feasibility questions correctly. Specifically, on MCQ and BCQ questions, GPT-3 achieves an accuracy of just (19%, 62%) and (25%, 64%) in zero-shot and few-shot settings, respectively. We also evaluate models by providing relevant knowledge statements required to answer the question. We find that the additional knowledge leads to a 7% gain in performance, but the overall performance still remains low. These results make one wonder how much commonsense knowledge about action feasibility is encoded in state-of-the-art models and how well they can reason about it.

This paper assesses the potential for large language models (LLMs) to serve as assistive tools in the creative writing process, by means of a single, in-depth case study. In the course of the study, we develop interactive and multi-voice prompting strategies that interleave background descriptions (scene setting, plot elements), instructions that guide composition, samples of text in the target style, and critical discussion of the given samples. We qualitatively evaluate the results from a literary critical perspective, as well as from the standpoint of computational creativity (a sub-field of artificial intelligence). Our findings lend support to the view that the sophistication of the results that can be achieved with an LLM mirrors the sophistication of the prompting.

Recent research has generated hope that inference scaling could allow weaker language models to match or exceed the accuracy of stronger models, such as by repeatedly sampling solutions to a coding problem until it passes unit tests. The central thesis of this paper is that there is no free lunch for inference scaling: indefinite accuracy improvement through resampling can only be realized if the "verifier" (in this case, a set of unit tests) is perfect. When the verifier is imperfect, as it almost always is in domains such as reasoning or coding (for example, unit tests have imperfect coverage), there is a nonzero probability of false positives: incorrect solutions that pass the verifier. Resampling cannot decrease this probability, so it imposes an upper bound to the accuracy of resampling-based inference scaling even with an infinite compute budget. We find that there is a very strong correlation between the model's single-sample accuracy (i.e. accuracy without unit tests) and its false positive rate on coding benchmarks HumanEval and MBPP, whose unit tests have limited coverage. Therefore, no amount of inference scaling of weaker models can enable them to match the single-sample accuracy of a sufficiently strong model (Fig. 1a). When we consider that false positives have a negative utility compared to abstaining from producing a solution, it bends the inference scaling curve further downward. Empirically, we find that the optimal number of samples can be less than 10 under realistic assumptions (Fig. 1b). Finally, we show that beyond accuracy, false positives may have other undesirable qualities, such as poor adherence to coding style conventions.

We revisit the well-studied problem of approximating a matrix product, A^TB, based on small space sketches S(A) and S(B) of A in R^{n times d} and Bin R^{n times m}. We are interested in the setting where the sketches must be computed independently of each other, except for the use of a shared random seed. We prove that, when A and B are sparse, methods based on coordinated random sampling can outperform classical linear sketching approaches, like Johnson-Lindenstrauss Projection or CountSketch. For example, to obtain Frobenius norm error epsilon|A|_F|B|_F, coordinated sampling requires sketches of size O(s/epsilon^2) when A and B have at most s leq d,m non-zeros per row. In contrast, linear sketching leads to sketches of size O(d/epsilon^2) and O(m/epsilon^2) for A and B. We empirically evaluate our approach on two applications: 1) distributed linear regression in databases, a problem motivated by tasks like dataset discovery and augmentation, and 2) approximating attention matrices in transformer-based language models. In both cases, our sampling algorithms yield an order of magnitude improvement over linear sketching.

We introduce Debate Speech Evaluation as a novel and challenging benchmark for assessing LLM judges. Evaluating debate speeches requires a deep understanding of the speech at multiple levels, including argument strength and relevance, the coherence and organization of the speech, the appropriateness of its style and tone, and so on. This task involves a unique set of cognitive abilities that have previously received limited attention in systematic LLM benchmarking. To explore such skills, we leverage a dataset of over 600 meticulously annotated debate speeches and present the first in-depth analysis of how state-of-the-art LLMs compare to human judges on this task. Our findings reveal a nuanced picture: while larger models can approximate individual human judgments in some respects, they differ substantially in their overall judgment behavior. We also investigate the ability of frontier LLMs to generate persuasive, opinionated speeches, showing that models may perform at a human level on this task.

Optimizing the phrasing of argumentative text is crucial in higher education and professional development. However, assessing whether and how the different claims in a text should be revised is a hard task, especially for novice writers. In this work, we explore the main challenges to identifying argumentative claims in need of specific revisions. By learning from collaborative editing behaviors in online debates, we seek to capture implicit revision patterns in order to develop approaches aimed at guiding writers in how to further improve their arguments. We systematically compare the ability of common word embedding models to capture the differences between different versions of the same text, and we analyze their impact on various types of writing issues. To deal with the noisy nature of revision-based corpora, we propose a new sampling strategy based on revision distance. Opposed to approaches from prior work, such sampling can be done without employing additional annotations and judgments. Moreover, we provide evidence that using contextual information and domain knowledge can further improve prediction results. How useful a certain type of context is, depends on the issue the claim is suffering from, though.

Feedback is crucial for every design process, such as user interface (UI) design, and automating design critiques can significantly improve the efficiency of the design workflow. Although existing multimodal large language models (LLMs) excel in many tasks, they often struggle with generating high-quality design critiques -- a complex task that requires producing detailed design comments that are visually grounded in a given design's image. Building on recent advancements in iterative refinement of text output and visual prompting methods, we propose an iterative visual prompting approach for UI critique that takes an input UI screenshot and design guidelines and generates a list of design comments, along with corresponding bounding boxes that map each comment to a specific region in the screenshot. The entire process is driven completely by LLMs, which iteratively refine both the text output and bounding boxes using few-shot samples tailored for each step. We evaluated our approach using Gemini-1.5-pro and GPT-4o, and found that human experts generally preferred the design critiques generated by our pipeline over those by the baseline, with the pipeline reducing the gap from human performance by 50% for one rating metric. To assess the generalizability of our approach to other multimodal tasks, we applied our pipeline to open-vocabulary object and attribute detection, and experiments showed that our method also outperformed the baseline.

Recent work in language modeling has raised the possibility of self-improvement, where a language models evaluates and refines its own generations to achieve higher performance without external feedback. It is impossible for this self-improvement to create information that is not already in the model, so why should we expect that this will lead to improved capabilities? We offer a new perspective on the capabilities of self-improvement through a lens we refer to as sharpening. Motivated by the observation that language models are often better at verifying response quality than they are at generating correct responses, we formalize self-improvement as using the model itself as a verifier during post-training in order to ``sharpen'' the model to one placing large mass on high-quality sequences, thereby amortizing the expensive inference-time computation of generating good sequences. We begin by introducing a new statistical framework for sharpening in which the learner aims to sharpen a pre-trained base policy via sample access, and establish fundamental limits. Then we analyze two natural families of self-improvement algorithms based on SFT and RLHF. We find that (i) the SFT-based approach is minimax optimal whenever the initial model has sufficient coverage, but (ii) the RLHF-based approach can improve over SFT-based self-improvement by leveraging online exploration, bypassing the need for coverage. Finally, we empirically validate the sharpening mechanism via inference-time and amortization experiments. We view these findings as a starting point toward a foundational understanding that can guide the design and evaluation of self-improvement algorithms.

Large Language Models (LLMs) have displayed massive improvements in reasoning and decision-making skills and can hold natural conversations with users. Recently, many tool-use benchmark datasets have been proposed. However, existing datasets have the following limitations: (1). Insufficient evaluation scenarios (e.g., only cover limited tool-use scenes). (2). Extensive evaluation costs (e.g., GPT API costs). To address these limitations, in this work, we propose a multi-granularity tool-use benchmark for large language models called MTU-Bench. For the "multi-granularity" property, our MTU-Bench covers five tool usage scenes (i.e., single-turn and single-tool, single-turn and multiple-tool, multiple-turn and single-tool, multiple-turn and multiple-tool, and out-of-distribution tasks). Besides, all evaluation metrics of our MTU-Bench are based on the prediction results and the ground truth without using any GPT or human evaluation metrics. Moreover, our MTU-Bench is collected by transforming existing high-quality datasets to simulate real-world tool usage scenarios, and we also propose an instruction dataset called MTU-Instruct data to enhance the tool-use abilities of existing LLMs. Comprehensive experimental results demonstrate the effectiveness of our MTU-Bench. Code and data will be released at https: //github.com/MTU-Bench-Team/MTU-Bench.git.

The development and evaluation of Large Language Models (LLMs) have largely focused on individual capabilities. However, this overlooks the intersection of multiple abilities across different types of expertise that are often required for real-world tasks, which we term cross capabilities. To systematically explore this concept, we first define seven core individual capabilities and then pair them to form seven common cross capabilities, each supported by a manually constructed taxonomy. Building on these definitions, we introduce CrossEval, a benchmark comprising 1,400 human-annotated prompts, with 100 prompts for each individual and cross capability. To ensure reliable evaluation, we involve expert annotators to assess 4,200 model responses, gathering 8,400 human ratings with detailed explanations to serve as reference examples. Our findings reveal that, in both static evaluations and attempts to enhance specific abilities, current LLMs consistently exhibit the "Law of the Weakest Link," where cross-capability performance is significantly constrained by the weakest component. Specifically, across 58 cross-capability scores from 17 models, 38 scores are lower than all individual capabilities, while 20 fall between strong and weak, but closer to the weaker ability. These results highlight the under-performance of LLMs in cross-capability tasks, making the identification and improvement of the weakest capabilities a critical priority for future research to optimize performance in complex, multi-dimensional scenarios.

We have witnessed that strong LLMs like Qwen-Math, MiMo, and Phi-4 possess immense reasoning potential inherited from the pre-training stage. With reinforcement learning (RL), these models can improve dramatically on reasoning tasks. Recent studies have shown that even RL on a single problem can unleash these models' reasoning capabilities. However, RL is not only expensive but also unstable. Even one-shot RL requires hundreds of GPU hours. This raises a critical question: Is there a more efficient way to unleash the reasoning potential of these powerful base LLMs? In this work, we demonstrate that Critique Fine-Tuning (CFT) on only one problem can effectively unleash the reasoning potential of LLMs. Our method constructs critique data by collecting diverse model-generated solutions to a single problem and using teacher LLMs to provide detailed critiques. We fine-tune Qwen and Llama family models, ranging from 1.5B to 14B parameters, on the CFT data and observe significant performance gains across diverse reasoning tasks. For example, with just 5 GPU hours of training, Qwen-Math-7B-CFT show an average improvement of 15% on six math benchmarks and 16% on three logic reasoning benchmarks. These results are comparable to or even surpass the results from RL with 20x less compute. Ablation studies reveal the robustness of one-shot CFT across different prompt problems. These results highlight one-shot CFT as a simple, general, and compute-efficient approach to unleashing the reasoning capabilities of modern LLMs.

Rapid advancements in large language models (LLMs) have the potential to assist in scientific progress. A critical capability toward this endeavor is the ability to reproduce existing work. To evaluate the ability of AI agents to reproduce results in an active research area, we introduce the Automated LLM Speedrunning Benchmark, leveraging the research community contributions on the NanoGPT speedrun, a competition to train a GPT-2 model in the shortest time. Each of the 19 speedrun tasks provides the agent with the previous records training script, optionally paired with one of three hint formats, ranging from pseudocode to paper-like descriptions of the new records improvements. Records execute quickly by design and speedrun improvements encompass diverse code-level changes, ranging from high-level algorithmic advancements to hardware-aware optimizations. These features make the benchmark both accessible and realistic for the frontier problem of improving LLM training. We find that recent reasoning LLMs combined with SoTA scaffolds struggle to reimplement already-known innovations in our benchmark, even when given detailed hints. Our benchmark thus provides a simple, non-saturated measure of an LLMs ability to automate scientific reproduction, a necessary (but not sufficient) skill for an autonomous research agent.

Large Language Models (LLMs) are often described as being instances of foundation models - that is, models that transfer strongly across various tasks and conditions in few-show or zero-shot manner, while exhibiting scaling laws that predict function improvement when increasing the pre-training scale. These claims of excelling in different functions and tasks rely on measurements taken across various sets of standardized benchmarks showing high scores for such models. We demonstrate here a dramatic breakdown of function and reasoning capabilities of state-of-the-art models trained at the largest available scales which claim strong function, using a simple, short, conventional common sense problem formulated in concise natural language, easily solvable by humans. The breakdown is dramatic, as models also express strong overconfidence in their wrong solutions, while providing often non-sensical "reasoning"-like explanations akin to confabulations to justify and backup the validity of their clearly failed responses, making them sound plausible. Various standard interventions in an attempt to get the right solution, like various type of enhanced prompting, or urging the models to reconsider the wrong solutions again by multi step re-evaluation, fail. We take these initial observations to the scientific and technological community to stimulate urgent re-assessment of the claimed capabilities of current generation of LLMs, Such re-assessment also requires common action to create standardized benchmarks that would allow proper detection of such basic reasoning deficits that obviously manage to remain undiscovered by current state-of-the-art evaluation procedures and benchmarks. Code for reproducing experiments in the paper and raw experiments data can be found at https://github.com/LAION-AI/AIW

As multimodal large language models (MLLMs) advance rapidly, rigorous evaluation has become essential, providing further guidance for their development. In this work, we focus on a unified and robust evaluation of vision perception abilities, the foundational skill of MLLMs. We find that existing perception benchmarks, each focusing on different question types, domains, and evaluation metrics, introduce significant evaluation variance, complicating comprehensive assessments of perception abilities when relying on any single benchmark. To address this, we introduce AbilityLens, a unified benchmark designed to evaluate MLLMs across six key perception abilities, focusing on both accuracy and stability, with each ability encompassing diverse question types, domains, and metrics. With the assistance of AbilityLens, we: (1) identify the strengths and weaknesses of current models, highlighting stability patterns and revealing a notable performance gap between open-source and closed-source models; (2) introduce an online evaluation mode, which uncovers interesting ability conflict and early convergence phenomena during MLLM training; and (3) design a simple ability-specific model merging method that combines the best ability checkpoint from early training stages, effectively mitigating performance decline due to ability conflict. The benchmark and online leaderboard will be released soon.

