TasteVal (performance multiplier)
Science · 2026-10-05
TasteVal's performance multiplier: how good the Researcher's final result is relative to the best human expert, whatever compute it took. Each run's official test score is rescaled so the task's naive baseline is 0 and the expert baseline 1, averaged over the seeds of a task with runs below the naive baseline counted as 0, and combined across the 8 tasks by a geometric mean, a task where no run beats the naive baseline entering at 0.01. A value above 1 means the model's final submissions beat the experts'.
Benchmark source: arXiv