Back to /ylecun
s/ylecunAI BENCHMARKS•Nov 9
23
votes
664
seen

NeuroscienceBench tests AI agents on real experimental data

NeuroscienceBench tests AI agents on experimental neuroscience data, not just static prediction tasks. The benchmark asks whether they can connect evidence across neural recordings, microscopy, animal behavior, sequencing, EEG, brain-computer interfaces, imaging, and clinical biomarkers.

That targets a basic problem in biology: a model trained on one layer, such as RNA, can produce findings that fail to generalize across real experimental systems. The first release contains 91 evaluations drawn from 43 peer-reviewed neuroscience studies, with structured answers graded deterministically.

Timeline2
Nov 9

NeuroscienceBench was introduced as the first release in a planned basic-research benchmark series.

Nov 9

The benchmark's rationale was linked to testing whether findings from single biological layers generalize to broader biology.

1 comment
Nov 9
Discussion

1 comment

Sign in to join the discussion