The Role of Feedback Alignment in Self-Distillation

Hugging Face Daily Papers Papers

Summary

This paper studies context design for self-distillation in language models, finding that step-aligned critique feedback significantly outperforms binary reward or reference solution conditioning, because it targets only erroneous tokens while preserving correct behavior.

Conditioning a language model on additional context, such as feedback on a previous attempt, typically improves its response. Self-distillation trains the model to retain this improvement when the context is not present. The method works by matching the model's output distribution under two settings: a student that sees only the question, and a self-teacher that also sees the context. What the model learns therefore depends on what context the self-teacher receives, yet the design of this context remains largely unexplored. We study context design for self-distillation by training a solver on feedback from a frozen critic. We compare three conditions: (i) a binary reward (GRPO), (ii) the reference solution, and (iii) a step-by-step critique aligned to the solver's reasoning trace. Step-aligned critique yields the largest gains, outperforming GRPO by 16.11 points and reference-solution-conditioned self-distillation by 5.27 points (Avg@12). Per-token advantage analysis reveals why: step-aligned feedback targets only the tokens where reasoning fails, leaving correct behavior intact. Conditioning on the reference solution, by contrast, pressures the model to change its behavior at every token (even correct steps) because an alternative derivation inevitably differs in phrasing and approach. This suggests that structural alignment between feedback and the solver's reasoning is a key driver of self-distillation effectiveness.
Original Article
View Cached Full Text

Cached at: 06/10/26, 05:46 PM

Paper page - The Role of Feedback Alignment in Self-Distillation

Source: https://huggingface.co/papers/2606.11173 Conditioning a language model on additional context, such as feedback on a previous attempt, typically improves its response. Self-distillation trains the model to retain this improvement when the context is not present. The method works by matching the model’s output distribution under two settings: a student that sees only the question, and a self-teacher that also sees the context. What the model learns therefore depends on what context the self-teacher receives, yet the design of this context remains largely unexplored.

We study context design for self-distillation by training a solver on feedback from a frozen critic. We compare three conditions: (i) a binary reward (GRPO), (ii) the reference solution, and (iii) a step-by-step critique aligned to the solver’s reasoning trace.

Step-aligned critique yields the largest gains, outperforming GRPO by 16.11 points and reference-solution-conditioned self-distillation by 5.27 points (Avg@12). Per-token advantage analysis reveals why: step-aligned feedback targets only the tokens where reasoning fails, leaving correct behavior intact. Conditioning on the reference solution, by contrast, pressures the model to change its behavior at every token (even correct steps) because an alternative derivation inevitably differs in phrasing and approach. This suggests that structural alignment between feedback and the solver’s reasoning is a key driver of self-distillation effectiveness.

Similar Articles

EchoDistill:Alignment Noisy-to-Clean Self-Distillation for Robust Audio LLMs

arXiv cs.CL

EchoDistill is an alignment-based noisy-to-clean self-distillation framework that improves the robustness of Audio Large Language Models (ALLMs) against real-world noise by using a frozen clean-audio teacher to guide the student model via group-relative policy optimization (GRPO). Experiments show significant improvements in semantic reliability and task performance under strong noise without additional inference costs.

Self-Distillation Enables Continual Learning [pdf]

Hacker News Top

Introduces Self-Distillation Fine-Tuning (SDFT), a method that enables on-policy learning from demonstrations to achieve continual learning without catastrophic forgetting, outperforming supervised fine-tuning.