File size: 3,888 Bytes
6f7e284 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 |
---
license: mit
datasets:
- agentica-org/DeepScaleR-Preview-Dataset
base_model:
- deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
tags:
- LRM
- hybrid_reasoning
- efficient_reasoning
---
# AdaptThink: LLM Can Learn When to Think
<p align="center">
🤗 <a href="https://huggingface.co/collections/THU-KEG/adaptthink-682a1059aa9f5102c4fa0470" target="_blank">HF Collections</a> • 💻 <a href="" target="_blank">Github Repo</a> • 📃 <a href="https://arxiv.org/abs/2505.13417" target="_blank">Paper</a>
</p>
## 🔍 Table of Contents
- [🤖️ AdaptThink](#adapt_think)
- [⚙️ Released Models](#model)
- [📊 Evaluation](#evaluation)
- [📝 Citation](#citation)
<a name="adapt_think"></a>
## 🤖️ AdaptThink
We present **AdapThink**, a novel reinforcement learning (RL) algorithm that enables reasoning models to adaptively choose between **Thinking** and **NoThinking** modes according to the difficulty of each input problem, thereby achieving automatic hybrid reasoning. Specifically, the model engages in thinking only when the problem is determined to be challenging; for other simple question, it will bypass the thinking process and directly produce a concise final solution. This approach substantially reduces inference costs while further improving overall performance.
<a name="model"></a>
## ⚙️ Released Models
### All Available Datasets and Models
We apply the AdaptThink algorithm on DeepSeek-R1-Distill-Qwen-1.5B with $\delta$ from 0 to 0.1, and DeepSeek-R1-Distill-Qwen-7B with $\delta=0.05$. A larger $\large$ results in a higher proportion of NoThinking responses, which reduces more inference costs but also diminish the resultant improvement in accuracy.
All the trained models are available on HuggingFace.
| Name | HF Repo |
|---|---|
| AdaptThink-1.5B-delta0 | [🤗 HF Repo](https://huggingface.co/THU-KEG/AdaptThink-1.5B-delta0) |
| AdaptThink-1.5B-delta0.01 | [🤗 HF Repo](https://huggingface.co/THU-KEG/AdaptThink-1.5B-delta0.01) |
| AdaptThink-1.5B-delta0.02 | [🤗 HF Repo](https://huggingface.co/THU-KEG/AdaptThink-1.5B-delta0.02) |
| AdaptThink-1.5B-delta0.05 | [🤗 HF Repo](https://huggingface.co/THU-KEG/AdaptThink-1.5B-delta0.05) |
| AdaptThink-1.5B-delta0.075 | [🤗 HF Repo](https://huggingface.co/THU-KEG/AdaptThink-1.5B-delta0.075) |
| AdaptThink-1.5B-delta0.1 | [🤗 HF Repo](https://huggingface.co/THU-KEG/AdaptThink-1.5B-delta0.1) |
| AdaptThink-7B-delta0.05 | [🤗 HF Repo](https://huggingface.co/THU-KEG/AdaptThink-7B-delta0.05) |
<a name="training"></a>
## 📊 Evaluation Results
We list our evaluation results as follows:
##### 1. Comparison with existing methods for efficient reasoning on mathematics datasets
##### 2. Nothinking responses ratio and accuracy across different difficulty levels on MATH500
##### 3. Comparison of different $\delta$ values
##### 4. Evaluation results on MMLU
<img width="1000" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/66cdd285c51a915bd5f2d017/19K2u6PNmYz3gx3JnHgn4.png">
<a name="citation"></a>
## 📝 Citation
If you find our work useful, please consider citing LongReward:
```
@article{zhang2025adapt_think,
title = {AdaptThink: LLM Can Learn When to Think}
author={Jiajie Zhang and Nianyi Lin and Lei Hou and Ling Feng and Juanzi Li},
journal={arXiv preprint arXiv: 2505.13417},
url={https://arxiv.org/abs/2505.13417}
year={2025}
}
```
|