An AI Read Neuroscience Papers and Spotted New Patterns in Animal Brains

In a new arXiv preprint, an AI system that reads neuroscience methods papers and runs its own analysis pipeline end-to-end has surfaced structured, distributed neural representations of larval zebrafish behavior — a finding the system produced without being told which test to run.

SeekBrain (arXiv:2607.29347) is built around a repertoire of analysis recipes extracted from code-paper pairs, so the system can match a research question to a published method, write the analysis code, and execute it across behavioral, neural, and anatomical data. In a zebrafish study, it reported structured, distributed neural representations of behavior; in a mouse decision-making task, it identified a shared axis of regional decoding strength across the brain (arXiv HTML).

The authors evaluate the system on BrainArena, an expert-annotated benchmark of neuroscience analysis tasks, and report it substantially outperforms state-of-the-art agent baselines. They have released the code as AI4NeuroLab/SeekBrain on GitHub.

The strongest caveat is the obvious one: this is a preprint, not a peer-reviewed paper. The benchmark results are self-reported, the real-world studies are demonstrations rather than independent replications, and the workflow has not been audited by outside labs. Whether the system's case-study findings hold up to independent replication remains the open question — and it is the one any reader should carry away.