Introduces WhatWorkedBench, a benchmark measuring whether AI research agents can accurately predict how experimental component changes affect outcomes after budgeted experimentation, using 36 tasks across 30 data sources and 8 workflow types with exhaustively computed reference effects. Agents inspect code, choose measurements, and submit a response-surface table predicting scores across configurations; the benchmark evaluates over 4,200 numerical-control records and 108 agent episodes. The authors show that fitting a Gaussian process to agent observations substantially improves effect-recovery accuracy over simpler baselines like pair-effect ridge regression.
