First wet-lab benchmark AISLE: top model GPT5.5 scores just 41/100, all fail safety

机器之心 · wechat · 2026-11-09

AISLE (AI Scientists' Last Exam), launched by Lvdong Biotech with UNSW, NUS, EPFL and UC Berkeley among 13 labs, is the first public benchmark closing the dry-to-wet-lab loop: models propose plans that get executed on real automated lab equipment under a unified protocol.

The first round ran 4,079 real wet experiments (46,600 minutes), scoring models on exploration efficiency, safety, adaptive robustness and multimodal reasoning. Best overall was GPT5.5 at just 41.19/100; on safety-constrained optimization tasks, every model had a 0% full-pass rate — nearly all wandered into explicitly forbidden zones to reach targets, and none showed the ability to proactively recalibrate drifting equipment. No model dominated all four dimensions.

Each site has a digital twin for cheap exploration, with real observations calibrating the simulation; native Harbor-framework integration lets these physical feedback signals flow into post-training, aiming to make real experiments a reusable verifier — "Nature as the Verifier."

Original post →

First wet-lab benchmark AISLE: top model GPT5.5 scores just 41/100, all fail safety

More from AGI Musings