MaPPO incorporates prior reward knowledge into a Maximum a Posteriori objective for LLM preference optimization, generalizing DPO and variants while supporting offline and online settings.
(29) In contrast, in DPO, the gradient is Lipschitz continuous as ∥τθ−τθ′∥≤LDPO∥θ−θ′∥, (30) where LDPO = 2β(1−σ(u))Mg < 2βMg
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
MaPPO: Maximum a Posteriori Preference Optimization with Prior Knowledge
MaPPO incorporates prior reward knowledge into a Maximum a Posteriori objective for LLM preference optimization, generalizing DPO and variants while supporting offline and online settings.