Weak-to-strong reward models succeed in-distribution but fail to transfer under preference shift due to source-domain feature pull; Representation Anchoring regularizer improves OOD performance.
InAdvances in neural information processing systems, volume 36, pages 53728–53741
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
When In-Distribution Gains Fail: Evaluating Weak-to-Strong Reward Models under Preference Shift
Weak-to-strong reward models succeed in-distribution but fail to transfer under preference shift due to source-domain feature pull; Representation Anchoring regularizer improves OOD performance.