Trustworthy-ML-Lab/ThinkEdit
[EMNLP 25] An effective and interpretable weight-editing method for mitigating overly short reasoning in LLMs, and a mechanistic study uncovering how reasoning length is encoded in the model’s representation space.
This project helps improve the performance of large language models (LLMs) on complex reasoning tasks by addressing overly short reasoning. It takes an existing LLM's responses to problems and identifies specific internal components responsible for insufficient 'thinking' steps. The output is a modified LLM that generates more complete and accurate reasoning, benefiting anyone using LLMs for tasks requiring detailed, multi-step problem-solving.
Use this if your LLM is producing correct answers on mathematical or logical problems but often skips intermediate steps, leading to less reliable or less transparent results.
Not ideal if you need to debug or modify an LLM's behavior for issues unrelated to reasoning length, such as factual inaccuracies, stylistic preferences, or ethical concerns.
Stars
17
Forks
1
Language
Python
License
—
Category
Last pushed
Dec 17, 2025
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/transformers/Trustworthy-ML-Lab/ThinkEdit"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
steering-vectors/steering-vectors
Steering vectors for transformer language models in Pytorch / Huggingface
jianghoucheng/AlphaEdit
AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models, ICLR 2025 (Outstanding Paper)
kmeng01/memit
Mass-editing thousands of facts into a transformer memory (ICLR 2023)
boyiwei/alignment-attribution-code
[ICML 2024] Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications
jianghoucheng/AnyEdit
AnyEdit: Edit Any Knowledge Encoded in Language Models, ICML 2025