ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models

• URL: http://arxiv.org/abs/2503.22048v2
• Date: Fri, 04 Apr 2025 21:17:46 GMT
• Title: ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models
• Authors: Chung-En Sun, Ge Yan, Tsui-Wei Weng,
• Abstract summary: This work investigates how reasoning length is embedded in the hidden representations of reasoning models.<n>We introduce ThinkEdit, a simple yet effective weight-editing approach to mitigate the issue of overly short reasoning.
• Score: 16.407923457296235
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Recent studies have shown that Large Language Models (LLMs) augmented with chain-of-thought (CoT) reasoning demonstrate impressive problem-solving abilities. However, in this work, we identify a recurring issue where these models occasionally generate overly short reasoning, leading to degraded performance on even simple mathematical problems. Specifically, we investigate how reasoning length is embedded in the hidden representations of reasoning models and its impact on accuracy. Our analysis reveals that reasoning length is governed by a linear direction in the representation space, allowing us to induce overly short reasoning by steering the model along this direction. Building on this insight, we introduce ThinkEdit, a simple yet effective weight-editing approach to mitigate the issue of overly short reasoning. We first identify a small subset of attention heads (approximately 2%) that predominantly drive short reasoning behavior. We then edit the output projection weights of these heads to suppress the short reasoning direction. With changes to only 0.1% of the model's parameters, ThinkEdit effectively reduces overly short reasoning and yields notable accuracy gains for short reasoning outputs (+5.44%), along with an overall improvement across multiple math benchmarks (+2.43%). Our findings provide new mechanistic insights into how reasoning length is controlled within LLMs and highlight the potential of fine-grained model interventions to improve reasoning quality. Our code is available at https://github.com/Trustworthy-ML-Lab/ThinkEdit

Related papers

• Between Underthinking and Overthinking: An Empirical Study of Reasoning Length and correctness in LLMs [52.405085773954596]
  We find that large language models (LLMs) tend to overthink simple problems, generating unnecessarily long outputs, and underthink harder ones. This indicates that models might misjudge problem difficulty and fail to calibrate their response length appropriately. Experiments show that the generation length can be significantly reduced while maintaining acceptable accuracy.
  arXiv  Detail & Related papers  (2025-04-30T18:48:06Z)
• ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning [1.170732359523702]
  Reasoning models such as OpenAI o3 and DeepSeek-R1 have demonstrated strong performance on reasoning-intensive tasks. Long reasoning traces can facilitate a more thorough exploration of solution paths for complex problems. We introduce ShorterBetter, a simple yet effective reinforcement learning methed that enables reasoning language models to discover their own optimal CoT lengths.
  arXiv  Detail & Related papers  (2025-04-30T07:04:19Z)
• Efficient Reasoning for LLMs through Speculative Chain-of-Thought [44.76494056102963]
  Large reasoning language models such as OpenAI-o1 and Deepseek-R1 have attracted widespread attention due to their impressive task-solving abilities. Existing methods for efficient reasoning mainly focus on reducing the number of model parameters or shortening the chain-of-thought length. We introduce Speculative Chain-of-Thought (SCoT), which reduces reasoning latency from another perspective by accelerated average reasoning speed.
  arXiv  Detail & Related papers  (2025-04-27T03:56:39Z)
• SEAL: Steerable Reasoning Calibration of Large Language Models for Free [58.190800043449336]
  Large Language Models (LLMs) have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism. Recent studies reveal substantial redundancy in the CoT reasoning traces, which negatively impacts model performance. We introduce SEAL, a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains.
  arXiv  Detail & Related papers  (2025-04-07T02:42:07Z)
• Process or Result? Manipulated Ending Tokens Can Mislead Reasoning LLMs to Ignore the Correct Reasoning Steps [39.759594479826454]
  We explore how vulnerable are reasoning models to subtle errors in their input reasoning chains.<n>We introduce "Compromising Thought" (CPT), a vulnerability where models presented with reasoning tokens containing manipulated calculation results tend to ignore correct reasoning steps and adopt incorrect results instead.<n>Our work enhances understanding of reasoning robustness and highlights security considerations for reasoning-intensive applications.
  arXiv  Detail & Related papers  (2025-03-25T03:43:11Z)
• Can Large Language Models Detect Errors in Long Chain-of-Thought Reasoning? [57.17826305464394]
  o1-like models produce the long Chain-of-Thought (CoT) reasoning steps to improve the reasoning abilities of existing Large Language Models (LLMs)<n>We introduce the DeltaBench, including the generated long CoTs from different o1-like models for different reasoning tasks.<n>Based on DeltaBench, we first perform fine-grained analysis of the generated long CoTs to discover the effectiveness and efficiency of different o1-like models.
  arXiv  Detail & Related papers  (2025-02-26T17:59:27Z)
• Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning [113.49074603075032]
  Recent studies have shown that making a model spend more time thinking through longer Chain of Thoughts (CoTs) enables it to gain significant improvements in complex reasoning tasks.<n>We explore whether scaling with longer CoTs can indeed impair the reasoning performance of Large Language Models (LLMs) in certain domains.
  arXiv  Detail & Related papers  (2025-02-25T10:48:05Z)
• The Relationship Between Reasoning and Performance in Large Language Models -- o3 (mini) Thinks Harder, Not Longer [1.474723404975345]
  We analyze chain-of-thought length across o1-mini and o3-mini variants on the Omni-MATH benchmark.<n>We find that o3-mini (m) achieves superior accuracy without requiring longer reasoning chains than o1-mini.<n>This accuracy drop is significantly smaller in more proficient models, suggesting that new generations of reasoning models use test-time compute more effectively.
  arXiv  Detail & Related papers  (2025-02-21T17:59:13Z)
• LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters! [53.84130385074551]
  Large reasoning models (LRMs) tackle complex reasoning problems by following long chain-of-thoughts (Long CoT)<n>We find that a Large Language model (LLM) can effectively learn Long CoT reasoning through data-efficient supervised fine-tuning (SFT) and parameter-efficient low-rank adaptation (LoRA)<n>With just 17k long CoT training samples, the Qwen2.5-32B-Instruct model achieves significant improvements on a wide range of math and coding benchmarks.
  arXiv  Detail & Related papers  (2025-02-11T08:48:48Z)
• O1-Pruner: Length-Harmonizing Fine-Tuning for O1-Like Reasoning Pruning [98.3430004984531]
  We propose Length-Harmonizing Fine-Tuning (O1-Pruner) to minimize reasoning overhead while maintaining accuracy.<n>Our code is coming soon at https://github.com/StarDewXXX/O1-Pruner.
  arXiv  Detail & Related papers  (2025-01-22T01:35:11Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.