The authors target a known failure mode in RLVR and on-policy self-distillation: when the policy cannot sample any valid rollouts, no effective supervision signal exists, so optimization stalls.
Guided self-exploration lets language models teach themselves better reasoning
SEPO uses ground-truth labels as hints to elicit correct in-distribution reasoning from the policy itself, then distills it back, outperforming RLVR and self-distillation baselines.
Chinese Tech
Wenhang Shi · Yiren Chen · Zhe Zhao · Jinhao Dong · Wei Lu · Xiaoyong Du
Renmin University of China · Peking University · Tencent
Research Digest··2 min read
The authors propose Self-Evolution Policy Optimization (SEPO), a reinforcement learning framework in which a large language model uses ground-truth labels as privileged hints to unlock its own exploration, then treats the resulting context-conditioned predictions as a teacher to distill the new solution into the standard policy.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.