Training a policy to prefer responses by optimizing only its own low-confidence (high-surprisal) tokens improves alignment over uniform token optimization in SimPO and DPO.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
ConfPO: Exploiting Policy Model Confidence for Critical Token Selection in Preference Optimization
Training a policy to prefer responses by optimizing only its own low-confidence (high-surprisal) tokens improves alignment over uniform token optimization in SimPO and DPO.