A hybrid online-plus-offline preference optimization algorithm, HPO, provably needs fewer samples than pure online or offline RLHF in linear MDP settings.
Apply Azuma Hoeffding with its sample average overt + γ terms
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Hybrid Preference Optimization for Alignment: Provably Faster Convergence Rates by Combining Offline Preferences with Online Exploration
A hybrid online-plus-offline preference optimization algorithm, HPO, provably needs fewer samples than pure online or offline RLHF in linear MDP settings.