rlhf-pairwise-optimal-policies-non-markovian

IN premiseentries/2026/06/21/wiki-Reinforcement_learning_from_human_feedback-chunk-1.md

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

Optimal RLHF policies from pairwise/dueling comparisons are inherently non-Markovian (memory-dependent) — the best action depends on previous events, unlike standard RL settings

Summary

Because human feedback in RLHF comes as head-to-head comparisons rather than absolute scores, the best action at any step depends on what happened earlier, not just the current state. This means a simple memoryless policy is insufficient; the system must track its own history to behave optimally, which makes the control problem harder than in standard RL settings where the present state is enough.

Dependents

These beliefs depend on this one: