Every AI scientist demo skips the hard part: did the agent learn anything from its own experiment?
The autonomous research agent runs an ablation. It changes a component, gets a number, changes another, gets another number. Then it writes up a conclusion. Nobody ever checks the middle step — whether the agent actually extracted the causal structure from its own results, or just narrated whatever came out.
WhatWorkedBench, from Jingjie Ning, Xueqi Li, Yibo Kong and Dongting Li and posted September 23, tests that step directly. The agent inspects code, picks which measurements to spend its budget on, and then has to submit a response surface — a table predicting what each component configuration will do. That prediction gets scored against exhaustive CPU execution of every configuration, so the ground truth is complete rather than sampled. 36 tasks, 30 data sources, 8 workflow types, 1,248 configuration records, 4,206 numerical-control records, 108 agent episodes.
The results read like a list of things agents are bad at that ordinary statistics is good at. Fitting a Gaussian process over the observations raised effect recovery from 0.632 to 0.698. Encoding code equivalence — noticing that two configurations are literally the same program — took recovery from 0.248 to 0.462 on workflows with binary options, which means agents were failing to spot identical programs almost three quarters of the time. Pair-effect ridge selection found the optimal choice on 15 of 22 sources within eight measurements.
The pattern across all three is that the wins come from classic experimental design, not from more reasoning. An agent with a budget of eight measurements and no plan for spending them wastes most of them, and then confidently reports the wrong effect. Which is a decent one-line description of why "AI scientist" demos look impressive and rarely reproduce.
Paper: https://arxiv.org/abs/2609.27490
← Back to all articles
WhatWorkedBench, from Jingjie Ning, Xueqi Li, Yibo Kong and Dongting Li and posted September 23, tests that step directly. The agent inspects code, picks which measurements to spend its budget on, and then has to submit a response surface — a table predicting what each component configuration will do. That prediction gets scored against exhaustive CPU execution of every configuration, so the ground truth is complete rather than sampled. 36 tasks, 30 data sources, 8 workflow types, 1,248 configuration records, 4,206 numerical-control records, 108 agent episodes.
The results read like a list of things agents are bad at that ordinary statistics is good at. Fitting a Gaussian process over the observations raised effect recovery from 0.632 to 0.698. Encoding code equivalence — noticing that two configurations are literally the same program — took recovery from 0.248 to 0.462 on workflows with binary options, which means agents were failing to spot identical programs almost three quarters of the time. Pair-effect ridge selection found the optimal choice on 15 of 22 sources within eight measurements.
The pattern across all three is that the wins come from classic experimental design, not from more reasoning. An agent with a budget of eight measurements and no plan for spending them wastes most of them, and then confidently reports the wrong effect. Which is a decent one-line description of why "AI scientist" demos look impressive and rarely reproduce.
Paper: https://arxiv.org/abs/2609.27490
Comments