ThinkEdit

This is the official repository for the paper: ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models. Here is our 5 min read project website.

🆕 Updates

• 2025/08/20 — 🎉 Our paper has been accepted to EMNLP 2025 main conference!
• 2025/06/06 — 🚀 Released improved ThinkEdit models with 4% short reasoning heads edited and introduced the larger ThinkEdit-deepseek-qwen-32b model. Please see our updated paper.

Overview

Set Up

pip install -r requirements.txt

If you want to skip all the steps and directly access the resulting output files, you can download them through:

gdown https://drive.google.com/uc?id=135ttUxdNc4gQkmbZlpr3ecbhA3_392PE

and then unzip the file

unzip ThinkEdit.zip

Steer along Reasoning Length Direction

Step 1: Generate responses for probing from GSM8k

First, collect the responses from the reasoning models and store them in responses/ for extracting hidden states later:

python generate response_gsm8k.py

Specify the --model argument: deepseek-qwen-1.5b, deepseek-llama3-8b, deepseek-qwen-14b, deepseek-qwen-32b.

Step 2: Extract the Reasoning Length Direction

Next, extract the layerwise directions from Self-Attn or MLP and store them in directions/:

python extract_thinking_length_directiongsm8k_attn.py
python extract_thinking_length_directiongsm8k_mlp.py

Specify the --model argument: deepseek-qwen-1.5b, deepseek-llama3-8b, deepseek-qwen-14b, deepseek-qwen-32b.

Step 3: Steer the reasoning length of the models

Finally, steer the models with the directions and observe changes in accuracy and reasoning length. To evaluate on 200 test examples from gsm8k and store the results in gsm8k_all_layer_thinking_length_steering_results/:

python thinking_length_steering_gsm8k.py

Specify the --model argument: deepseek-qwen-1.5b, deepseek-llama3-8b, deepseek-qwen-14b, deepseek-qwen-32b.

--control argument options: thinking_length_attn, thinking_length_mlp.

The steering strength alpha (--direction_weight): we use -0.08 -0.07 ... 0.07 0.08 in our paper.

Similarly, to evaluate 140 Level-5 examples from MATH and store the results in math_level5_all_layer_thinking_length_steering_results/:

python thinking_length_steering_math_level6.py

Specify arguments accordingly.

To steer only one layer each time and store the results in gsm8k_layerwise_thinking_length_steering_results/:

python thinking_length_layerwise_steering_gsm8k.py

Specify arguments accordingly. Use --layer to specify the layer and set --direction_weight to -1 or 1 (as in our paper). Running the layerwise analysis can take considerable time. We suggest using automate_layerwise_steering_jobs.sh to handle the jobs; please modify the script based on your hardware.

ThinkEdit models: Weight editing short reasoning heads

Step 1: Find the short reasoning heads

First, identify the short reasoning heads by calculating their per-head contribution to the short reasoning direction:

python find_short_thinking_attn_heads.py

Specify the --model argument: deepseek-qwen-1.5b, deepseek-llama3-8b, deepseek-qwen-14b, deepseek-qwen-32b.

This will output a list of short reasoning heads and a heatmap figure of every head's contribution.

Step 2: Perform Weight Editing

Next, perform weight editing to the o_proj layer of the short reasoning heads and store the model under ThinkEdit_models/:

python get_ThinkEdit_models.py

Specify the --model argument: deepseek-qwen-1.5b, deepseek-llama3-8b, deepseek-qwen-14b, deepseek-qwen-32b.

We have provided ThinkEdit models on the Huggingface repo:

• cesun/ThinkEdit-deepseek-qwen-32b
• cesun/ThinkEdit-deepseek-qwen-14b
• cesun/ThinkEdit-deepseek-llama3-8b
• cesun/ThinkEdit-deepseek-qwen-1.5b

You can skip this step and our evaluation script will directly download the models from Huggingface.

Step 3: Evaluate the performance of the ThinkEdit models

Finally, evaluate the performance of the original and ThinkEdit models and store the results under ThinkEdit_model_evaluation_results/. We use vllm to speed up evaluation:

CUDA_VISIBLE_DEVICES={your available gpus} python evaluate_ThinkEdit_models.py

Specify the --model argument: deepseek-qwen-32b, deepseek-qwen-14b, deepseek-llama3-8b, deepseek-qwen-1.5b, ThinkEdit-deepseek-qwen-32b, ThinkEdit-deepseek-qwen-14b, ThinkEdit-deepseek-llama3-8b, ThinkEdit-deepseek-qwen-1.5b.

--dataset argument: gsm8k, mmlu_elementary_math, MATH-500, MATH-level1, MATH-level5.

--n_samples argument: we set this to 10 in our paper, meaning each question is evaluated 10 times.

--tensor_parallel_size argument: set this according to your number of GPUs; it should be a factor of the number of attention heads in each model. We recommend setting it to 4.

After you have all the results, run:

python analyze_ThinkEdit_performance.py

to generate the plots and tables presented in our paper. You can expect output similar to the results shown below.

Results

1. Overall Accuracy

Model	GSM8K	MMLU Elem.	MATH-Lv1	MATH-Lv5	MATH500
ds-qwen-32b	92.97 ± 0.39	95.93 ± 0.83	96.41 ± 0.45	91.27 ± 0.53	91.62 ± 0.58
ThinkEdit-ds-qwen-32b	95.25 ± 0.25	98.02 ± 0.31	96.02 ± 0.42	91.31 ± 0.50	91.60 ± 0.65
ds-qwen-14b	90.80 ± 0.36	95.08 ± 0.65	96.32 ± 0.35	90.25 ± 0.72	91.48 ± 0.55
ThinkEdit-ds-qwen-14b	93.78 ± 0.50	96.56 ± 0.84	96.38 ± 0.52	91.03 ± 0.44	91.92 ± 0.63
ds-llama3-8b	82.26 ± 0.91	96.01 ± 0.62	93.46 ± 0.84	85.49 ± 0.83	87.26 ± 1.16
ThinkEdit-ds-llama3-8b	89.44 ± 0.55	96.19 ± 0.73	94.44 ± 0.31	86.49 ± 0.54	88.06 ± 1.09
ds-qwen-1.5b	79.15 ± 1.08	68.52 ± 1.56	93.00 ± 0.33	75.48 ± 0.90	82.22 ± 1.29
ThinkEdit-ds-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 Elem.	MATH-Lv1	MATH-Lv5	MATH500
ds-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-ds-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
ds-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-ds-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
ds-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-ds-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
ds-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-ds-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

Cite this work

Chung-En Sun, Ge Yan, Tsui-Wei Weng, "ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models", EMNLP 2025

@article{ThinkEdit,
   title={ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models},
   author={Sun, Chung-En and Yan, Ge and Weng, Tsui-Wei},
   journal={EMNLP},
   year={2025}
}