Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?
This addresses a problem for LLM developers and researchers by revealing that self-distillation can harm out-of-domain robustness in reasoning tasks, highlighting the need to optimize reasoning behavior beyond just reinforcing correct answers.
The paper investigates why self-distillation sometimes degrades the reasoning capability of LLMs, particularly in mathematical reasoning, where it reduces response length and performance by suppressing epistemic verbalization (uncertainty expression), leading to performance drops of up to 40% in models like Qwen3-8B.
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 Qwen3-8B, 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.