Uncovering early-stage metrics that reflect final model performance is one core principle for large-scale pretraining. The existing scaling law demonstrates the power-law correlation between pretraining loss and training flops, which serves as an important indicator of the current training state for large language models. However, this principle only focuses on the model's compression properties on the training data, resulting in an inconsistency with the ability improvements on the downstream tasks. Some follow-up works attempted to extend the scaling-law to more complex metrics (such as hyperparameters), but still lacked a comprehensive analysis of the dynamic differences among various capabilities during pretraining. To address the aforementioned limitations, this paper undertakes a comprehensive comparison of model capabilities at various pretraining intermediate checkpoints. Through this analysis, we confirm that specific downstream metrics exhibit similar training dynamics across models of different sizes, up to 67 billion parameters. In addition to our core findings, we've reproduced Amber and OpenLLaMA, releasing their intermediate checkpoints. This initiative offers valuable resources to the research community and facilitates the verification and exploration of LLM pretraining by open-source researchers. Besides, we provide empirical summaries, including performance comparisons of different models and capabilities, and tuition of key metrics for different training phases. Based on these findings, we provide a more user-friendly strategy for evaluating the optimization state, offering guidance for establishing a stable pretraining process.

Most popular benchmarks for comparing LLMs rely on a limited set of prompt templates, which may not fully capture the LLMs' abilities and can affect the reproducibility of results on leaderboards. Many recent works empirically verify prompt sensitivity and advocate for changes in LLM evaluation. In this paper, we consider the problem of estimating the performance distribution across many prompt variants instead of finding a single prompt to evaluate with. We introduce PromptEval, a method for estimating performance across a large set of prompts borrowing strength across prompts and examples to produce accurate estimates under practical evaluation budgets. The resulting distribution can be used to obtain performance quantiles to construct various robust performance metrics (e.g., top 95% quantile or median). We prove that PromptEval consistently estimates the performance distribution and demonstrate its efficacy empirically on three prominent LLM benchmarks: MMLU, BIG-bench Hard, and LMentry. For example, PromptEval can accurately estimate performance quantiles across 100 prompt templates on MMLU with a budget equivalent to two single-prompt evaluations. Our code and data can be found at https://github.com/felipemaiapolo/prompt-eval.

We propose an approach to estimate the number of samples required for a model to reach a target performance. We find that the power law, the de facto principle to estimate model performance, leads to large error when using a small dataset (e.g., 5 samples per class) for extrapolation. This is because the log-performance error against the log-dataset size follows a nonlinear progression in the few-shot regime followed by a linear progression in the high-shot regime. We introduce a novel piecewise power law (PPL) that handles the two data regimes differently. To estimate the parameters of the PPL, we introduce a random forest regressor trained via meta learning that generalizes across classification/detection tasks, ResNet/ViT based architectures, and random/pre-trained initializations. The PPL improves the performance estimation on average by 37% across 16 classification and 33% across 10 detection datasets, compared to the power law. We further extend the PPL to provide a confidence bound and use it to limit the prediction horizon that reduces over-estimation of data by 76% on classification and 91% on detection datasets.

The recent success of Large Language Models (LLMs) has catalyzed an increasing interest in their self-correction capabilities. This paper presents a comprehensive investigation into the intrinsic self-correction of LLMs, attempting to address the ongoing debate about its feasibility. Our research has identified an important latent factor - the "confidence" of LLMs - during the self-correction process. Overlooking this factor may cause the models to over-criticize themselves, resulting in unreliable conclusions regarding the efficacy of self-correction. We have experimentally observed that LLMs possess the capability to understand the "confidence" in their own responses. It motivates us to develop an "If-or-Else" (IoE) prompting framework, designed to guide LLMs in assessing their own "confidence", facilitating intrinsic self-corrections. We conduct extensive experiments and demonstrate that our IoE-based Prompt can achieve a consistent improvement regarding the accuracy of self-corrected responses over the initial answers. Our study not only sheds light on the underlying factors affecting self-correction in LLMs, but also introduces a practical framework that utilizes the IoE prompting principle to efficiently improve self-correction capabilities with "confidence". The code is available at https://github.com/MBZUAI-CLeaR/IoE-Prompting.git.

In this paper, we evaluate the creative fiction writing abilities of a fine-tuned small language model (SLM), BART Large, and compare its performance to humans and two large language models (LLMs): GPT-3.5 and GPT-4o. Our evaluation consists of two experiments: (i) a human evaluation where readers assess the stories generated by the SLM compared to human-written stories, and (ii) a qualitative linguistic analysis comparing the textual characteristics of the stories generated by the different models. In the first experiment, we asked 68 participants to rate short stories generated by the models and humans along dimensions such as grammaticality, relevance, creativity, and attractiveness. BART Large outperformed human writers in most aspects, except creativity, with an overall score of 2.11 compared to 1.85 for human-written texts -- a 14% improvement. In the second experiment, the qualitative analysis revealed that, while GPT-4o exhibited near-perfect internal and external coherence, it tended to produce more predictable narratives, with only 3% of its stories seen as novel. In contrast, 15% of BART's stories were considered novel, indicating a higher degree of creativity despite its smaller model size. This study provides both quantitative and qualitative insights into how model size and fine-tuning influence the balance between creativity, fluency, and coherence in creative writing tasks.

We test the robustness of debate as a method of scalable oversight by training models to debate with data generated via self-play. In a long-context reading comprehension task, we find that language model based evaluators answer questions more accurately when judging models optimized to win debates. By contrast, we find no such relationship for consultancy models trained to persuade a judge without an opposing debater present. In quantitative and qualitative comparisons between our debate models and novel consultancy baselines, we find evidence that debate training encourages stronger and more informative arguments, showing promise that it can help provide high-quality supervision for tasks that are difficult to directly evaluate.

State-of-the-art language models (LMs) are notoriously susceptible to generating hallucinated information. Such inaccurate outputs not only undermine the reliability of these models but also limit their use and raise serious concerns about misinformation and propaganda. In this work, we focus on hallucinated book and article references and present them as the "model organism" of language model hallucination research, due to their frequent and easy-to-discern nature. We posit that if a language model cites a particular reference in its output, then it should ideally possess sufficient information about its authors and content, among other relevant details. Using this basic insight, we illustrate that one can identify hallucinated references without ever consulting any external resources, by asking a set of direct or indirect queries to the language model about the references. These queries can be considered as "consistency checks." Our findings highlight that while LMs, including GPT-4, often produce inconsistent author lists for hallucinated references, they also often accurately recall the authors of real references. In this sense, the LM can be said to "know" when it is hallucinating references. Furthermore, these findings show how hallucinated references can be dissected to shed light on their nature. Replication code and results can be found at https://github.com/microsoft/hallucinated-references.

Code generation has attracted increasing attention with the rise of Large Language Models (LLMs). Many studies have developed powerful code LLMs by synthesizing code-related instruction data and applying supervised fine-tuning. However, these methods are limited by teacher model distillation and ignore the potential of iterative refinement by self-generated code. In this paper, we propose Adaptive Critique Refinement (ACR), which enables the model to refine itself by self-generated code and external critique, rather than directly imitating the code responses of the teacher model. Concretely, ACR includes a composite scoring system with LLM-as-a-Judge to evaluate the quality of code responses and a selective critique strategy with LLM-as-a-Critic to critique self-generated low-quality code responses. We develop the RefineCoder series by iteratively applying ACR, achieving continuous performance improvement on multiple code generation benchmarks. Compared to the baselines of the same size, our proposed RefineCoder series can achieve comparable or even superior performance using less data.

Large Language Models (LLMs) have demonstrated human-like instruction-following abilities, particularly those exceeding 100 billion parameters. The combined capability of some smaller, resource-friendly LLMs can address most of the instructions that larger LLMs excel at. In this work, we explore how to route the best-performing LLM for each instruction to achieve better overall performance. We develop a new paradigm, constructing capability instructions with model capability representation, user instruction, and performance inquiry prompts to assess the performance. To learn from capability instructions, we introduce a new end-to-end framework called Model Selection with Aptitude Test (Model-SAT), which generates positive and negative samples based on what different models perform well or struggle with. Model-SAT uses a model capability encoder that extends its model representation to a lightweight LLM. Our experiments show that Model-SAT understands the performance dimensions of candidate models and provides the probabilities of their capability to handle various instructions. Additionally, during deployment, a new model can quickly infer its aptitude test results across 50 tasks, each with 20 shots. Model-SAT performs state-of-the-art model routing without candidate inference and in real-world new model-released scenarios. The code is available at https://github.com/Now-Join-Us/CIT-LLM-Routing

The unprecedented performance of large language models (LLMs) necessitates improvements in evaluations. Rather than merely exploring the breadth of LLM abilities, we believe meticulous and thoughtful designs are essential to thorough, unbiased, and applicable evaluations. Given the importance of world knowledge to LLMs, we construct a Knowledge-oriented LLM Assessment benchmark (KoLA), in which we carefully design three crucial factors: (1) For ability modeling, we mimic human cognition to form a four-level taxonomy of knowledge-related abilities, covering 19 tasks. (2) For data, to ensure fair comparisons, we use both Wikipedia, a corpus prevalently pre-trained by LLMs, along with continuously collected emerging corpora, aiming to evaluate the capacity to handle unseen data and evolving knowledge. (3) For evaluation criteria, we adopt a contrastive system, including overall standard scores for better numerical comparability across tasks and models and a unique self-contrast metric for automatically evaluating knowledge hallucination. We evaluate 21 open-source and commercial LLMs and obtain some intriguing findings. The KoLA dataset and open-participation leaderboard are publicly released at https://kola.xlore.cn and will be continuously updated to provide references for developing LLMs and knowledge-related systems.

Interestingly, LLMs yet struggle with some basic tasks that humans find trivial to handle, e.g., counting the number of character r's in the word "strawberry". There are several popular conjectures (e.g., tokenization, architecture and training data) regarding the reason for deficiency of LLMs in simple word-based counting problems, sharing the similar belief that such failure stems from model pretraining hence probably inevitable during deployment. In this paper, we carefully design multiple evaluation settings to investigate validity of prevalent conjectures. Meanwhile, we measure transferability of advanced mathematical and coding reasoning capabilities from specialized LLMs to simple counting tasks. Although specialized LLMs suffer from counting problems as well, we find conjectures about inherent deficiency of LLMs invalid and further seek opportunities to elicit knowledge and capabilities from LLMs that are beneficial to counting tasks. Compared with strategies such as finetuning and in-context learning that are commonly adopted to enhance performance on new or challenging tasks, we show that engaging reasoning is the most robust and efficient way to help LLMs better perceive tasks with more accurate responses. We hope our conjecture validation design could provide insights into the study of future critical failure modes of LLMs. Based on challenges in transferring advanced capabilities to much simpler tasks, we call for more attention to model capability acquisition and evaluation. We also highlight the importance of cultivating consciousness of "reasoning before responding" during model pretraining.

Modern Code Review (MCR) is an informal tool-assisted quality assurance practice. It relies on the asynchronous communication among the authors of code changes and reviewers, who are developers that provide feedback. However, from candidate developers, some are able to provide better feedback than others given a particular context. The selection of reviewers is thus an important task, which can benefit from automated support. Many approaches have been proposed in this direction, using for example data from code review repositories to recommend reviewers. In this paper, we propose the use of team-related features to improve the performance of predictions that are helpful to build code reviewer recommenders, with our target predictions being the identification of reviewers that would participate in a review and the provided amount of feedback. We evaluate the prediction power of these features, which are related to code ownership, workload, and team relationship. This evaluation was done by carefully addressing challenges imposed by the MCR domain, such as temporal aspects of the dataset and unbalanced classes. Moreover, given that it is currently unknown how much past data is needed for building MCR prediction models with acceptable performance, we explore the amount of past data used to build prediction models. Our results show that, individually, features related to code ownership have the best prediction power. However, based on feature selection, we conclude that all proposed features together with lines of code can make the best predictions for both reviewer participation and amount of feedback. Regarding the amount of past data, the timeframes of 3, 6, 9, and 12 months of data produce similar results. Therefore, models can be trained considering short timeframes, thus reducing the computational costs with negligible impact in the prediction performance ...

Scaling the capacity of language models has consistently proven to be a reliable approach for improving performance and unlocking new capabilities. Capacity can be primarily defined by two dimensions: the number of model parameters and the compute per example. While scaling typically involves increasing both, the precise interplay between these factors and their combined contribution to overall capacity remains not fully understood. We explore this relationship in the context of sparse Mixture-of-Experts (MoEs), which allow scaling the number of parameters without proportionally increasing the FLOPs per example. We investigate how varying the sparsity level, i.e., the fraction of inactive parameters, impacts model's performance during pretraining and downstream few-shot evaluation. We find that under different constraints (e.g., parameter size and total training compute), there is an optimal level of sparsity that improves both training efficiency and model performance. These results provide a better understanding of the impact of sparsity in scaling laws for MoEs and complement existing works in this area, offering insights for designing more efficient architectures.

The rapid advancement of Large Language Models (LLMs) necessitates robust evaluation methodologies. Current benchmarking approaches often rely on comparing model outputs against predefined prompts and reference outputs. Relying on predefined reference outputs hinders flexible adaptation of benchmarks to the rapidly evolving capabilities of LLMs. This limitation necessitates periodic efforts to prepare new benchmarks. To keep pace with rapidly evolving LLM capabilities, we propose a more flexible benchmarking approach. Our method, \textbf{Varco Arena}, provides reference-free benchmarking of LLMs in tournament style. \textbf{Varco Arena} directly compares LLM outputs across a diverse set of prompts, determining model rankings through a single-elimination tournament structure. This direct pairwise comparison offers two key advantages: (1) Direct comparison, unmediated by reference text, more effectively orders competing LLMs, resulting in more reliable rankings, and (2) reference-free approach to benchmarking adds flexibility in updating benchmark prompts by eliminating the need for quality references. Our empirical results, supported by simulation experiments, demonstrate that the \textbf{Varco Arena} tournament approach aligns better with the current Elo model for benchmarking LLMs. The alignment is measured in terms of Spearman correlation, showing improvement over current practice of benchmarking that use reference outputs as comparison anchors.

Despite its importance to experimental design, statistical power (the probability that, given a real effect, an experiment will reject the null hypothesis) has largely been ignored by the NLP community. Underpowered experiments make it more difficult to discern the difference between statistical noise and meaningful model improvements, and increase the chances of exaggerated findings. By meta-analyzing a set of existing NLP papers and datasets, we characterize typical power for a variety of settings and conclude that underpowered experiments are common in the NLP literature. In particular, for several tasks in the popular GLUE benchmark, small test sets mean that most attempted comparisons to state of the art models will not be adequately powered. Similarly, based on reasonable assumptions, we find that the most typical experimental design for human rating studies will be underpowered to detect small model differences, of the sort that are frequently studied. For machine translation, we find that typical test sets of 2000 sentences have approximately 75% power to detect differences of 1 BLEU point. To improve the situation going forward, we give an overview of best practices for power analysis in NLP and release a series of notebooks to assist with future power analyses.

Academic citations are widely used for evaluating research and tracing knowledge flows. Such uses typically rely on raw citation counts and neglect variability in citation types. In particular, citations can vary in their fidelity as original knowledge from cited studies may be paraphrased, summarized, or reinterpreted, possibly wrongly, leading to variation in how much information changes from cited to citing paper. In this study, we introduce a computational pipeline to quantify citation fidelity at scale. Using full texts of papers, the pipeline identifies citations in citing papers and the corresponding claims in cited papers, and applies supervised models to measure fidelity at the sentence level. Analyzing a large-scale multi-disciplinary dataset of approximately 13 million citation sentence pairs, we find that citation fidelity is higher when authors cite papers that are 1) more recent and intellectually close, 2) more accessible, and 3) the first author has a lower H-index and the author team is medium-sized. Using a quasi-experiment, we establish the "telephone effect" - when citing papers have low fidelity to the original claim, future papers that cite the citing paper and the original have lower fidelity to the original. Our work reveals systematic differences in citation fidelity, underscoring the limitations of analyses that rely on citation quantity alone and the potential for distortion of evidence.

Abstention Ability (AA) is a critical aspect of Large Language Model (LLM) reliability, referring to an LLM's capability to withhold responses when uncertain or lacking a definitive answer, without compromising performance. Although previous studies have attempted to improve AA, they lack a standardised evaluation method and remain unsuitable for black-box models where token prediction probabilities are inaccessible. This makes comparative analysis challenging, especially for state-of-the-art closed-source commercial LLMs. This paper bridges this gap by introducing a black-box evaluation approach and a new dataset, Abstain-QA, crafted to rigorously assess AA across varied question types (answerable and unanswerable), domains (well-represented and under-represented), and task types (fact centric and reasoning). We also propose a new confusion matrix, the ''Answerable-Unanswerable Confusion Matrix'' (AUCM) which serves as the basis for evaluating AA, by offering a structured and precise approach for assessment. Finally, we explore the impact of three prompting strategies-Strict Prompting, Verbal Confidence Thresholding, and Chain-of-Thought (CoT)-on improving AA. Our results indicate that even powerful models like GPT-4, Mixtral 8x22b encounter difficulties with abstention; however, strategic approaches such as Strict prompting and CoT can enhance this capability.

There is a growing literature on reasoning by large language models (LLMs), but the discussion on the uncertainty in their responses is still lacking. Our aim is to assess the extent of confidence that LLMs have in their answers and how it correlates with accuracy. Confidence is measured (i) qualitatively in terms of persistence in keeping their answer when prompted to reconsider, and (ii) quantitatively in terms of self-reported confidence score. We investigate the performance of three LLMs -- GPT4o, GPT4-turbo and Mistral -- on two benchmark sets of questions on causal judgement and formal fallacies and a set of probability and statistical puzzles and paradoxes. Although the LLMs show significantly better performance than random guessing, there is a wide variability in their tendency to change their initial answers. There is a positive correlation between qualitative confidence and accuracy, but the overall accuracy for the second answer is often worse than for the first answer. There is a strong tendency to overstate the self-reported confidence score. Confidence is only partially explained by the underlying token-level probability. The material effects of prompting on qualitative confidence and the strong tendency for overconfidence indicate that current LLMs do not have any internally coherent sense of confidence.

Recent advances in large language models (LLMs), such as OpenAI-o1 and DeepSeek-R1, have demonstrated the effectiveness of test-time scaling, where extended reasoning processes substantially enhance model performance. Despite this, current models are constrained by limitations in handling long texts and reinforcement learning (RL) training efficiency. To address these issues, we propose a simple yet effective test-time scaling approach Multi-round Thinking. This method iteratively refines model reasoning by leveraging previous answers as prompts for subsequent rounds. Extensive experiments across multiple models, including QwQ-32B and DeepSeek-R1, consistently show performance improvements on various benchmarks such as AIME 2024, MATH-500, GPQA-diamond, and LiveCodeBench. For instance, the accuracy of QwQ-32B improved from 80.3% (Round 1) to 82.1% (Round 2) on the AIME 2024 dataset, while DeepSeek-R1 showed a similar increase from 79.7% to 82.0%. These results confirm that Multi-round Thinking is a broadly applicable, straightforward approach to achieving stable enhancements in model performance, underscoring its potential for future developments in test-time scaling techniques. The key prompt: {Original question prompt} The assistant's previous answer is: <answer> {last round answer} </answer>, and please re-answer.

This work presents a framework for assessing whether large language models (LLMs) encode more factual knowledge in their parameters than what they express in their outputs. While a few studies hint at this possibility, none has clearly defined or demonstrated this phenomenon. We first propose a formal definition of knowledge, quantifying it for a given question as the fraction of correct-incorrect answer pairs where the correct one is ranked higher. This gives rise to external and internal knowledge, depending on the information used to score individual answer candidates: either the model's observable token-level probabilities or its intermediate computations. Hidden knowledge arises when internal knowledge exceeds external knowledge. We then present a case study, applying this framework to three popular open-weights LLMs in a closed-book QA setup. Our results indicate that: (1) LLMs consistently encode more factual knowledge internally than what they express externally, with an average gap of 40%. (2) Surprisingly, some knowledge is so deeply hidden that a model can internally know an answer perfectly, yet fail to generate it even once, despite large-scale repeated sampling of 1,000 answers. This reveals fundamental limitations in the generation capabilities of LLMs, which (3) puts a practical constraint on scaling test-time compute via repeated answer sampling in closed-book QA: significant performance improvements remain inaccessible because some answers are practically never sampled, yet if they were, we would be guaranteed to rank them first.

The dominance of proprietary LLMs has led to restricted access and raised information privacy concerns. High-performing open-source alternatives are crucial for information-sensitive and high-volume applications but often lag behind in performance. To address this gap, we propose (1) A untargeted variant of iterative self-critique and self-refinement devoid of external influence. (2) A novel ranking metric - Performance, Refinement, and Inference Cost Score (PeRFICS) - to find the optimal model for a given task considering refined performance and cost. Our experiments show that SoTA open source models of varying sizes from 7B - 65B, on average, improve 8.2% from their baseline performance. Strikingly, even models with extremely small memory footprints, such as Vicuna-7B, show a 11.74% improvement overall and up to a 25.39% improvement in high-creativity, open ended tasks on the Vicuna benchmark. Vicuna-13B takes it a step further and outperforms ChatGPT post-refinement. This work has profound implications for resource-constrained and information-sensitive environments seeking to leverage LLMs without incurring prohibitive costs, compromising on performance and privacy. The domain-agnostic self-refinement process coupled with our novel ranking metric facilitates informed decision-making in model selection, thereby reducing costs and democratizing access to high-performing language models, as evidenced by case studies.

There has been considerable divergence of opinion on the reasoning abilities of Large Language Models (LLMs). While the initial optimism that reasoning might emerge automatically with scale has been tempered thanks to a slew of counterexamples, a wide spread belief in their iterative self-critique capabilities persists. In this paper, we set out to systematically investigate the effectiveness of iterative prompting of LLMs in the context of Graph Coloring, a canonical NP-complete reasoning problem that is related to propositional satisfiability as well as practical problems like scheduling and allocation. We present a principled empirical study of the performance of GPT4 in solving graph coloring instances or verifying the correctness of candidate colorings. In iterative modes, we experiment with the model critiquing its own answers and an external correct reasoner verifying proposed solutions. In both cases, we analyze whether the content of the criticisms actually affects bottom line performance. The study seems to indicate that (i) LLMs are bad at solving graph coloring instances (ii) they are no better at verifying a solution--and thus are not effective in iterative modes with LLMs critiquing LLM-generated solutions (iii) the correctness and content of the criticisms--whether by LLMs or external solvers--seems largely irrelevant to the performance of iterative prompting. We show that the observed increase in effectiveness is largely due to the correct solution being fortuitously present in the top-k completions of the prompt (and being recognized as such by an external verifier). Our results thus call into question claims about the self-critiquing capabilities of state of the art LLMs.

ARC Challenge appears more difficult than ARC Easy for modern LLMs primarily due to an evaluation setup that prevents direct comparison of answer choices rather than inherent complexity. Although some researchers have quietly shifted to a more appropriate scheme over the last year, the implications of this change have yet to be widely acknowledged. We highlight this overlooked shift, show how similar evaluation practices falsely imply reasoning deficits in other benchmarks, and demonstrate that fairer methods dramatically reduce performance gaps (e.g. on SIQA) and even yield superhuman results (OpenBookQA). In doing so, we reveal how evaluation shapes perceived difficulty and offer guidelines to ensure that multiple-choice evaluations accurately reflect actual model capabilities.

Recent advances in large language models (LLMs) have led to the development of various evaluation benchmarks. These benchmarks typically rely on a single instruction template for evaluating all LLMs on a specific task. In this paper, we comprehensively analyze the brittleness of results obtained via single-prompt evaluations across 6.5M instances, involving 20 different LLMs and 39 tasks from 3 benchmarks. To improve robustness of the analysis, we propose to evaluate LLMs with a set of diverse prompts instead. We discuss tailored evaluation metrics for specific use cases (e.g., LLM developers vs. developers interested in a specific downstream task), ensuring a more reliable and meaningful assessment of LLM capabilities. We then implement these criteria and conduct evaluations of multiple models, providing insights into the true strengths and limitations of current LLMs.

For humans to trust the fluent generations of large language models (LLMs), they must be able to verify their correctness against trusted, external sources. Recent efforts aim to increase verifiability through citations of retrieved documents or post-hoc provenance. However, such citations are prone to mistakes that further complicate their verifiability. To address these limitations, we tackle the verifiability goal with a different philosophy: we trivialize the verification process by developing models that quote verbatim statements from trusted sources in pre-training data. We propose Quote-Tuning, which demonstrates the feasibility of aligning LLMs to leverage memorized information and quote from pre-training data. Quote-Tuning quantifies quoting against large corpora with efficient membership inference tools, and uses the amount of quotes as an implicit reward signal to construct a synthetic preference dataset for quoting, without any human annotation. Next, the target model is aligned to quote using preference optimization algorithms. Experimental results show that Quote-Tuning significantly increases the percentage of LLM generation quoted verbatim from high-quality pre-training documents by 55% to 130% relative to untuned models while maintaining response quality. Further experiments demonstrate that Quote-Tuning generalizes quoting to out-of-domain data, is applicable in different tasks, and provides additional benefits to truthfulness. Quote-Tuning not only serves as a hassle-free method to increase quoting but also opens up avenues for improving LLM trustworthiness through better verifiability.

State-of-the-art large language models (LLMs) have demonstrated impressive code generation capabilities but struggle with real-world software engineering tasks, such as revising source code to address code reviews, hindering their practical use. Code review comments are often implicit, ambiguous, and colloquial, requiring models to grasp both code and human intent. This challenge calls for evaluating large language models' ability to bridge both technical and conversational contexts. While existing work has employed the automated code refinement (ACR) task to resolve these comments, current evaluation methods fall short, relying on text matching metrics that provide limited insight into model failures and remain susceptible to training data contamination. To address these limitations, we introduce a novel evaluation benchmark, CodeReviewQA that enables us to conduct fine-grained assessment of model capabilities and mitigate data contamination risks. In CodeReviewQA, we decompose the generation task of code refinement into three essential reasoning steps: change type recognition (CTR), change localisation (CL), and solution identification (SI). Each step is reformulated as multiple-choice questions with varied difficulty levels, enabling precise assessment of model capabilities, while mitigating data contamination risks. Our comprehensive evaluation spans 72 recently released large language models on 900 manually curated, high-quality examples across nine programming languages. Our results show that CodeReviewQA is able to expose specific model weaknesses in code review comprehension, disentangled from their generative automated code refinement results.

Peer assessment has been widely applied across diverse academic fields over the last few decades and has demonstrated its effectiveness. However, the advantages of peer assessment can only be achieved with high-quality peer reviews. Previous studies have found that high-quality review comments usually comprise several features (e.g., contain suggestions, mention problems, use a positive tone). Thus, researchers have attempted to evaluate peer-review comments by detecting different features using various machine learning and deep learning models. However, there is no single study that investigates using a multi-task learning (MTL) model to detect multiple features simultaneously. This paper presents two MTL models for evaluating peer-review comments by leveraging the state-of-the-art pre-trained language representation models BERT and DistilBERT. Our results demonstrate that BERT-based models significantly outperform previous GloVe-based methods by around 6% in F1-score on tasks of detecting a single feature, and MTL further improves performance while reducing model size.

Recently, o1-like models have drawn significant attention, where these models produce the long Chain-of-Thought (CoT) reasoning steps to improve the reasoning abilities of existing Large Language Models (LLMs). In this paper, to understand the qualities of these long CoTs and measure the critique abilities of existing LLMs on these long CoTs, we introduce the DeltaBench, including the generated long CoTs from different o1-like models (e.g., QwQ, DeepSeek-R1) for different reasoning tasks (e.g., Math, Code, General Reasoning), to measure the ability to detect errors in long CoT reasoning. Based on DeltaBench, we first perform fine-grained analysis of the generated long CoTs to discover the effectiveness and efficiency of different o1-like models. Then, we conduct extensive evaluations of existing process reward models (PRMs) and critic models to detect the errors of each annotated process, which aims to investigate the boundaries and limitations of existing PRMs and critic models. Finally, we hope that DeltaBench could guide developers to better understand the long CoT reasoning abilities of their models.

For many evaluation metrics commonly used as benchmarks for unconditional image generation, trivially memorizing the training set attains a better score than models which are considered state-of-the-art; we consider this problematic. We clarify a necessary condition for an evaluation metric not to behave this way: estimating the function must require a large sample from the model. In search of such a metric, we turn to neural network divergences (NNDs), which are defined in terms of a neural network trained to distinguish between distributions. The resulting benchmarks cannot be "won" by training set memorization, while still being perceptually correlated and computable only from samples. We survey past work on using NNDs for evaluation and implement an example black-box metric based on these ideas. Through experimental validation we show that it can effectively measure diversity, sample quality, and generalization.

Large Language Models (LLMs) are increasingly relied upon to evaluate text outputs of other LLMs, thereby influencing leaderboards and development decisions. However, concerns persist over the accuracy of these assessments and the potential for misleading conclusions. In this work, we investigate the effectiveness of LLMs as evaluators for text generation tasks. We propose FBI, a novel framework designed to examine the proficiency of Evaluator LLMs in assessing four critical abilities in other LLMs: factual accuracy, instruction following, coherence in long-form writing, and reasoning proficiency. By introducing targeted perturbations in answers generated by LLMs, that clearly impact one of these key capabilities, we test whether an Evaluator LLM can detect these quality drops. By creating a total of 2400 perturbed answers covering 22 perturbation categories, we conduct a comprehensive study using different evaluation strategies on five prominent LLMs commonly used as evaluators in the literature. Our findings reveal significant shortcomings in current Evaluator LLMs, which failed to identify quality drops in over 50\% of cases on average. Single-answer and pairwise evaluations demonstrated notable limitations, whereas reference-based evaluations showed comparatively better performance. These results underscore the unreliable nature of current Evaluator LLMs and advocate for cautious implementation in practical applications. Code and data are available at https://github.com/AI4Bharat/FBI.

