Benchmark granularity magnifies ranking attack: 18 task scores quadruple selection loss

sanmikoyejo · x · 2026-10-01

Score granularity drives the benchmark attack: on LiveBench, returning 18 per-task scores instead of one aggregate score almost quadruples the mean model-selection loss after just 8 router submissions.

Original post →

More from Research

Research channel →