This paper introduces Self-Distilled Policy Gradient (SDPG), a novel framework that enhances sparse-reward reinforcement learning through on-policy self-distillation. SDPG integrates group-relative verifier advantages, exact full-vocabulary self-distillation, and KL regularization, demonstrating improved stability and performance over existing baselines.
Self-Distillation Zero (SD-Zero) is a novel post-training method designed to be more training sample-efficient than traditional reinforcement learning, without requiring external teachers or high-quality demonstrations. It operates by having a single model act as both a Generator and a Reviser, using the Reviser's improved responses and token distributions to provide dense supervision for the Generator through on-policy self-distillation.
EchoDistill is an alignment-based self-distillation framework designed to make Audio Large Language Models (ALLMs) robust to real-world noise. It leverages a frozen clean-audio teacher to guide an inference-time noisy-audio student, optimizing responses via group-relative policy optimization and token-level consistency.
This content from the Hugging Face Journal Club discusses an "embarrassingly simple" self-distillation method that significantly improves code generation. It highlights advancements in leveraging large language models for programming tasks.