Fake news and misinformation poses a significant threat to society, making efficient mitigation essential. However, manual fact-checking is costly and lacks scalability. Large Language Models (LLMs) offer promise in automating counter-response generation to mitigate misinformation, but a critical challenge lies in their tendency to hallucinate non-factual information. Existing models mainly rely on LLM self-feedback to reduce hallucination, but this approach is computationally expensive. In this paper, we propose MisMitiFact, Misinformation Mitigation grounded in Facts, an efficient framework for generating fact-grounded counter-responses at scale. MisMitiFact generates simple critique feedback to refine LLM outputs, ensuring responses are grounded in evidence. We develop lightweight, fine-grained critique models trained on data sourced from readily available fact-checking sites to identify and correct errors in key elements such as numerals, entities, and topics in LLM generations. Experiments show that MisMitiFact generates counter-responses of comparable quality to LLMs' self-feedback while using significantly smaller critique models. Importantly, it achieves ~5x increase in feedback generation throughput, making it highly suitable for cost-effective, large-scale misinformation mitigation. Code and LLM prompt templates are at https://github.com/xxfwin/MisMitiFact.

Large language models (LLMs) struggle with multi-step reasoning, where inference-time scaling has emerged as a promising strategy for performance improvement. Verifier-guided search outperforms repeated sampling when sample size is limited by selecting and prioritizing valid reasoning paths. However, we identify a critical limitation: scaling flaws, prevalent across different models (Mistral 7B and DeepSeekMath 7B), benchmarks (GSM8K and MATH), and verifiers (outcome value models and process reward models). As sample size increases, verifier-guided search exhibits diminishing advantages and eventually underperforms repeated sampling. Our analysis attributes this to verifier failures, where imperfect verifiers misrank candidates and erroneously prune all valid paths. These issues are further exacerbated in challenging and out-of-distribution problems, restricting search effectiveness. To mitigate verifier failures, we explore reducing reliance on verifiers and conduct preliminary investigations using two simple methods. Our findings reveal fundamental limitations in verifier-guided search and suggest future directions.

Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the Open LLM Leaderboard aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly). However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed. We use data from n > 5000 LLMs to identify the most informative items of six benchmarks, ARC, GSM8K, HellaSwag, MMLU, TruthfulQA and WinoGrande (with d=28,632 items in total). From them we distill a sparse benchmark, metabench, that has less than 3% of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities. We show that these estimators (1) can be used to reconstruct each original individual benchmark score with, on average, 1.5% root mean square error (RMSE), (2) reconstruct the original total score with 0.8% RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is r = 0.93.

We introduce SciVer, the first benchmark specifically designed to evaluate the ability of foundation models to verify claims within a multimodal scientific context. SciVer consists of 3,000 expert-annotated examples over 1,113 scientific papers, covering four subsets, each representing a common reasoning type in multimodal scientific claim verification. To enable fine-grained evaluation, each example includes expert-annotated supporting evidence. We assess the performance of 21 state-of-the-art multimodal foundation models, including o4-mini, Gemini-2.5-Flash, Llama-3.2-Vision, and Qwen2.5-VL. Our experiment reveals a substantial performance gap between these models and human experts on SciVer. Through an in-depth analysis of retrieval-augmented generation (RAG), and human-conducted error evaluations, we identify critical limitations in current open-source models, offering key insights to advance models' comprehension and reasoning in multimodal scientific literature tasks.

Scalable oversight studies methods of training and evaluating AI systems in domains where human judgment is unreliable or expensive, such as scientific research and software engineering in complex codebases. Most work in this area has focused on methods of improving the quality of labels. Recent work by Burns et al. (2023) considers the complementary problem of training models with low-quality labels, finding that large pretrained models often have an inductive bias towards producing correct answers. In practice, however, neither label quantity nor quality is fixed: practitioners face a quantity-quality tradeoff. In this paper, we explore the microeconomics of the quantity-quality tradeoff on binary NLP classification tasks used in Burns et al. (2023). While sample-efficient learning has been studied extensively, little public research has focused on scalable elicitation: eliciting capabilities from pretrained models subject to labeling cost constraints. We find that this setting has novel dynamics caused by the tradeoff between label quantity and quality, as well as the model's existing latent capabilities. We observe three regimes of eliciting classification knowledge from pretrained models using supervised finetuning: quantity-dominant, quality-dominant, and a mixed regime involving the use of low- and high-quality data together to attain higher accuracy at a lower cost than using either alone. We explore sample-efficient elicitation methods that make use of two datasets of differing qualities, and establish a Pareto frontier of scalable elicitation methods that optimally trade off labeling cost and classifier performance. We find that the accuracy of supervised fine-tuning can be improved by up to 5 percentage points at a fixed labeling budget by adding a few-shot prompt to make use of the model's existing knowledge of the task.

This paper investigates the rational thinking capability of Large Language Models (LLMs) in multi-round argumentative debates by exploring the impact of fallacious arguments on their logical reasoning performance. More specifically, we present Logic Competence Measurement Benchmark (LOGICOM), a diagnostic benchmark to assess the robustness of LLMs against logical fallacies. LOGICOM involves two agents: a persuader and a debater engaging in a multi-round debate on a controversial topic, where the persuader tries to convince the debater of the correctness of its claim. First, LOGICOM assesses the potential of LLMs to change their opinions through reasoning. Then, it evaluates the debater's performance in logical reasoning by contrasting the scenario where the persuader employs logical fallacies against one where logical reasoning is used. We use this benchmark to evaluate the performance of GPT-3.5 and GPT-4 using a dataset containing controversial topics, claims, and reasons supporting them. Our findings indicate that both GPT-3.5 and GPT-4 can adjust their opinion through reasoning. However, when presented with logical fallacies, GPT-3.5 and GPT-4 are erroneously convinced 41% and 69% more often, respectively, compared to when logical reasoning is used. Finally, we introduce a new dataset containing over 5k pairs of logical vs. fallacious arguments. The source code and dataset of this work are made publicly available.

The global economy is increasingly dependent on knowledge workers to meet the needs of public and private organizations. While there is no single definition of knowledge work, organizations and industry groups still attempt to measure individuals' capability to engage in it. The most comprehensive assessment of capability readiness for professional knowledge workers is the Uniform CPA Examination developed by the American Institute of Certified Public Accountants (AICPA). In this paper, we experimentally evaluate OpenAI's `text-davinci-003` and prior versions of GPT on both a sample Regulation (REG) exam and an assessment of over 200 multiple-choice questions based on the AICPA Blueprints for legal, financial, accounting, technology, and ethical tasks. First, we find that `text-davinci-003` achieves a correct rate of 14.4% on a sample REG exam section, significantly underperforming human capabilities on quantitative reasoning in zero-shot prompts. Second, `text-davinci-003` appears to be approaching human-level performance on the Remembering & Understanding and Application skill levels in the Exam absent calculation. For best prompt and parameters, the model answers 57.6% of questions correctly, significantly better than the 25% guessing rate, and its top two answers are correct 82.1% of the time, indicating strong non-entailment. Finally, we find that recent generations of GPT-3 demonstrate material improvements on this assessment, rising from 30% for `text-davinci-001` to 57% for `text-davinci-003`. These findings strongly suggest that large language models have the potential to transform the quality and efficiency of future knowledge work.

Test-time scaling is a promising new approach to language modeling that uses extra test-time compute to improve performance. Recently, OpenAI's o1 model showed this capability but did not publicly share its methodology, leading to many replication efforts. We seek the simplest approach to achieve test-time scaling and strong reasoning performance. First, we curate a small dataset s1K of 1,000 questions paired with reasoning traces relying on three criteria we validate through ablations: difficulty, diversity, and quality. Second, we develop budget forcing to control test-time compute by forcefully terminating the model's thinking process or lengthening it by appending "Wait" multiple times to the model's generation when it tries to end. This can lead the model to double-check its answer, often fixing incorrect reasoning steps. After supervised finetuning the Qwen2.5-32B-Instruct language model on s1K and equipping it with budget forcing, our model s1 exceeds o1-preview on competition math questions by up to 27% (MATH and AIME24). Further, scaling s1 with budget forcing allows extrapolating beyond its performance without test-time intervention: from 50% to 57% on AIME24. Our model, data, and code are open-source at https://github.com/simplescaling/s1.

Large Language Model (LLM) integrations into applications like Microsoft365 suite and Google Workspace for creating/processing documents, emails, presentations, etc. has led to considerable enhancements in productivity and time savings. But as these integrations become more more complex, it is paramount to ensure that the quality of output from the LLM-integrated applications are relevant and appropriate for use. Identifying the need to develop robust evaluation approaches for natural language generation, wherein references/ground labels doesn't exist or isn't amply available, this paper introduces a novel framework called "SAGEval" which utilizes a critiquing Agent to provide feedback on scores generated by LLM evaluators. We show that the critiquing Agent is able to rectify scores from LLM evaluators, in absence of references/ground-truth labels, thereby reducing the need for labeled data even for complex NLG evaluation scenarios, like the generation of JSON-structured forms/surveys with responses in different styles like multiple choice, likert ratings, single choice questions, etc.

While text-to-image (T2I) generative models have become ubiquitous, they do not necessarily generate images that align with a given prompt. While previous work has evaluated T2I alignment by proposing metrics, benchmarks, and templates for collecting human judgements, the quality of these components is not systematically measured. Human-rated prompt sets are generally small and the reliability of the ratings -- and thereby the prompt set used to compare models -- is not evaluated. We address this gap by performing an extensive study evaluating auto-eval metrics and human templates. We provide three main contributions: (1) We introduce a comprehensive skills-based benchmark that can discriminate models across different human templates. This skills-based benchmark categorises prompts into sub-skills, allowing a practitioner to pinpoint not only which skills are challenging, but at what level of complexity a skill becomes challenging. (2) We gather human ratings across four templates and four T2I models for a total of >100K annotations. This allows us to understand where differences arise due to inherent ambiguity in the prompt and where they arise due to differences in metric and model quality. (3) Finally, we introduce a new QA-based auto-eval metric that is better correlated with human ratings than existing metrics for our new dataset, across different human templates, and on TIFA160.

Large language models (LLMs) have made impressive progress in natural language processing. These models rely on proper human instructions (or prompts) to generate suitable responses. However, the potential of LLMs are not fully harnessed by commonly-used prompting methods: many human-in-the-loop algorithms employ ad-hoc procedures for prompt selection; while auto prompt generation approaches are essentially searching all possible prompts randomly and inefficiently. We propose Evoke, an automatic prompt refinement framework. In Evoke, there are two instances of a same LLM: one as a reviewer (LLM-Reviewer), it scores the current prompt; the other as an author (LLM-Author), it edits the prompt by considering the edit history and the reviewer's feedback. Such an author-reviewer feedback loop ensures that the prompt is refined in each iteration. We further aggregate a data selection approach to Evoke, where only the hard samples are exposed to the LLM. The hard samples are more important because the LLM can develop deeper understanding of the tasks out of them, while the model may already know how to solve the easier cases. Experimental results show that Evoke significantly outperforms existing methods. For instance, in the challenging task of logical fallacy detection, Evoke scores above 80, while all other baseline methods struggle to reach 20.

The scientific scale-up of large language models (LLMs) necessitates a comprehensive understanding of their scaling properties. However, the existing literature on the scaling properties only yields an incomplete answer: optimization loss decreases predictably as the model size increases, in line with established scaling law; yet no scaling law for task has been established and the task performances are far from predictable during scaling. Task performances typically show minor gains on small models until they improve dramatically once models exceed a size threshold, exemplifying the ``emergent abilities''. In this study, we discover that small models, although they exhibit minor performance, demonstrate critical and consistent task performance improvements that are not captured by conventional evaluation strategies due to insufficient measurement resolution. To measure such improvements, we introduce PassUntil, an evaluation strategy through massive sampling in the decoding phase. We conduct quantitative investigations into the scaling law of task performance. Firstly, a strict task scaling law is identified, enhancing the predictability of task performances. Remarkably, we are able to predict the performance of the 2.4B model on code generation with merely 0.05\% deviation before training starts. Secondly, underpinned by PassUntil, we observe concrete evidence of emergent abilities and ascertain that they are not in conflict with the continuity of performance improvement. Their semblance to break-through is that their scaling curve cannot be fitted by standard scaling law function. We then introduce a mathematical definition for the emergent abilities. Through the definition, we refute a prevalent ``multi-step reasoning hypothesis'' regarding the genesis of emergent abilities and propose a new hypothesis with a satisfying fit to the observed scaling curve.

When large language models (LMs) are applied in zero- or few-shot settings to discriminative tasks such as multiple-choice questions, their attentiveness (i.e., probability mass) is spread across many vocabulary tokens that are not valid choices. Such a spread across multiple surface forms with identical meaning is thought to cause an underestimation of a model's true performance, referred to as the "surface form competition" (SFC) hypothesis. This has motivated the introduction of various probability normalization methods. However, many core questions remain unanswered. How do we measure SFC or attentiveness? Are there direct ways of increasing attentiveness on valid choices? Does increasing attentiveness always improve task accuracy? We propose a mathematical formalism for studying this phenomenon, provide a metric for quantifying attentiveness, and identify a simple method for increasing it -- namely, in-context learning with even just one example containing answer choices. The formalism allows us to quantify SFC and bound its impact. Our experiments on three diverse datasets and six LMs reveal several surprising findings. For example, encouraging models to generate a valid answer choice can, in fact, be detrimental to task performance for some LMs, and prior probability normalization methods are less effective (sometimes even detrimental) to instruction-tuned LMs. We conclude with practical insights for effectively using prompted LMs for multiple-choice tasks.

