Research map / Large language models / Reasoning and chain of thought
Characterizing and Mitigating Reasoning Drift in Large Language Models
Yufeng Zhang, Xuepeng Wang, Lingxiang Wu, Jinqiao Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Analyzes thousands of chain-of-thought reasoning paths to characterize "Reasoning Drift," a failure mode where LLMs lock into flawed reasoning patterns, and proposes mitigation methods.
Problem
Chain-of-thought reasoning in LLMs is unreliable because stochastic generation can lock models into flawed reasoning patterns (reasoning drift).
Approach
Empirically characterizes reasoning drift across thousands of sampled reasoning paths, showing its manifestation is an interplay of factors; mitigation details are not stated in the available abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply reasoning-aware activation steering | R1-Distill-Llama-8B · AIME2024 | accuracy: 34.99 -> 55.56 | Improved |
| apply reasoning-aware activation steering | R1-Distill-Qwen-14B · AIME2025 | accuracy: 26.67 -> 36.67 | Improved |
| apply fine-grained library of steering vectors instead of the single uncertainty-management vector (CAAum) (ablation) | R1-Distill-Qwen-14B · GPQA-Diamond | accuracy: 52.02 -> 57.57 | Improved |
| apply single uncertainty-management vector (CAAum) alone (ablation) | R1-Distill-Qwen-14B · AIME2024 and GPQA-Diamond | accuracy: 54.44 -> 53.33 (AIME2024); 55.55 -> 52.02 (GPQA-D) | Hurt |
| match steering vector to source and target category (vs random source or random target) (ablation) | RAAS · GSM8K, AIME, GPQA-Diamond | accuracy | Improved |
| apply fully random steering vectors (ablation) | RAAS · GSM8K, AIME, GPQA-Diamond | accuracy | Hurt |
| transfer vectors from distilled models to non-distilled base models | Llama3.1-8B · GSM8K | accuracy: 83.91 -> 85.22 | Improved |
| transfer vectors from distilled models to non-distilled base models | Qwen2.5-14B · GPQA-Diamond | accuracy: 34.27 -> 35.18 | Improved |
| apply math-derived steering vectors to non-math tasks | R1-Distill-Qwen-14B · ReClor | accuracy: 65.97 -> 66.51 | Improved |
| apply math-derived steering vectors to non-math tasks | R1-Distill-Llama-8B · LogiQA | accuracy: 58.80 -> 61.00 | Improved |
| apply RAAS vs PiCSAR on AIME2025 | R1-Distill-Llama-8B · AIME2025 | accuracy: 31.70 vs 33.33 | No clear effect |
Closest papers by meaning
- Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring (NeurIPS 2026 · similarity 0.92)
- Probing the Trajectories of Reasoning Traces in Large Language Models (NeurIPS 2026 · similarity 0.91)
- Modeling Hierarchical Thinking in Large Reasoning Models (ICML 2026 · similarity 0.91)
- LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning (NeurIPS 2026 · similarity 0.90)
- Calibrating Reasoning in Language Models with Internal Consistency (NeurIPS 2024 · similarity 0.90)
- From Passive to Active Reasoning: Can Large Language Models Ask the Right Questions under Incomplete Information? (ICML 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
