Calibrated multiaccuracy shows sharp randomized-vs-deterministic sample complexity gap
Aaroth · x · 2026-11-03
Aaron Roth adds the key takeaway: randomization isn't needed for the closely related multicalibration problem, but for calibrated multiaccuracy there is a sharp polynomial separation in minimax sample complexity between deterministic and randomized predictors — unusual for single-distribution learning. The attached figure shows the group-budget rate landscape: randomized tight rate max(5/2, 2+β), deterministic 8/3+β/3, and multicalibration at 3, all meeting at A ≈ ε^{-1}.
More from Research
- 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
- 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
- PrimaCons benchmark exposes cross-primitive inconsistency in Jev decision-making — Xu Liu · 2026-11-03
- MACR: confidence-guided relay multi-agent framework cuts token overhead 6x — Xu Liu · 2026-11-03