TwinLab-Bench

AI scientists design their own experiments,
freely interacting with twins of real laboratories.

GitHub Paper Leaderboard
52tasks100+action types45/ 100, the best model
01The film

Inside a twin lab

02Four of the 52

PhD-level problems, handed over by real labs

Each sheet follows one run of Opus 5.5: what it looked at, what it thought, the experiment it ran in the twin lab.

03Task sources

Each task targets a finding of a real lab

The tasks come from journal papers and lab notebooks.

04How a task is run

One run, with three agents

1 / 6
05Results

Frontier models fail

The best scores 45 out of 100. Mean score against the cost of the actor agent on the full benchmark.

06Where it sits

Two axes for science benchmarks

07Scorecard

Against other science benchmarks

Task Dashboard

52 tasks · one run of Opus 5.5 each ·