Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes
Paper • 2603.25562 • Published • 19
Papers LLM training tricks
Note Tricks to make On-Policy Distillation better
-
Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe
Paper • 2604.13016 • Published • 108 -
From P(y|x) to P(y): Investigating Reinforcement Learning in Pre-train Space
Paper • 2604.14142 • Published • 30
TIP: Token Importance in On-Policy Distillation
Paper • 2604.14084 • Published • 15Note Informative tokens come from two regions: positions with high student entropy, and positions with low student entropy plus high teacher–student divergence, where the student is overconfident and wrong
-
Co-Evolving Policy Distillation
Paper • 2604.27083 • Published • 67 -
Beyond SFT-to-RL: Pre-alignment via Black-Box On-Policy Distillation for Multimodal RL
Paper • 2604.28123 • Published • 49 -
Efficient Pre-Training with Token Superposition
Paper • 2605.06546 • Published • 46
Self-Distilled Agentic Reinforcement Learning
Paper • 2605.15155 • Published • 111Note Gating for on-policy self distillation