Hugging Face Daily Papers·2026-09-29 08:00· 4 天前AI 评分44

模型在何处改变主意：用于可验证奖励高效强化学习的 Hindsight-Divergence Localization

Where the Model Changes Its Mind: Hindsight-Divergence Localization for Efficient Reinforcement Learning with Verifiable Rewards

AI 导读

Hindsight-Divergence Localization（HDL）用事后 token log-likelihood 变化选择分支点，以降低 RLVR 中完整轨迹采样成本。