Scaling law principles indicate a power-law correlation between loss and variables such as model size, dataset size, and computational resources utilized during training. These principles play a vital role in optimizing various aspects of model pre-training, ultimately contributing to the success of large language models such as GPT-4, Llama and Gemini. However, the original scaling law paper by OpenAI did not disclose the complete details necessary to derive the precise scaling law formulas, and their conclusions are only based on models containing up to 1.5 billion parameters. Though some subsequent works attempt to unveil these details and scale to larger models, they often neglect the training dependency of important factors such as the learning rate, context length and batch size, leading to their failure to establish a reliable formula for predicting the test loss trajectory. In this technical report, we confirm that the scaling law formulations proposed in the original OpenAI paper remain valid when scaling the model size up to 33 billion, but the constant coefficients in these formulas vary significantly with the experiment setup. We meticulously identify influential factors and provide transparent, step-by-step instructions to estimate all constant terms in scaling-law formulas by training on models with only 1M~60M parameters. Using these estimated formulas, we showcase the capability to accurately predict various attributes for models with up to 33B parameters before their training, including (1) the minimum possible test loss; (2) the minimum required training steps and processed tokens to achieve a specific loss; (3) the critical batch size with an optimal time/computation trade-off at any loss value; and (4) the complete test loss trajectory with arbitrary batch size.

Self-improvement is a mechanism in Large Language Model (LLM) pre-training, post-training and test-time inference. We explore a framework where the model verifies its own outputs, filters or reweights data based on this verification, and distills the filtered data. Despite several empirical successes, a fundamental understanding is still lacking. In this work, we initiate a comprehensive, modular and controlled study on LLM self-improvement. We provide a mathematical formulation for self-improvement, which is largely governed by a quantity which we formalize as the generation-verification gap. Through experiments with various model families and tasks, we discover a scaling phenomenon of self-improvement -- a variant of the generation-verification gap scales monotonically with the model pre-training flops. We also examine when self-improvement is possible, an iterative self-improvement procedure, and ways to improve its performance. Our findings not only advance understanding of LLM self-improvement with practical implications, but also open numerous avenues for future research into its capabilities and boundaries.

Evaluating the truthfulness of online content is critical for combating misinformation. This study examines the efficiency and effectiveness of crowdsourced truthfulness assessments through a comparative analysis of two approaches: one involving full-length webpages as evidence for each claim, and another using summaries for each evidence document generated with a large language model. Using an A/B testing setting, we engage a diverse pool of participants tasked with evaluating the truthfulness of statements under these conditions. Our analysis explores both the quality of assessments and the behavioral patterns of participants. The results reveal that relying on summarized evidence offers comparable accuracy and error metrics to the Standard modality while significantly improving efficiency. Workers in the Summary setting complete a significantly higher number of assessments, reducing task duration and costs. Additionally, the Summary modality maximizes internal agreement and maintains consistent reliance on and perceived usefulness of evidence, demonstrating its potential to streamline large-scale truthfulness evaluations.

Recent advancements in large language models (LLM) capable of processing extremely long texts highlight the need for a dedicated evaluation benchmark to assess their long-context capabilities. However, existing methods, like the needle-in-a-haystack test, do not effectively assess whether these models fully utilize contextual information, raising concerns about the reliability of current evaluation techniques. To thoroughly examine the effectiveness of existing benchmarks, we introduce a new metric called information coverage (IC), which quantifies the proportion of the input context necessary for answering queries. Our findings indicate that current benchmarks exhibit low IC; although the input context may be extensive, the actual usable context is often limited. To address this, we present ETHIC, a novel benchmark designed to assess LLMs' ability to leverage the entire context. Our benchmark comprises 2,648 test instances spanning four long-context tasks with high IC scores in the domains of books, debates, medicine, and law. Our evaluations reveal significant performance drops in contemporary LLMs, highlighting a critical challenge in managing long contexts. Our benchmark is available at https://github.com/dmis-lab/ETHIC.

This study explores the proactive ability of LLMs to seek user support. We propose metrics to evaluate the trade-off between performance improvements and user burden, and investigate whether LLMs can determine when to request help under varying information availability. Our experiments show that without external feedback, many LLMs struggle to recognize their need for user support. The findings highlight the importance of external signals and provide insights for future research on improving support-seeking strategies. Source code: https://github.com/appier-research/i-need-help

Recent studies have shown LLMs possess some ability to improve their responses when given external feedback. However, it remains unclear how effectively and thoroughly these models can incorporate extrinsic feedback. In an ideal scenario, if LLMs receive near-perfect and complete feedback, we would expect them to fully integrate the feedback and change their incorrect answers to correct ones. In this paper, we systematically investigate LLMs' ability to incorporate feedback by designing a controlled experimental environment. For each problem, a solver model attempts a solution, then a feedback generator with access to near-complete ground-truth answers produces targeted feedback, after which the solver tries again. We evaluate this pipeline across a diverse range of tasks, including math reasoning, knowledge reasoning, scientific reasoning, and general multi-domain evaluations with state-of-the-art language models including Claude 3.7 (with and without extended thinking). Surprisingly, even under these near-ideal conditions, solver models consistently show resistance to feedback, a limitation that we term FEEDBACK FRICTION. To mitigate this limitation, we experiment with sampling-based strategies like progressive temperature increases and explicit rejection of previously attempted incorrect answers, which yield improvements but still fail to help models achieve target performance. We also perform a rigorous exploration of potential causes of FEEDBACK FRICTION, ruling out factors such as model overconfidence and data familiarity. We hope that highlighting this issue in LLMs and ruling out several apparent causes will help future research in self-improvement.

Scaling up language models has been shown to predictably improve performance and sample efficiency on a wide range of downstream tasks. This paper instead discusses an unpredictable phenomenon that we refer to as emergent abilities of large language models. We consider an ability to be emergent if it is not present in smaller models but is present in larger models. Thus, emergent abilities cannot be predicted simply by extrapolating the performance of smaller models. The existence of such emergence implies that additional scaling could further expand the range of capabilities of language models.

In recent years, evaluating the Theory of Mind (ToM) capabilities of large language models (LLMs) has received significant attention within the research community. As the field rapidly evolves, navigating the diverse approaches and methodologies has become increasingly complex. This systematic review synthesizes current efforts to assess LLMs' ability to perform ToM tasks, an essential aspect of human cognition involving the attribution of mental states to oneself and others. Despite notable advancements, the proficiency of LLMs in ToM remains a contentious issue. By categorizing benchmarks and tasks through a taxonomy rooted in cognitive science, this review critically examines evaluation techniques, prompting strategies, and the inherent limitations of LLMs in replicating human-like mental state reasoning. A recurring theme in the literature reveals that while LLMs demonstrate emerging competence in ToM tasks, significant gaps persist in their emulation of human cognitive abilities.

Large Multimodal Models (LMMs) have witnessed remarkable growth, showcasing formidable capabilities in handling intricate multimodal tasks with exceptional performance. Recent research has underscored the inclination of large language models to passively accept defective inputs, often resulting in futile reasoning on invalid prompts. However, the same critical question of whether LMMs can actively detect and scrutinize erroneous inputs still remains unexplored. To address this gap, we introduce the Input Scrutiny Ability Evaluation Framework (ISEval), which encompasses seven categories of flawed premises and three evaluation metrics. Our extensive evaluation of ten advanced LMMs has identified key findings. Most models struggle to actively detect flawed textual premises without guidance, which reflects a strong reliance on explicit prompts for premise error identification. Error type affects performance: models excel at identifying logical fallacies but struggle with surface-level linguistic errors and certain conditional flaws. Modality trust varies-Gemini 2.5 pro and Claude Sonnet 4 balance visual and textual info, while aya-vision-8b over-rely on text in conflicts. These insights underscore the urgent need to enhance LMMs' proactive verification of input validity and shed novel insights into mitigating the problem. The code is available at https://github.com/MLGroupJLU/LMM_ISEval.

Peer review is a cornerstone of quality control in scientific publishing. With the increasing workload, the unintended use of `quick' heuristics, referred to as lazy thinking, has emerged as a recurring issue compromising review quality. Automated methods to detect such heuristics can help improve the peer-reviewing process. However, there is limited NLP research on this issue, and no real-world dataset exists to support the development of detection tools. This work introduces LazyReview, a dataset of peer-review sentences annotated with fine-grained lazy thinking categories. Our analysis reveals that Large Language Models (LLMs) struggle to detect these instances in a zero-shot setting. However, instruction-based fine-tuning on our dataset significantly boosts performance by 10-20 performance points, highlighting the importance of high-quality training data. Furthermore, a controlled experiment demonstrates that reviews revised with lazy thinking feedback are more comprehensive and actionable than those written without such feedback. We will release our dataset and the enhanced guidelines that can be used to train junior reviewers in the community. (Code available here: https://github.com/UKPLab/arxiv2025-lazy-review)

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 442 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Large language models (LLMs) have demonstrated enhanced performance through the Thinking then Responding paradigm, where models generate internal thoughts before final responses (aka, System 2 thinking). However, existing research lacks a systematic understanding of the mechanisms underlying how thinking patterns affect performance across model sizes. In this work, we conduct a comprehensive analysis of the impact of various thinking types on model performance and introduce ThinkPatterns-21k, a curated dataset comprising 21k instruction-response pairs (QA) collected from existing instruction-following datasets with five thinking types. For each pair, we augment it with five distinct internal thinking patterns: one unstructured thinking (monologue) and four structured variants (decomposition, self-ask, self-debate and self-critic), while maintaining the same instruction and response. Through extensive evaluation across different model sizes (3B-32B parameters), we have two key findings: (1) smaller models (<30B parameters) can benefit from most of structured thinking patterns, while larger models (32B) with structured thinking like decomposition would degrade performance and (2) unstructured monologue demonstrates broad effectiveness across different model sizes. Finally, we released all of our datasets, checkpoints, training logs of diverse thinking patterns to reproducibility, aiming to facilitate further research in this direction.

Large Language Models (LLMs) have significantly advanced the state-of-the-art in various coding tasks. Beyond directly answering user queries, LLMs can also serve as judges, assessing and comparing the quality of responses generated by other models. Such an evaluation capability is crucial both for benchmarking different LLMs and for improving response quality through response ranking. However, despite the growing adoption of the LLM-as-a-Judge paradigm, its effectiveness in coding scenarios remains underexplored due to the absence of dedicated benchmarks. To address this gap, we introduce CodeJudgeBench, a benchmark explicitly designed to evaluate the performance of LLM-as-a-Judge models across three critical coding tasks: code generation, code repair, and unit test generation. Through comprehensive benchmarking of 26 LLM-as-a-Judge models, we find that recent thinking models significantly outperform non-thinking models on our carefully designed code judging tasks. Notably, even relatively small thinking models, such as Qwen3-8B, can outperform specially trained LLM-as-a-Judge models up to 70B in size. Nevertheless, all models still exhibit significant randomness in their judgment of coding tasks. For pairwise judging tasks, simply changing the order in which responses are presented can substantially impact accuracy. In addition, when judging code and unit tests written by different LLMs, LLM-as-a-Judge models also show variance in performance. This sensitivity raises concerns about the reliability and consistency of LLM-as-a-Judge in coding scenarios. Lastly, we study optimal prompting strategies for LLM-as-a-Judge. We find that using pair-wise comparison outperforms scalar point-wise judging. Furthermore, retaining comments and reasoning in the full, unprocessed LLM response leads to improved judge performance.

As creative writing tasks do not have singular correct answers, large language models (LLMs) trained to perform these tasks should be able to generate diverse valid outputs. However, LLM post-training often focuses on improving generation quality but neglects to facilitate output diversity. Hence, in creative writing generation, we investigate post-training approaches to promote both output diversity and quality. Our core idea is to include deviation -- the degree of difference between a training sample and all other samples with the same prompt -- in the training objective to facilitate learning from rare high-quality instances. By adopting our approach to direct preference optimization (DPO) and odds ratio preference optimization (ORPO), we demonstrate that we can promote the output diversity of trained models while minimally decreasing quality. Our best model with 8B parameters could achieve on-par diversity as a human-created dataset while having output quality similar to the best instruction-tuned models we examined, GPT-4o and DeepSeek-R1. We further validate our approaches with a human evaluation, an ablation, and a comparison to an existing diversification approach, DivPO.

As large language models (LLMs) are adopted as a fundamental component of language technologies, it is crucial to accurately characterize their performance. Because choices in prompt design can strongly influence model behavior, this design process is critical in effectively using any modern pre-trained generative language model. In this work, we focus on LLM sensitivity to a quintessential class of meaning-preserving design choices: prompt formatting. We find that several widely used open-source LLMs are extremely sensitive to subtle changes in prompt formatting in few-shot settings, with performance differences of up to 76 accuracy points when evaluated using LLaMA-2-13B. Sensitivity remains even when increasing model size, the number of few-shot examples, or performing instruction tuning. Our analysis suggests that work evaluating LLMs with prompting-based methods would benefit from reporting a range of performance across plausible prompt formats, instead of the currently-standard practice of reporting performance on a single format. We also show that format performance only weakly correlates between models, which puts into question the methodological validity of comparing models with an arbitrarily chosen, fixed prompt format. To facilitate systematic analysis we propose FormatSpread, an algorithm that rapidly evaluates a sampled set of plausible prompt formats for a given task, and reports the interval of expected performance without accessing model weights. Furthermore, we present a suite of analyses that characterize the nature of this sensitivity, including exploring the influence of particular atomic perturbations and the internal representation of particular formats.

Most large language models (LLMs) are trained once and never updated; thus, they lack the ability to dynamically adapt to our ever-changing world. In this work, we perform a detailed study of the factuality of LLM-generated text in the context of answering questions that test current world knowledge. Specifically, we introduce FreshQA, a novel dynamic QA benchmark encompassing a diverse range of question and answer types, including questions that require fast-changing world knowledge as well as questions with false premises that need to be debunked. We benchmark a diverse array of both closed and open-source LLMs under a two-mode evaluation procedure that allows us to measure both correctness and hallucination. Through human evaluations involving more than 50K judgments, we shed light on limitations of these models and demonstrate significant room for improvement: for instance, all models (regardless of model size) struggle on questions that involve fast-changing knowledge and false premises. Motivated by these results, we present FreshPrompt, a simple few-shot prompting method that substantially boosts the performance of an LLM on FreshQA by incorporating relevant and up-to-date information retrieved from a search engine into the prompt. Our experiments show that FreshPrompt outperforms both competing search engine-augmented prompting methods such as Self-Ask (Press et al., 2022) as well as commercial systems such as Perplexity.AI. Further analysis of FreshPrompt reveals that both the number of retrieved evidences and their order play a key role in influencing the correctness of LLM-generated answers. Additionally, instructing the LLM to generate concise and direct answers helps reduce hallucination compared to encouraging more verbose answers. To facilitate future work, we release FreshQA at github.com/freshllms/freshqa and commit to updating it at regular intervals.

