Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?

TL;DR

Self-distillation can improve LLM performance but may degrade reasoning ability by suppressing uncertainty expression, especially affecting out-of-distribution generalization.

Contribution

This paper uncovers how self-distillation suppresses epistemic verbalization, leading to performance drops in mathematical reasoning and OOD tasks.

Findings

Performance drops of up to 40% across multiple models.

Rich conditioning suppresses uncertainty, harming OOD performance.

Proper uncertainty expression is vital for robust reasoning.

Abstract

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.