cesun/ThinkEdit-deepseek-llama3-8b

Repository for:

ThinkEdit-deepseek-llama3-8b

(We also release ThinkEdit versions for ThinkEdit-deepseek-qwen-1.5b, ThinkEdit-deepseek-qwen-14b, and ThinkEdit-deepseek-qwen-32b.)

Authors: Chung-En Sun, Ge Yan, Tsui-Wei Weng
Paper: ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models

Github: https://github.com/Trustworthy-ML-Lab/ThinkEdit

---

Introduction

Reasoning-augmented models sometimes fail by generating overly short, abstract chain-of-thought (CoT) reasoning, hurting their accuracy.

ThinkEdit is a lightweight weight-editing method that:

• Identifies ~4% of "short reasoning" attention heads
• Edits only ~0.2% of total parameters
• Removes the "short reasoning" direction from their output
• Boosts performance, especially on cases with short reasoning traces

---

Full Performance Results

1. Overall Accuracy

Model	GSM8K	MMLU Elementary Math	MATH-Level1	MATH-Level5	MATH-500
deepseek-qwen-32b	92.97 ± 0.39	95.93 ± 0.83	96.41 ± 0.45	91.27 ± 0.53	91.62 ± 0.58
ThinkEdit-deepseek-qwen-32b	95.25 ± 0.25	98.02 ± 0.31	96.02 ± 0.42	91.31 ± 0.50	91.60 ± 0.65
deepseek-qwen-14b	90.80 ± 0.36	95.08 ± 0.65	96.32 ± 0.35	90.25 ± 0.72	91.48 ± 0.55
ThinkEdit-deepseek-qwen-14b	93.78 ± 0.50	96.56 ± 0.84	96.38 ± 0.52	91.03 ± 0.44	91.92 ± 0.63
deepseek-llama3-8b	82.26 ± 0.91	96.01 ± 0.62	93.46 ± 0.84	85.49 ± 0.83	87.26 ± 1.16
ThinkEdit-deepseek-llama3-8b	89.44 ± 0.55	96.19 ± 0.73	94.44 ± 0.31	86.49 ± 0.54	88.06 ± 1.09
deepseek-qwen-1.5b	79.15 ± 1.08	68.52 ± 1.56	93.00 ± 0.33	75.48 ± 0.90	82.22 ± 1.29
ThinkEdit-deepseek-qwen-1.5b	84.56 ± 0.79	90.66 ± 0.97	93.66 ± 0.62	75.05 ± 0.82	82.24 ± 0.89

---

2. Accuracy on Short Reasoning Cases (Top 5% / 10% / 20%)

Model	GSM8K	MMLU Elementary Math	MATH-Level1	MATH-Level5	MATH-500
deepseek-qwen-32b	98.31 / 97.18 / 96.20	97.78 / 97.03 / 95.87	100.00 / 100.00 / 98.97	93.03 / 96.36 / 97.35	86.40 / 92.00 / 94.00
ThinkEdit-deepseek-qwen-32b	98.92 / 97.71 / 97.83	97.78 / 97.57 / 97.20	100.00 / 100.00 / 98.74	98.03 / 98.64 / 97.99	92.00 / 94.40 / 95.80
deepseek-qwen-14b	96.31 / 95.65 / 92.93	93.89 / 96.22 / 95.60	99.52 / 99.30 / 97.70	89.39 / 94.32 / 96.25	86.40 / 91.40 / 93.50
ThinkEdit-deepseek-qwen-14b	96.31 / 96.18 / 96.77	97.78 / 95.14 / 96.53	99.53 / 98.62 / 98.67	96.67 / 97.88 / 98.11	91.20 / 93.20 / 95.00
deepseek-llama3-8b	88.92 / 87.18 / 85.82	97.22 / 96.49 / 96.80	97.14 / 94.88 / 94.83	78.64 / 88.79 / 93.41	82.00 / 81.40 / 88.30
ThinkEdit-deepseek-llama3-8b	97.08 / 95.27 / 93.95	97.78 / 98.65 / 97.87	100.00 / 99.30 / 98.62	95.61 / 96.89 / 97.12	92.80 / 93.60 / 94.40
deepseek-qwen-1.5b	88.46 / 87.48 / 85.02	62.78 / 62.16 / 60.53	97.62 / 95.12 / 93.91	91.52 / 95.00 / 95.72	82.40 / 89.80 / 93.40
ThinkEdit-deepseek-qwen-1.5b	92.62 / 92.90 / 92.32	87.78 / 88.11 / 88.67	95.71 / 95.58 / 96.44	95.15 / 96.59 / 97.27	90.80 / 92.00 / 94.20

---

Usage

The usage of ThinkEdit models is exactly the same as the original deepseek-distilled models.

Citation

@misc{sun2025thinkedit,
      title={ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models}, 
      author={Chung-En Sun and Ge Yan and Tsui-Wei Weng},
      year={2025},
      eprint={2503.22048},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.22048}, 
}