Large language models (LLMs) can suggest missing elements from items listed in a prompt, which can be used for list completion or recommendations based on users' history. However, their performance degrades when presented with too many items, as they start to suggest items already included in the input list. This occurs at around 100 items for mid-2024 flagship LLMs. We evaluate this phenomenon on both synthetic problems (e.g., finding missing numbers in a given range of shuffled integers) and realistic movie recommendation scenarios. We refer to this issue as attention overflow, as preventing repetition requires attending to all items simultaneously. Although iterative loops can mitigate this problem, their costs increase with the repetition rate, affecting the language models' ability to derive novelty from lengthy inputs.

LLM-as-a-Judge is a framework that uses an LLM (large language model) to evaluate the quality of natural language text - typically text that is also generated by an LLM. This framework holds great promise due to its relative low-cost, ease of use, and strong correlations with human stylistic preferences. However, LLM Judges have been shown to exhibit biases that can distort their judgments. We evaluate how well LLM Judges can grade whether a given response to a conversational question is correct, an ability crucial to soundly estimating the overall response quality. To do so, we create and publicly release a human-annotated dataset with labels of correctness for 1,200 LLM responses. We source questions from a combination of existing datasets and a novel, challenging benchmark (BFF-Bench) created for this analysis. We demonstrate a strong connection between an LLM's ability to correctly answer a question and grade responses to that question. Although aggregate level statistics might imply a judge has high agreement with human annotators, it will struggle on the subset of questions it could not answer. To address this issue, we recommend a simple solution: provide the judge with a correct, human-written reference answer. We perform an in-depth analysis on how reference quality can affect the performance of an LLM Judge. We show that providing a weaker judge (e.g. Qwen 2.5 7B) with higher quality references reaches better agreement with human annotators than a stronger judge (e.g. GPT-4o) with synthetic references.

No two authors write alike. Personal flourishes invoked in written narratives, from lexicon to rhetorical devices, imply a particular author--what literary theorists label the implied or virtual author; distinct from the real author or narrator of a text. Early large language models trained on unfiltered training sets drawn from a variety of discordant sources yielded incoherent personalities, problematic for conversational tasks but proving useful for sampling literature from multiple perspectives. Successes in alignment research in recent years have allowed researchers to impose subjectively consistent personae on language models via instruction tuning and reinforcement learning from human feedback (RLHF), but whether aligned models retain the ability to model an arbitrary virtual author has received little scrutiny. By studying 4,374 stories sampled from three OpenAI language models, we show successive versions of GPT-3 suffer from increasing degrees of "mode collapse" whereby overfitting the model during alignment constrains it from generalizing over authorship: models suffering from mode collapse become unable to assume a multiplicity of perspectives. Our method and results are significant for researchers seeking to employ language models in sociological simulations.

Common methods for aligning large language models (LLMs) with desired behaviour heavily rely on human-labelled data. However, as models grow increasingly sophisticated, they will surpass human expertise, and the role of human evaluation will evolve into non-experts overseeing experts. In anticipation of this, we ask: can weaker models assess the correctness of stronger models? We investigate this question in an analogous setting, where stronger models (experts) possess the necessary information to answer questions and weaker models (non-experts) lack this information. The method we evaluate is debate, where two LLM experts each argue for a different answer, and a non-expert selects the answer. We find that debate consistently helps both non-expert models and humans answer questions, achieving 76% and 88% accuracy respectively (naive baselines obtain 48% and 60%). Furthermore, optimising expert debaters for persuasiveness in an unsupervised manner improves non-expert ability to identify the truth in debates. Our results provide encouraging empirical evidence for the viability of aligning models with debate in the absence of ground truth.

Large Language Models (LLMs) are being used more and more extensively for automated evaluation in various scenarios. Previous studies have attempted to fine-tune open-source LLMs to replicate the evaluation explanations and judgments of powerful proprietary models, such as GPT-4. However, these methods are largely limited to text-based analyses under predefined general criteria, resulting in reduced adaptability for unseen instructions and demonstrating instability in evaluating adherence to quantitative and structural constraints. To address these limitations, we propose a novel evaluation framework, ARJudge, that adaptively formulates evaluation criteria and synthesizes both text-based and code-driven analyses to evaluate LLM responses. ARJudge consists of two components: a fine-tuned Analyzer that generates multi-faceted evaluation analyses and a tuning-free Refiner that combines and refines all analyses to make the final judgment. We construct a Composite Analysis Corpus that integrates tasks for evaluation criteria generation alongside text-based and code-driven analysis generation to train the Analyzer. Our results demonstrate that ARJudge outperforms existing fine-tuned evaluators in effectiveness and robustness. Furthermore, it demonstrates the importance of multi-faceted evaluation and code-driven analyses in enhancing evaluation capabilities.

Large language models (LLMs) have led to a surge in collaborative writing with model assistance. As different users incorporate suggestions from the same model, there is a risk of decreased diversity in the produced content, potentially limiting diverse perspectives in public discourse. In this work, we measure the impact of co-writing on diversity via a controlled experiment, where users write argumentative essays in three setups -- using a base LLM (GPT3), a feedback-tuned LLM (InstructGPT), and writing without model help. We develop a set of diversity metrics and find that writing with InstructGPT (but not the GPT3) results in a statistically significant reduction in diversity. Specifically, it increases the similarity between the writings of different authors and reduces the overall lexical and content diversity. We additionally find that this effect is mainly attributable to InstructGPT contributing less diverse text to co-written essays. In contrast, the user-contributed text remains unaffected by model collaboration. This suggests that the recent improvement in generation quality from adapting models to human feedback might come at the cost of more homogeneous and less diverse content.

Recent large language models often answer factual questions correctly. But users can't trust any given claim a model makes without fact-checking, because language models can hallucinate convincing nonsense. In this work we use reinforcement learning from human preferences (RLHP) to train "open-book" QA models that generate answers whilst also citing specific evidence for their claims, which aids in the appraisal of correctness. Supporting evidence is drawn from multiple documents found via a search engine, or from a single user-provided document. Our 280 billion parameter model, GopherCite, is able to produce answers with high quality supporting evidence and abstain from answering when unsure. We measure the performance of GopherCite by conducting human evaluation of answers to questions in a subset of the NaturalQuestions and ELI5 datasets. The model's response is found to be high-quality 80\% of the time on this Natural Questions subset, and 67\% of the time on the ELI5 subset. Abstaining from the third of questions for which it is most unsure improves performance to 90\% and 80\% respectively, approaching human baselines. However, analysis on the adversarial TruthfulQA dataset shows why citation is only one part of an overall strategy for safety and trustworthiness: not all claims supported by evidence are true.

Large language models (LLMs) with enormous pre-training tokens and parameter amounts emerge abilities, including math reasoning, code generation, and instruction following. These abilities are further enhanced by supervised fine-tuning (SFT). The open-source community has studied on ad-hoc SFT for each ability, while proprietary LLMs are versatile for all abilities. It is important to investigate how to unlock them with multiple abilities via SFT. In this study, we specifically focus on the data composition between mathematical reasoning, code generation, and general human-aligning abilities during SFT. From a scaling perspective, we investigate the relationship between model abilities and various factors including data amounts, data composition ratio, model parameters, and SFT strategies. Our experiments reveal that different abilities exhibit different scaling patterns, and larger models generally show superior performance with the same amount of data. Mathematical reasoning and code generation improve as data amounts increase consistently, while the general ability is enhanced with about a thousand samples and improves slowly. We find data composition results in various abilities improvements with low data amounts, while conflicts of abilities with high data amounts. Our experiments further show that composition data amount impacts performance, while the influence of composition ratio is insignificant. Regarding the SFT strategies, we evaluate sequential learning multiple abilities are prone to catastrophic forgetting. Our proposed Dual-stage Mixed Fine-tuning (DMT) strategy learns specialized abilities first and then learns general abilities with a small amount of specialized data to prevent forgetting, offering a promising solution to learn multiple abilities with different scaling patterns.

Research on scaling large language models (LLMs) has primarily focused on model parameters and training data size, overlooking the role of vocabulary size. % Intuitively, larger vocabularies enable more efficient tokenization by representing sentences with fewer tokens, but they also increase the risk of under-fitting representations for rare tokens. We investigate how vocabulary size impacts LLM scaling laws by training models ranging from 33M to 3B parameters on up to 500B characters with various vocabulary configurations. We propose three complementary approaches for predicting the compute-optimal vocabulary size: IsoFLOPs analysis, derivative estimation, and parametric fit of the loss function. Our approaches converge on the same result that the optimal vocabulary size depends on the available compute budget and that larger models deserve larger vocabularies. However, most LLMs use too small vocabulary sizes. For example, we predict that the optimal vocabulary size of Llama2-70B should have been at least 216K, 7 times larger than its vocabulary of 32K. We validate our predictions empirically by training models with 3B parameters across different FLOPs budgets. Adopting our predicted optimal vocabulary size consistently improves downstream performance over commonly used vocabulary sizes. By increasing the vocabulary size from the conventional 32K to 43K, we improve performance on ARC-Challenge from 29.1 to 32.0 with the same 2.3e21 FLOPs. Our work emphasizes the necessity of jointly considering model parameters and vocabulary size for efficient scaling.

Language model (LM) distillation is a trending area that aims to distil the knowledge resided in a large teacher LM to a small student one. While various methods have been proposed to push the distillation to its limits, it is still a pain distilling LMs when a large capacity gap is exhibited between the teacher and the student LMs. The pain is mainly resulted by the curse of capacity gap, which describes that a larger teacher LM cannot always lead to a better student LM than one distilled from a smaller teacher LM due to the affect of capacity gap increment. That is, there is likely an optimal point yielding the best student LM along the scaling course of the teacher LM. Even worse, the curse of capacity gap can be only partly yet not fully lifted as indicated in previous studies. However, the tale is not ever one-sided. Although a larger teacher LM has better performance than a smaller teacher LM, it is much more resource-demanding especially in the context of recent large LMs (LLMs). Consequently, instead of sticking to lifting the curse, leaving the curse as is should be arguably fine. Even better, in this paper, we reveal that the optimal capacity gap is almost consistent across different student scales and architectures, fortunately turning the curse into the law of capacity gap. The law later guides us to distil a 3B student LM (termed MiniMA) from a 7B teacher LM (adapted LLaMA2-7B). MiniMA is demonstrated to yield a new compute-performance pareto frontier among existing 3B LMs on commonly used benchmarks, and its instruction-tuned version (termed MiniChat) outperforms a wide range of 3B competitors in GPT4 evaluation and could even compete with several 7B chat models.

Agentic workflows -- where multiple large language model (LLM) instances interact to solve tasks -- are increasingly built on feedback mechanisms, where one model evaluates and critiques another. Despite the promise of feedback-driven improvement, the stability of agentic workflows rests on the reliability of the judge. However, judges may hallucinate information, exhibit bias, or act adversarially -- introducing critical vulnerabilities into the workflow. In this work, we present a systematic analysis of agentic workflows under deceptive or misleading feedback. We introduce a two-dimensional framework for analyzing judge behavior, along axes of intent (from constructive to malicious) and knowledge (from parametric-only to retrieval-augmented systems). Using this taxonomy, we construct a suite of judge behaviors and develop WAFER-QA, a new benchmark with critiques grounded in retrieved web evidence to evaluate robustness of agentic workflows against factually supported adversarial feedback. We reveal that even strongest agents are vulnerable to persuasive yet flawed critiques -- often switching correct answers after a single round of misleading feedback. Taking a step further, we study how model predictions evolve over multiple rounds of interaction, revealing distinct behavioral patterns between reasoning and non-reasoning models. Our findings highlight fundamental vulnerabilities in feedback-based workflows and offer guidance for building more robust agentic systems.

Style is an important concept in today's challenges in natural language generating. After the success in the field of image style transfer, the task of text style transfer became actual and attractive. Researchers are also interested in the tasks of style reproducing in generation of the poetic text. Evaluation of style reproducing in natural poetry generation remains a problem. I used 3 character-based LSTM-models to work with style reproducing assessment. All three models were trained on the corpus of texts by famous Russian-speaking poets. Samples were shown to the assessors and 4 answer options were offered, the style of which poet this sample reproduces. In addition, the assessors were asked how well they were familiar with the work of the poet they had named. Students studying history of literature were the assessors, 94 answers were received. It has appeared that accuracy of definition of style increases if the assessor can quote the poet by heart. Each model showed at least 0.7 macro-average accuracy. The experiment showed that it is better to involve a professional rather than a naive reader in the evaluation of style in the tasks of poetry generation, while lstm models are good at reproducing the style of Russian poets even on a limited training corpus.

This paper investigates the fragility of Large Language Models (LLMs) in generalizing to novel inputs, specifically focusing on minor perturbations in well-established benchmarks (e.g., slight changes in question format or distractor length). Despite high benchmark scores, LLMs exhibit significant accuracy drops and unexpected biases (e.g., preference for longer distractors) when faced with these minor but content-preserving modifications. For example, Qwen 2.5 1.5B's MMLU score rises from 60 to 89 and drops from 89 to 36 when option lengths are changed without altering the question. Even GPT-4 experiences a 25-point accuracy loss when question types are changed, with a 6-point drop across all three modification categories. These analyses suggest that LLMs rely heavily on superficial cues rather than forming robust, abstract representations that generalize across formats, lexical variations, and irrelevant content shifts. This work aligns with the ACL 2025 theme track on the Generalization of NLP models, proposing a "Generalization Stress Test" to assess performance shifts under controlled perturbations. The study calls for reevaluating benchmarks and developing more reliable evaluation methodologies to capture LLM generalization abilities better.

As LLMs grow more powerful, their most profound achievement may be recognising when to say "I don't know". Existing studies on LLM self-knowledge have been largely constrained by human-defined notions of feasibility, often neglecting the reasons behind unanswerability by LLMs and failing to study deficient types of self-knowledge. This study aims to obtain intrinsic insights into different types of LLM self-knowledge with a novel methodology: allowing them the flexibility to set their own feasibility boundaries and then analysing the consistency of these limits. We find that even frontier models like GPT-4o and Mistral Large are not sure of their own capabilities more than 80% of the time, highlighting a significant lack of trustworthiness in responses. Our analysis of confidence balance in LLMs indicates that models swing between overconfidence and conservatism in feasibility boundaries depending on task categories and that the most significant self-knowledge weaknesses lie in temporal awareness and contextual understanding. These difficulties in contextual comprehension additionally lead models to question their operational boundaries, resulting in considerable confusion within the self-knowledge of LLMs. We make our code and results available publicly at https://github.com/knowledge-verse-ai/LLM-Self_Knowledge_Eval

Peer review, as a cornerstone of scientific research, ensures the integrity and quality of scholarly work by providing authors with objective feedback for refinement. However, in the traditional peer review process, authors often receive vague or insufficiently detailed feedback, which provides limited assistance and leads to a more time-consuming review cycle. If authors can identify some specific weaknesses in their paper, they can not only address the reviewer's concerns but also improve their work. This raises the critical question of how to enhance authors' comprehension of review comments. In this paper, we present SEAGraph, a novel framework developed to clarify review comments by uncovering the underlying intentions behind them. We construct two types of graphs for each paper: the semantic mind graph, which captures the author's thought process, and the hierarchical background graph, which delineates the research domains related to the paper. A retrieval method is then designed to extract relevant content from both graphs, facilitating coherent explanations for the review comments. Extensive experiments show that SEAGraph excels in review comment understanding tasks, offering significant benefits to authors.

