WhatWorkedBench: Benchmarking Experimental Understanding in AI Agents
AI research agents need reliable knowledge of how their experiments change outcomes. We introduce WhatWorkedBench to measure experimental understanding, the accuracy of predictions...