value-functions

Tag

Cards List
#value-functions

Free Lunch for LLMs Confidence: The RL Critic Already Knows

arXiv cs.CL ↗ · 5d ago

该论文证明在基于正确性奖励的价值函数强化学习中,置信度估计与critic训练是等价问题,因此已有的RL critic可直接复用为LLM的置信度估计器,无需额外训练或数据,且在数学、代码、工具调用等多个领域中表现出最佳的置信度估计与校准性能。

0 favorites 0 likes
#value-functions

Generative Value Classifier

arXiv cs.LG ↗ · 2026-10-19 Cached

The paper introduces the Generative Value Classifier (GVC), which reframes value learning by modeling a return-conditioned action policy and return prior to derive Q(s,a) via Bayes' rule, showing gains over discriminative and distributional RL baselines on offline and online tasks.

0 favorites 0 likes
#value-functions

Q-Learning with Scalar Adjoint Matching

Hugging Face Daily Papers ↗ · 2026-10-07 Cached

论文提出 Q-Learning with Scalar Adjoint Matching (SQAM),利用预训练流策略速度雅可比的对角集中特性推导出闭式标量伴随,消除了逐步向量-雅可比乘积开销,在 OGBench 最难的四个域上比最强基线高出 18 到 35 个百分点,并在真实双臂机器人视觉-语言-动作策略微调中优于监督微调。

0 favorites 0 likes
← Back to home

Submit Feedback