Get Started
Home
Topics
Search
Library
Research questionHow can on-policy LLM reasoning turn unreliable skill-derived signals into reliable token-level supervision?On-policy self-distillation needs dense token-level signals, but skill-derived privileged information can be irrelevant or misleading. The difficulty is determining when those signals should guide a reasoning rollout rather than reinforce incorrect behavior.
AI
LLM Pretraining & Post-training
Machine Learning
Reasoning
Reinforcement Learning
Latest papersRecent research connected to this question, newest first.Skill-Conditioned Gated Self-Distillation for LLM ReasoningThe setting uses an experience-derived skill bank and a verifier for LLM mathematical reasoning. Reported evidence covers AIME24, AIME25, and HMMT25, with comparisons against GRPO and answer-conditioned on-policy self-distillation; broader tasks and deployment behavior are not established.research paper · Sep 3, 2026
Related questions
How sparse can token-level supervision be in on-policy post-training without weakening language-model reasoning?How can reinforcement learning give LLMs useful intermediate feedback when rewards reveal only final correctness?How can LLM agents jointly adapt reasoning policies and hierarchical skill libraries during reinforcement learning?Should reasoning LLMs combine on-policy distillation and verifiable-reward RL jointly or sequentially during post-training?