Efficient Agentic LLM

Value Function Estimation

1. Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference ICLR 2025. Paper

   Qining Zhang, Lei Ying

   Moti: Reward function construction bottleneck: RLHF -> DPO -> GRPO.

   Design: Directly apply policy-gradient through ZO-based value function estimation.