DRO-REBEL delivers scalable DRO-based online LLM alignment via relative reward regression, with O~(sqrt(d/n)) parameter error bounds and first parametric O~(d/n) rate under preference shift.
Lemma 9(Beck [2017], Theorem 5.25; Existence and uniqueness of minimizer).Let f:E→(−∞,∞] be proper, closed, andσ-strongly convex withσ >0
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
-
Online Distributionally Robust LLM Alignment via Regression to Relative Reward
DRO-REBEL delivers scalable DRO-based online LLM alignment via relative reward regression, with O~(sqrt(d/n)) parameter error bounds and first parametric O~(d/n) rate under preference shift.