SeekBrain: a 29-author multi-agent system that builds its analysis recipes from code-paper pairs

arXiv·medium signalXBlueskyLinkedInCopy link

Posted July 31 (arXiv 2607.29347, lead author Jiamin Wu), SeekBrain targets fragmented neuroscience workflows by dynamically constructing a repertoire of analysis recipes extracted from code-paper pairs, then using domain-grounded hierarchical planning to generate hypotheses and analytical pipelines across heterogeneous datasets. It reports substantial gains over agent baselines on the BrainArena benchmark, and in real deployment integrated behavioral, neural and anatomical data to surface structured distributed neural representations of larval zebrafish behavior. The transferable idea for builders is the recipe-extraction step: mining executable procedures from paired code and prose is a general way to give a domain agent priors that a system prompt cannot encode.