Self-distillation can harm LLM reasoning by suppressing uncertainty, study finds

A new research paper explores how self-distillation, a technique used to improve large language models (LLMs), can sometimes degrade their mathematical reasoning capabilities. The study, published on arXiv, found that this degradation occurs because the process suppresses "epistemic verbalization," where models express uncertainty during reasoning. By controlling the richness of conditioning context and task coverage, researchers demonstrated that suppressing uncertainty expression can lead to rapid in-domain optimization but harms out-of-domain performance. This effect was observed across several models, including Qwen3 and Olmo3, with performance drops of up to 40%. The findings emphasize the importance of allowing models to express uncertainty for robust reasoning, rather than solely focusing on reinforcing correct answer traces. AI

IMPACT This research highlights a potential drawback of self-distillation in LLMs, suggesting that optimizing for correct answers may inadvertently reduce robustness by suppressing uncertainty expression.

RANK_REASON Research paper published on arXiv detailing findings about LLM capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• DeepSeek-Distill-Qwen-7B
• Jeonghye Kim
• LLMs
• Olmo3-7B-Instruct
• Qwen3
• Self-distillation

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →