Iterative RLHF collapses because policies exploit reward-model blind spots in the feedback loop; foresighted policy optimization restores the missing influence term and prevents the collapse.
Both updates use the 8-bit paged AdamW optimizer [Dettmers et al., 2023, Loshchilov and Hutter, 2019] with gradients accumulated over four steps before each optimizer step
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Explaining and Preventing Alignment Collapse in Iterative RLHF
Iterative RLHF collapses because policies exploit reward-model blind spots in the feedback loop; foresighted policy optimization restores the missing influence term and prevents the collapse.