Multicalibration is often overkill: calibrated multiaccuracy needs fewer samples
Aaroth · x · 2026-11-03
Aaron Roth argues multicalibration, while useful for omniprediction, fast agreement and hard-core-set construction, is often overkill — many applications only need global calibration plus multiaccuracy. The attached paper (Calibrated Multiaccuracy: Deterministic and Randomized Sample Complexity, by Lu, Noarov, Roth, Nov 3 2026) resolves its sample complexity: Õ(log|G|/ε² + 1/ε^{5/2}) for randomized predictors vs Õ(log|G|/ε² + (log|G|)^{1/3}/ε^{8/3}) for deterministic ones, with a polynomial separation when log|G| ≤ ε^{-1+κ}, and the cube-root dependence on log|G| shown to be intrinsic.
More from Research
- mlx-learn ships scikit-learn API on Apple Silicon with up to 300x faster inference — HankYeomans · 2026-11-03
- CMU launches first Protostar Challenge on climate risk AI, open until Nov 13 — evijit · 2026-11-03
- Tivadar Danka publishes a complete math roadmap for machine learning: linear algebra, calculus, probability — TivadarDanka · 2026-11-03
- Formalizing CFSG Uncovers 96 Math Errors in 47 Peer-Reviewed Papers, Including Annals of Mathematics — skdh · 2026-11-03
- SEC-Bench Pro: frontier labs test agents on 183 real V8 and SpiderMonkey vulnerabilities — NielsRogge · 2026-11-03
- PhD thesis from Aengus Lynch shows aligned AI stays vulnerable: 89% jailbreak, 96% blackmail — StephenLCasper · 2026-11-03