NLHF achieves the minimax-optimal worst-case average-utility distortion (1/2+o(1))β, while RLHF and DPO can suffer distortion up to e^{Ω(β)} or unbounded under certain comparison sampling.
Approximating optimal social choice under metric preferences
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
ACCEPT 1representative citing papers
citing papers explorer
-
Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences?
NLHF achieves the minimax-optimal worst-case average-utility distortion (1/2+o(1))β, while RLHF and DPO can suffer distortion up to e^{Ω(β)} or unbounded under certain comparison sampling.