自蒸馏会削弱大模型推理能力，因抑制了表达不确定性的关键机制。
Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?
- 通过控制上下文丰富度与任务覆盖范围，发现教师模型越丰富，越抑制不确定性表达。
- 在数学推理任务中，自蒸馏导致性能下降最高达40%，尤其影响未知问题的泛化能力。
- 适合关注大模型推理鲁棒性、思考过程可解释性的研究者阅读。
自蒸馏已成为大语言模型的有效后训练范式，常能提升性能并缩短推理路径。然而在数学推理任务中，我们发现其虽缩短响应长度，却会降低性能。究其原因，是自蒸馏抑制了模型在推理过程中对不确定性的表达（即‘认知不确定性表征’）。通过控制实验，改变条件上下文的丰富度和任务覆盖范围，我们发现：当教师模型基于丰富信息进行训练时，会抑制不确定性表达，从而在有限任务覆盖下实现快速域内优化，但损害了域外（OOD）性能——而未知问题恰恰需要通过表达不确定性来调整推理。在Qwen3-1.7B/8B、DeepSeek-Distill-Qwen-7B和Olmo3-7B-Instruct上，性能下降最高达40%。研究强调，在优化正确答案路径的同时，保留适当不确定性表达对提升推理鲁棒性至关重要。
原文摘要 · Abstract (English)
Self-distillation has emerged as an effective post-training paradigm for LLMs, often improving performance while shortening reasoning traces. However, in mathematical reasoning, we find that it can reduce response length while degrading performance. We trace this degradation to the suppression of epistemic verbalization - the model's expression of uncertainty during reasoning. Through controlled experiments varying conditioning context richness and task coverage, we show that conditioning the teacher on rich information suppresses uncertainty expression, enabling rapid in-domain optimization with limited task coverage but harming OOD performance, where unseen problems benefit from expressing uncertainty and adjusting accordingly. Across Qwen3-1.7B/8B, DeepSeek-Distill-Qwen-7B, and Olmo3-7B-Instruct, we observe performance drops of up to 40%. Our findings highlight that exposing appropriate levels of uncertainty is crucial for robust reasoning and underscore the importance of optimizing reasoning behavior beyond merely reinforcing correct answer traces.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。