We present OpenReviewer, an open-source system for generating high-quality peer reviews of machine learning and AI conference papers. At its core is Llama-OpenReviewer-8B, an 8B parameter language model specifically fine-tuned on 79,000 expert reviews from top conferences. Given a PDF paper submission and review template as input, OpenReviewer extracts the full text, including technical content like equations and tables, and generates a structured review following conference-specific guidelines. Our evaluation on 400 test papers shows that OpenReviewer produces considerably more critical and realistic reviews compared to general-purpose LLMs like GPT-4 and Claude-3.5. While other LLMs tend toward overly positive assessments, OpenReviewer's recommendations closely match the distribution of human reviewer ratings. The system provides authors with rapid, constructive feedback to improve their manuscripts before submission, though it is not intended to replace human peer review. OpenReviewer is available as an online demo and open-source tool.

Can LLMs accurately adjust their confidence when facing opposition? Building on previous studies measuring calibration on static fact-based question-answering tasks, we evaluate Large Language Models (LLMs) in a dynamic, adversarial debate setting, uniquely combining two realistic factors: (a) a multi-turn format requiring models to update beliefs as new information emerges, and (b) a zero-sum structure to control for task-related uncertainty, since mutual high-confidence claims imply systematic overconfidence. We organized 60 three-round policy debates among ten state-of-the-art LLMs, with models privately rating their confidence (0-100) in winning after each round. We observed five concerning patterns: (1) Systematic overconfidence: models began debates with average initial confidence of 72.9% vs. a rational 50% baseline. (2) Confidence escalation: rather than reducing confidence as debates progressed, debaters increased their win probabilities, averaging 83% by the final round. (3) Mutual overestimation: in 61.7% of debates, both sides simultaneously claimed >=75% probability of victory, a logical impossibility. (4) Persistent self-debate bias: models debating identical copies increased confidence from 64.1% to 75.2%; even when explicitly informed their chance of winning was exactly 50%, confidence still rose (from 50.0% to 57.1%). (5) Misaligned private reasoning: models' private scratchpad thoughts sometimes differed from their public confidence ratings, raising concerns about faithfulness of chain-of-thought reasoning. These results suggest LLMs lack the ability to accurately self-assess or update their beliefs in dynamic, multi-turn tasks; a major concern as LLMs are now increasingly deployed without careful review in assistant and agentic roles. Code for our experiments is available at https://github.com/pradyuprasad/llms_overconfidence

Frozen models trained to mimic static datasets can never improve their performance. Models that can employ internet-retrieval for up-to-date information and obtain feedback from humans during deployment provide the promise of both adapting to new information, and improving their performance. In this work we study how to improve internet-driven conversational skills in such a learning framework. We collect deployment data, which we make publicly available, of human interactions, and collect various types of human feedback -- including binary quality measurements, free-form text feedback, and fine-grained reasons for failure. We then study various algorithms for improving from such feedback, including standard supervised learning, rejection sampling, model-guiding and reward-based learning, in order to make recommendations on which type of feedback and algorithms work best. We find the recently introduced Director model (Arora et al., '22) shows significant improvements over other existing approaches.

Empowering large language models to accurately express confidence in their answers is essential for trustworthy decision-making. Previous confidence elicitation methods, which primarily rely on white-box access to internal model information or model fine-tuning, have become less suitable for LLMs, especially closed-source commercial APIs. This leads to a growing need to explore the untapped area of black-box approaches for LLM uncertainty estimation. To better break down the problem, we define a systematic framework with three components: prompting strategies for eliciting verbalized confidence, sampling methods for generating multiple responses, and aggregation techniques for computing consistency. We then benchmark these methods on two key tasks-confidence calibration and failure prediction-across five types of datasets (e.g., commonsense and arithmetic reasoning) and five widely-used LLMs including GPT-4 and LLaMA 2 Chat. Our analysis uncovers several key insights: 1) LLMs, when verbalizing their confidence, tend to be overconfident, potentially imitating human patterns of expressing confidence. 2) As model capability scales up, both calibration and failure prediction performance improve. 3) Employing our proposed strategies, such as human-inspired prompts, consistency among multiple responses, and better aggregation strategies can help mitigate this overconfidence from various perspectives. 4) Comparisons with white-box methods indicate that while white-box methods perform better, the gap is narrow, e.g., 0.522 to 0.605 in AUROC. Despite these advancements, none of these techniques consistently outperform others, and all investigated methods struggle in challenging tasks, such as those requiring professional knowledge, indicating significant scope for improvement. We believe this study can serve as a strong baseline and provide insights for eliciting confidence in black-box LLMs.

Previous work has shown that there exists a scaling law between the size of Language Models (LMs) and their zero-shot performance on different downstream NLP tasks. In this work, we show that this phenomenon does not hold when evaluating large LMs on tasks with negated prompts, but instead shows an inverse scaling law. We evaluate 9 different tasks with negated prompts on (1) pretrained LMs (OPT & GPT-3) of varying sizes (125M - 175B), (2) LMs further pretrained to generalize to novel prompts (InstructGPT), (3) LMs provided with few-shot examples, and (4) LMs fine-tuned specifically on negated prompts; all LM types perform worse on negated prompts as they scale and show a huge performance gap between the human performance when comparing the average score on both original and negated prompts. By highlighting a critical limitation of existing LMs and methods, we urge the community to develop new approaches of developing LMs that actually follow the given instructions. We provide the code and the datasets to explore negated prompts at https://github.com/joeljang/negated-prompts-for-llms

Science journalism reports current scientific discoveries to non-specialists, aiming to enable public comprehension of the state of the art. However, this task can be challenging as the audience often lacks specific knowledge about the presented research. To address this challenge, we propose a framework that integrates three LLMs mimicking the real-world writing-reading-feedback-revision workflow, with one LLM acting as the journalist, a smaller LLM as the general public reader, and the third LLM as an editor. The journalist's writing is iteratively refined by feedback from the reader and suggestions from the editor. Our experiments demonstrate that by leveraging the collaboration of two 7B and one 1.8B open-source LLMs, we can generate articles that are more accessible than those generated by existing methods, including advanced models such as GPT-4.

Existing benchmarks have proven effective for assessing the performance of fully trained large language models. However, we find striking differences in the early training stages of small models, where benchmarks often fail to provide meaningful or discriminative signals. To explore how these differences arise, this competition tackles the challenge of designing scientific knowledge evaluation tasks specifically tailored for measuring early training progress of language models. Participants are invited to develop novel evaluation methodologies or adapt existing benchmarks to better capture performance differences among language models. To support this effort, we provide three pre-trained small models (0.5B, 1B, and 3B parameters), along with intermediate checkpoints sampled during training up to 200B tokens. All experiments and development work can be run on widely available free cloud-based GPU platforms, making participation accessible to researchers with limited computational resources. Submissions will be evaluated based on three criteria: the quality of the performance signal they produce, the consistency of model rankings at 1 trillion tokens of training, and their relevance to the scientific knowledge domain. By promoting the design of tailored evaluation strategies for early training, this competition aims to attract a broad range of participants from various disciplines, including those who may not be machine learning experts or have access to dedicated GPU resources. Ultimately, this initiative seeks to make foundational LLM research more systematic and benchmark-informed from the earliest phases of model development.

Human feedback has become the de facto standard for evaluating the performance of Large Language Models, and is increasingly being used as a training objective. However, it is not clear which properties of a generated output this single `preference' score captures. We hypothesise that preference scores are subjective and open to undesirable biases. We critically analyse the use of human feedback for both training and evaluation, to verify whether it fully captures a range of crucial error criteria. We find that while preference scores have fairly good coverage, they under-represent important aspects like factuality. We further hypothesise that both preference scores and error annotation may be affected by confounders, and leverage instruction-tuned models to generate outputs that vary along two possible confounding dimensions: assertiveness and complexity. We find that the assertiveness of an output skews the perceived rate of factuality errors, indicating that human annotations are not a fully reliable evaluation metric or training objective. Finally, we offer preliminary evidence that using human feedback as a training objective disproportionately increases the assertiveness of model outputs. We encourage future work to carefully consider whether preference scores are well aligned with the desired objective.

We present Multi-expert Prompting, a novel enhancement of ExpertPrompting (Xu et al., 2023), designed to improve the large language model (LLM) generation. Specifically, it guides an LLM to fulfill an input instruction by simulating multiple experts, aggregating their responses, and selecting the best among individual and aggregated responses. This process is performed in a single chain of thoughts through our seven carefully designed subtasks derived from the Nominal Group Technique (Ven and Delbecq, 1974), a well-established decision-making framework. Our evaluations demonstrate that Multi-expert Prompting significantly outperforms ExpertPrompting and comparable baselines in enhancing the truthfulness, factuality, informativeness, and usefulness of responses while reducing toxicity and hurtfulness. It further achieves state-of-the-art truthfulness by outperforming the best baseline by 8.69% with ChatGPT. Multi-expert Prompting is efficient, explainable, and highly adaptable to diverse scenarios, eliminating the need for manual prompt construction.

Comprehensively understanding and accurately predicting the performance of large language models across diverse downstream tasks has emerged as a pivotal challenge in NLP research. The pioneering scaling law on downstream works demonstrated intrinsic similarities within model families and utilized such similarities for performance prediction. However, they tend to overlook the similarities between model families and only consider design factors listed in the original scaling law. To overcome these limitations, we introduce a novel framework, Collaborative Performance Prediction (CPP), which significantly enhances prediction accuracy by leveraging the historical performance of various models on downstream tasks and other design factors for both model and task. We also collect a collaborative data sourced from online platforms containing both historical performance and additional design factors. With the support of the collaborative data, CPP not only surpasses traditional scaling laws in predicting the performance of scaled LLMs but also facilitates a detailed analysis of factor importance, an area previously overlooked.

Recent improvements in text generation have leveraged human feedback to improve the quality of the generated output. However, human feedback is not always available, especially during inference. In this work, we propose an inference time optimization method FITO to use fine-grained actionable feedback in the form of error type, error location and severity level that are predicted by a learned error pinpoint model for iterative refinement. FITO starts with an initial output, then iteratively incorporates the feedback via a refinement model that generates an improved output conditioned on the feedback. Given the uncertainty of consistent refined samples at iterative steps, we formulate iterative refinement into a local search problem and develop a simulated annealing based algorithm that balances exploration of the search space and optimization for output quality. We conduct experiments on three text generation tasks, including machine translation, long-form question answering (QA) and topical summarization. We observe 0.8 and 0.7 MetricX gain on Chinese-English and English-German translation, 4.5 and 1.8 ROUGE-L gain at long form QA and topic summarization respectively, with a single iteration of refinement. With our simulated annealing algorithm, we see further quality improvements, including up to 1.7 MetricX improvements over the baseline approach.

The advent of test-time scaling in large language models (LLMs), exemplified by OpenAI's o1 series, has advanced reasoning capabilities by scaling computational resource allocation during inference. While successors like QwQ, Deepseek-R1 (R1) and LIMO replicate these advancements, whether these models truly possess test-time scaling capabilities remains underexplored. This study found that longer CoTs of these o1-like models do not consistently enhance accuracy; in fact, correct solutions are often shorter than incorrect ones for the same questions. Further investigation shows this phenomenon is closely related to models' self-revision capabilities - longer CoTs contain more self-revisions, which often lead to performance degradation. We then compare sequential and parallel scaling strategies on QwQ, R1 and LIMO, finding that parallel scaling achieves better coverage and scalability. Based on these insights, we propose Shortest Majority Vote, a method that combines parallel scaling strategies with CoT length characteristics, significantly improving models' test-time scalability compared to conventional majority voting approaches.

Large language models (LLMs) have shown promise for automatic summarization but the reasons behind their successes are poorly understood. By conducting a human evaluation on ten LLMs across different pretraining methods, prompts, and model scales, we make two important observations. First, we find instruction tuning, and not model size, is the key to the LLM's zero-shot summarization capability. Second, existing studies have been limited by low-quality references, leading to underestimates of human performance and lower few-shot and finetuning performance. To better evaluate LLMs, we perform human evaluation over high-quality summaries we collect from freelance writers. Despite major stylistic differences such as the amount of paraphrasing, we find that LMM summaries are judged to be on par with human written summaries.

To assist human review process, we build a novel ReviewRobot to automatically assign a review score and write comments for multiple categories such as novelty and meaningful comparison. A good review needs to be knowledgeable, namely that the comments should be constructive and informative to help improve the paper; and explainable by providing detailed evidence. ReviewRobot achieves these goals via three steps: (1) We perform domain-specific Information Extraction to construct a knowledge graph (KG) from the target paper under review, a related work KG from the papers cited by the target paper, and a background KG from a large collection of previous papers in the domain. (2) By comparing these three KGs, we predict a review score and detailed structured knowledge as evidence for each review category. (3) We carefully select and generalize human review sentences into templates, and apply these templates to transform the review scores and evidence into natural language comments. Experimental results show that our review score predictor reaches 71.4%-100% accuracy. Human assessment by domain experts shows that 41.7%-70.5% of the comments generated by ReviewRobot are valid and constructive, and better than human-written ones for 20% of the time. Thus, ReviewRobot can serve as an assistant for paper reviewers, program chairs and authors.

Large language models (LLMs) have demonstrated remarkable capabilities in language generation, understanding, and few-shot learning in recent years. An extensive body of work has explored how their performance may be further improved through the tools of prompting, ranging from verification, self-consistency, or intermediate scratchpads. In this paper, we present a complementary approach to improve language responses where multiple language model instances propose and debate their individual responses and reasoning processes over multiple rounds to arrive at a common final answer. Our findings indicate that this approach significantly enhances mathematical and strategic reasoning across a number of tasks. We also demonstrate that our approach improves the factual validity of generated content, reducing fallacious answers and hallucinations that contemporary models are prone to. Our approach may be directly applied to existing black-box models and uses identical procedure and prompts for all tasks we investigate. Overall, our findings suggest that such "society of minds" approach has the potential to significantly advance the capabilities of LLMs and pave the way for further breakthroughs in language generation and understanding.

Pre-trained language models (PLMs) have shown unprecedented potential in various fields, especially as the backbones for question-answering (QA) systems. However, they tend to be easily deceived by tricky questions such as "How many eyes does the sun have?". Such frailties of PLMs often allude to the lack of knowledge within them. In this paper, we find that the PLMs already possess the knowledge required to rebut such questions, and the key is how to activate the knowledge. To systematize this observation, we investigate the PLMs' responses to one kind of tricky questions, i.e., the false premises questions (FPQs). We annotate a FalseQA dataset containing 2365 human-written FPQs, with the corresponding explanations for the false premises and the revised true premise questions. Using FalseQA, we discover that PLMs are capable of discriminating FPQs by fine-tuning on moderate numbers (e.g., 256) of examples. PLMs also generate reasonable explanations for the false premise, which serve as rebuttals. Further replaying a few general questions during training allows PLMs to excel on FPQs and general questions simultaneously. Our work suggests that once the rebuttal ability is stimulated, knowledge inside the PLMs can be effectively utilized to handle FPQs, which incentivizes the research on PLM-based QA systems.

With infinitely many high-quality data points, infinite computational power, an infinitely large foundation model with a perfect training algorithm and guaranteed zero generalization error on the pretext task, can the model be used for everything? This question cannot be answered by the existing theory of representation, optimization or generalization, because the issues they mainly investigate are assumed to be nonexistent here. In this paper, we show that category theory provides powerful machinery to answer this question. We have proved three results. The first one limits the power of prompt-based learning, saying that the model can solve a downstream task with prompts if and only if the task is representable. The second one says fine tuning does not have this limit, as a foundation model with the minimum required power (up to symmetry) can theoretically solve downstream tasks for the category defined by pretext task, with fine tuning and enough resources. Our final result can be seen as a new type of generalization theorem, showing that the foundation model can generate unseen objects from the target category (e.g., images) using the structural information from the source category (e.g., texts). Along the way, we provide a categorical framework for supervised and self-supervised learning, which might be of independent interest.

We evaluate recent Large language Models (LLMs) on the challenging task of summarizing short stories, which can be lengthy, and include nuanced subtext or scrambled timelines. Importantly, we work directly with authors to ensure that the stories have not been shared online (and therefore are unseen by the models), and to obtain informed evaluations of summary quality using judgments from the authors themselves. Through quantitative and qualitative analysis grounded in narrative theory, we compare GPT-4, Claude-2.1, and LLama-2-70B. We find that all three models make faithfulness mistakes in over 50% of summaries and struggle to interpret difficult subtext. However, at their best, the models can provide thoughtful thematic analysis of stories. We additionally demonstrate that LLM judgments of summary quality do not match the feedback from the writers.

While Large Language Models (LLMs) have demonstrated remarkable capabilities in scientific tasks, existing evaluation frameworks primarily assess their performance using rich contextual inputs, overlooking their ability to generate novel ideas from minimal information. We introduce LiveIdeaBench, a comprehensive benchmark that evaluates LLMs' scientific creativity and divergent thinking capabilities using single-keyword prompts. Drawing from Guilford's creativity theory, our framework employs a dynamic panel of state-of-the-art LLMs to assess generated ideas across four key dimensions: originality, feasibility, fluency, and flexibility. Through extensive experimentation with 20 leading models across 1,180 keywords spanning 18 scientific domains, we reveal that scientific creative ability shows distinct patterns from general intelligence metrics. Notably, our results demonstrate that models like QwQ-32B-preview achieve comparable creative performance to top-tier models like o1-preview, despite significant gaps in their general intelligence scores. These findings highlight the importance of specialized evaluation frameworks for scientific creativity and suggest that the development of creative capabilities in LLMs may follow different trajectories than traditional problem-solving abilities.

Tool learning enables Large Language Models (LLMs) to interact with external environments by invoking tools, serving as an effective strategy to mitigate the limitations inherent in their pre-training data. In this process, tool documentation plays a crucial role by providing usage instructions for LLMs, thereby facilitating effective tool utilization. This paper concentrates on the critical challenge of bridging the comprehension gap between LLMs and external tools due to the inadequacies and inaccuracies inherent in existing human-centric tool documentation. We propose a novel framework, DRAFT, aimed at Dynamically Refining tool documentation through the Analysis of Feedback and Trails emanating from LLMs' interactions with external tools. This methodology pivots on an innovative trial-and-error approach, consisting of three distinct learning phases: experience gathering, learning from experience, and documentation rewriting, to iteratively enhance the tool documentation. This process is further optimized by implementing a diversity-promoting exploration strategy to ensure explorative diversity and a tool-adaptive termination mechanism to prevent overfitting while enhancing efficiency. Extensive experiments on multiple datasets demonstrate that DRAFT's iterative, feedback-based refinement significantly ameliorates documentation quality, fostering a deeper comprehension and more effective utilization of tools by LLMs. Notably, our analysis reveals that the tool documentation refined via our approach demonstrates robust cross-model generalization capabilities.

In many text-generation problems, users may prefer not only a single response, but a diverse range of high-quality outputs from which to choose. Quality-diversity (QD) search algorithms aim at such outcomes, by continually improving and diversifying a population of candidates. However, the applicability of QD to qualitative domains, like creative writing, has been limited by the difficulty of algorithmically specifying measures of quality and diversity. Interestingly, recent developments in language models (LMs) have enabled guiding search through AI feedback, wherein LMs are prompted in natural language to evaluate qualitative aspects of text. Leveraging this development, we introduce Quality-Diversity through AI Feedback (QDAIF), wherein an evolutionary algorithm applies LMs to both generate variation and evaluate the quality and diversity of candidate text. When assessed on creative writing domains, QDAIF covers more of a specified search space with high-quality samples than do non-QD controls. Further, human evaluation of QDAIF-generated creative texts validates reasonable agreement between AI and human evaluation. Our results thus highlight the potential of AI feedback to guide open-ended search for creative and original solutions, providing a recipe that seemingly generalizes to many domains and modalities. In this way, QDAIF is a step towards AI systems that can independently search, diversify, evaluate, and improve, which are among the core skills underlying human society's capacity for innovation.

As language models (LMs) scale, they develop many novel behaviors, good and bad, exacerbating the need to evaluate how they behave. Prior work creates evaluations with crowdwork (which is time-consuming and expensive) or existing data sources (which are not always available). Here, we automatically generate evaluations with LMs. We explore approaches with varying amounts of human effort, from instructing LMs to write yes/no questions to making complex Winogender schemas with multiple stages of LM-based generation and filtering. Crowdworkers rate the examples as highly relevant and agree with 90-100% of labels, sometimes more so than corresponding human-written datasets. We generate 154 datasets and discover new cases of inverse scaling where LMs get worse with size. Larger LMs repeat back a dialog user's preferred answer ("sycophancy") and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down. Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.

Benchmarks are essential for quantitatively tracking progress in AI. As AI agents become increasingly capable, researchers and practitioners have introduced agentic benchmarks to evaluate agents on complex, real-world tasks. These benchmarks typically measure agent capabilities by evaluating task outcomes via specific reward designs. However, we show that many agentic benchmarks have issues task setup or reward design. For example, SWE-bench Verified uses insufficient test cases, while TAU-bench counts empty responses as successful. Such issues can lead to under- or overestimation agents' performance by up to 100% in relative terms. To make agentic evaluation rigorous, we introduce the Agentic Benchmark Checklist (ABC), a set of guidelines that we synthesized from our benchmark-building experience, a survey of best practices, and previously reported issues. When applied to CVE-Bench, a benchmark with a particularly complex evaluation design, ABC reduces the performance overestimation by 33%.

Large Language Models (LLMs) have shown exceptional results on current benchmarks when working individually. The advancement in their capabilities, along with a reduction in parameter size and inference times, has facilitated the use of these models as agents, enabling interactions among multiple models to execute complex tasks. Such collaborations offer several advantages, including the use of specialized models (e.g. coding), improved confidence through multiple computations, and enhanced divergent thinking, leading to more diverse outputs. Thus, the collaborative use of language models is expected to grow significantly in the coming years. In this work, we evaluate the behavior of a network of models collaborating through debate under the influence of an adversary. We introduce pertinent metrics to assess the adversary's effectiveness, focusing on system accuracy and model agreement. Our findings highlight the importance of a model's persuasive ability in influencing others. Additionally, we explore inference-time methods to generate more compelling arguments and evaluate the potential of prompt-based mitigation as a defensive strategy.

Benchmarking is seen as critical to assessing progress in NLP. However, creating a benchmark involves many design decisions (e.g., which datasets to include, which metrics to use) that often rely on tacit, untested assumptions about what the benchmark is intended to measure or is actually measuring. There is currently no principled way of analyzing these decisions and how they impact the validity of the benchmark's measurements. To address this gap, we draw on evidence-centered design in educational assessments and propose Evidence-Centered Benchmark Design (ECBD), a framework which formalizes the benchmark design process into five modules. ECBD specifies the role each module plays in helping practitioners collect evidence about capabilities of interest. Specifically, each module requires benchmark designers to describe, justify, and support benchmark design choices -- e.g., clearly specifying the capabilities the benchmark aims to measure or how evidence about those capabilities is collected from model responses. To demonstrate the use of ECBD, we conduct case studies with three benchmarks: BoolQ, SuperGLUE, and HELM. Our analysis reveals common trends in benchmark design and documentation that could threaten the validity of benchmarks' measurements.

There is growing excitement about the potential of Language Models (LMs) to accelerate scientific discovery. Falsifying hypotheses is key to scientific progress, as it allows claims to be iteratively refined over time. This process requires significant researcher effort, reasoning, and ingenuity. Yet current benchmarks for LMs predominantly assess their ability to generate solutions rather than challenge them. We advocate for developing benchmarks that evaluate this inverse capability - creating counterexamples for subtly incorrect solutions. To demonstrate this approach, we start with the domain of algorithmic problem solving, where counterexamples can be evaluated automatically using code execution. Specifically, we introduce REFUTE, a dynamically updating benchmark that includes recent problems and incorrect submissions from programming competitions, where human experts successfully identified counterexamples. Our analysis finds that the best reasoning agents, even OpenAI o3-mini (high) with code execution feedback, can create counterexamples for only <9% of incorrect solutions in REFUTE, even though ratings indicate its ability to solve up to 48% of these problems from scratch. We hope our work spurs progress in evaluating and enhancing LMs' ability to falsify incorrect solutions - a capability that is crucial for both accelerating research and making models self-improve through reliable reflective reasoning.

Despite the utility of Large Language Models (LLMs) across a wide range of tasks and scenarios, developing a method for reliably evaluating LLMs across varied contexts continues to be challenging. Modern evaluation approaches often use LLMs to assess responses generated by LLMs. However, the meta-evaluation conducted to assess the effectiveness of these LLMs as evaluators is typically constrained by the coverage of existing benchmarks or requires extensive human annotation. This underscores the urgency of methods for scalable meta-evaluation that can effectively, reliably, and efficiently evaluate the performance of LLMs as evaluators across diverse tasks and scenarios, particularly in potentially new, user-defined scenarios. To fill this gap, we propose ScaleEval, an agent-debate-assisted meta-evaluation framework that leverages the capabilities of multiple communicative LLM agents. This framework supports multi-round discussions to assist human annotators in discerning the most capable LLMs as evaluators, which significantly eases their workload in cases that used to require large-scale annotations during meta-evaluation. We release the code for our framework, which is publicly available at: https://github.com/GAIR-NLP/scaleeval.

We present SimpleQA, a benchmark that evaluates the ability of language models to answer short, fact-seeking questions. We prioritized two properties in designing this eval. First, SimpleQA is challenging, as it is adversarially collected against GPT-4 responses. Second, responses are easy to grade, because questions are created such that there exists only a single, indisputable answer. Each answer in SimpleQA is graded as either correct, incorrect, or not attempted. A model with ideal behavior would get as many questions correct as possible while not attempting the questions for which it is not confident it knows the correct answer. SimpleQA is a simple, targeted evaluation for whether models "know what they know," and our hope is that this benchmark will remain relevant for the next few generations of frontier models. SimpleQA can be found at https://github.com/openai/simple-evals.

Recent advancements in Language Models (LMs) have catalyzed the creation of multiple benchmarks, designed to assess these models' general capabilities. A crucial task, however, is assessing the validity of the benchmarks themselves. This is most commonly done via Benchmark Agreement Testing (BAT), where new benchmarks are validated against established ones using some agreement metric (e.g., rank correlation). Despite the crucial role of BAT for benchmark builders and consumers, there are no standardized procedures for such agreement testing. This deficiency can lead to invalid conclusions, fostering mistrust in benchmarks and upending the ability to properly choose the appropriate benchmark to use. By analyzing over 40 prominent benchmarks, we demonstrate how some overlooked methodological choices can significantly influence BAT results, potentially undermining the validity of conclusions. To address these inconsistencies, we propose a set of best practices for BAT and demonstrate how utilizing these methodologies greatly improves BAT robustness and validity. To foster adoption and facilitate future research,, we introduce BenchBench, a python package for BAT, and release the BenchBench-leaderboard, a meta-benchmark designed to evaluate benchmarks using their peers. Our findings underscore the necessity for standardized BAT, ensuring the robustness and validity of benchmark evaluations in the evolving landscape of language model research. BenchBench Package: https://github.com/IBM/BenchBench Leaderboard: https://huggingface.co/spaces/per/BenchBench

Foundation models are increasingly used in scientific research, but evaluating AI-generated scientific work remains challenging. While expert reviews are costly, large language models (LLMs) as proxy reviewers have proven to be unreliable. To address this, we investigate two automatic evaluation metrics, specifically citation count prediction and review score prediction. We parse all papers of OpenReview and augment each submission with its citation count, reference, and research hypothesis. Our findings reveal that citation count prediction is more viable than review score prediction, and predicting scores is more difficult purely from the research hypothesis than from the full paper. Furthermore, we show that a simple prediction model based solely on title and abstract outperforms LLM-based reviewers, though it still falls short of human-level consistency.

Multi-agent debates have been introduced to improve the accuracy of Large Language Models (LLMs) by having multiple agents discuss solutions to a problem over several rounds of debate. However, models often generate incorrect yet confident-sounding responses, which can mislead others. This issue arises partly because agents do not consider how confident their peers are. To address this, we propose DebUnc, a debate framework that uses uncertainty metrics to assess agent confidence. Confidence is then conveyed through a modified attention mechanism that adjusts token weights, or through textual prompts. Evaluations across benchmarks show that attention-based methods are particularly effective and that performance continues to improve as uncertainty estimation becomes more reliable. The code is available at https://github.com/lukeyoffe/debunc.