Paper 2603.24472
Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?
- Published
- Mar 2026
- Research lab
- Independent
- Citations
- 61
- GitHub
- 75 stars
01 In brief
Summary
This paper investigates why self-distillation can degrade the mathematical reasoning performance of LLMs, despite often improving other domains.
The authors find that self-distillation suppresses 'epistemic verbalization'—the expression of uncertainty during reasoning—which is crucial for robust problem-solving.
Through controlled experiments, they show that conditioning a teacher model on richer information (e.g., full solutions) leads to more concise, confident reasoning with fewer uncertainty markers, which can harm out-of-distribution (OOD) performance.
The degradation is amplified when task coverage is broad, as unseen problems benefit from uncertainty expression.
Across models like Qwen3-1.7B/8B, DeepSeek-Distill-Qwen-7B, and Olmo3-7B-Instruct, performance drops up to 40% on benchmarks like AIME24.
The study highlights that optimizing for correct answers alone is insufficient; preserving uncertainty-aware reasoning is essential for generalization.
02 From the paper
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.