Reinforcement learning from human feedback

Technique · Training · 1 item since Apr 2026

RL using rewards learned from human preference judgments between model outputs.

Organizations: UC Berkeley UT Austin

Items

Apr 15
Value Gradient Flow recasts behavior-regularized RL
UC Berkeley and UT Austin treat the optimization behind RLHF and offline RL as optimal transport solved by discrete gradient flow from the reference distribution, dropping explicit policy parameterization and beating PPO and DPO variants on D4RL, OGBench, and LLM RL tasks.