cesun/ThinkEdit-deepseek-qwen-14b

Repository for:

ThinkEdit-deepseek-qwen-14b

(We also release ThinkEdit versions for ThinkEdit-deepseek-qwen-1.5b and ThinkEdit-deepseek-llama3-8b.)

Authors: Chung-En Sun, Ge Yan, Tsui-Wei Weng
Paper: ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models
Code: 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 ~2% of "short reasoning" attention heads
• Edits only ~0.1% 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-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	93.50 ± 0.31	96.53 ± 0.54	96.50 ± 0.46	91.15 ± 0.59	91.78 ± 0.58
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	88.97 ± 0.78	96.08 ± 0.86	94.12 ± 0.47	85.91 ± 0.48	87.60 ± 0.81
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	83.34 ± 0.79	86.24 ± 1.12	93.89 ± 0.76	74.94 ± 0.85	82.74 ± 0.77

---

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

Model	GSM8K	MMLU Elementary Math	MATH-Level1	MATH-Level5	MATH-500
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	96.62 / 96.03 / 96.12	96.11 / 96.22 / 96.27	100.00 / 99.77 / 98.85	95.76 / 97.65 / 98.07	89.60 / 92.60 / 94.70
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	92.46 / 92.37 / 92.05	77.22 / 80.54 / 79.73	96.19 / 95.81 / 97.36	93.79 / 95.83 / 95.80	92.80 / 94.40 / 94.90

---

3. Reasoning Lengths (Top 5% / 10% / 20% Shortest Responses)

Model	GSM8K	MMLU Elementary Math	MATH-Level1	MATH-Level5	MATH-500
deepseek-qwen-14b	76.6 / 86.5 / 99.1	65.8 / 72.2 / 80.6	93.7 / 114.3 / 188.6	628.8 / 858.4 / 1125.9	198.7 / 434.3 / 697.0
ThinkEdit-deepseek-qwen-14b	95.4 / 106.3 / 120.2	79.1 / 87.1 / 98.7	125.1 / 150.2 / 243.4	698.5 / 906.6 / 1157.2	270.2 / 492.6 / 733.3
deepseek-llama3-8b	73.0 / 83.1 / 96.6	371.0 / 438.1 / 518.2	80.3 / 97.2 / 130.3	617.9 / 854.9 / 1126.5	159.5 / 357.5 / 644.5
ThinkEdit-deepseek-llama3-8b	93.2 / 106.9 / 127.4	396.5 / 464.2 / 543.2	137.4 / 173.3 / 277.1	791.2 / 954.8 / 1185.1	305.2 / 506.3 / 737.6
deepseek-qwen-1.5b	78.8 / 89.4 / 103.0	61.6 / 68.5 / 77.6	88.8 / 110.3 / 219.7	804.6 / 1017.9 / 1314.0	249.7 / 506.5 / 760.7
ThinkEdit-deepseek-qwen-1.5b	97.2 / 109.4 / 126.3	75.9 / 85.0 / 99.5	127.9 / 174.1 / 416.4	818.0 / 984.5 / 1214.3	435.0 / 612.9 / 800.6

---

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}, 
}