Institution

UCLA

Public research university in Los Angeles, USA; its machine learning groups work on optimization, LLM reasoning, and reinforcement learning theory.

Reinforcement Learning · UCLA

SDPG: Self-Distilled Policy Gradient for Sparse-Reward RL

SDPG adds a full-vocabulary self-distillation loss to verifier RL, learning from a hint-conditioned teacher. On Qwen3-4B it lifts AIME25 from 0.242 (GRPO) to 0.327 and AMC23 from 0.714 to 0.870 last-checkpoint.