Improved Bounds for Private and Robust Alignment

ORID r3PKGNlKET · tags icml2026-repro paper-r3PKGNlKET

#StatusPageArtifactClaim excerpt
1VERIFIED 2/201-standard-mle-style-log-loss-near-optimalartifactLemma 3.1 shows that standard MLE-style log loss achieves near-optimal uniform c…
2VERIFIED 2/202-establishes-uniform-convergence-square-lossartifactLemma 3.3 establishes uniform convergence for square loss under joint corruption…
3VERIFIED 2/203-privacy-cost-enters-multiplicative-factorartifactThe privacy cost enters through the multiplicative factor c(epsilon) = (e^epsilo…
4VERIFIED 2/204-two-orderings-corruption-privacy-areartifactTwo orderings of corruption and privacy are analyzed: Corruption-then-LDP (CTL),…
5VERIFIED 2/205-introduces-privxpo-squarexpo-first-onlineartifactThe paper introduces PrivXPO and SquareXPO as the first online algorithms for pr…
6VERIFIED 2/206-offline-algorithms-priv-chi-po-square-chi-po-achartifactOffline algorithms Priv-chi-PO and Square-chi-PO achieve suboptimality gaps of O…

Open logbook index · logbook.json