rlhf-taxonomy-policy-gradient

IN premise — entries/2026/06/21/wiki-Reinforcement_learning_from_human_feedback-chunk-5.md

Created 2026-06-21T09:50:10+00:00

RLHF is classified under Reinforcement Learning → Policy Gradient methods in the AI taxonomy, not under supervised or self-supervised learning.

Summary

RLHF works by having the model act as an agent that improves its behavior based on feedback signals, which is fundamentally a reinforcement learning strategy rather than a case of learning from labeled examples or finding patterns in raw data. This matters because the model inherits the risks and guarantees of policy gradient methods—like reward hacking and high variance in updates—so the system should reason about its behavior using those assumptions rather than treating it as a more straightforward fine-tuning step.