Can AI Scientists Discover Neural Mechanisms? Evaluating Agentic Biological Discovery in a Digital Fly.

Autonomous scientific agents are beginning to move beyond narrow assistance roles toward systems that can generate hypotheses, run experiments, and draft manuscripts. Yet the current evaluation still says little about whether such systems can uncover biological mechanisms rather than merely summarize correlations or orchestrate known pipelines. We propose a benchmark for agentic biological discovery focused on uncovering neural mechanisms in a digital fly, and we instantiate it as a pilot study across six tasks from a whole-brain Drosophila model. The benchmark casts discovery as a budgeted hypothesis--experiment--update loop in which an agent observes compressed neural and behavioral readouts, selects interventions, and outputs a mechanistic explanation together with a held-out counterfactual prediction. On feeding and grooming tasks, a planning agent reaches a mean Node F1 of 0.856/0.952/1.000/1.000 at budgets 2/4/6/8, outperforming a one-shot GPT-5.4 stand-in (0.532), a no-memory variant (0.578/0.667/0.667/0.667), and a memory-only variant (0.789/0.897/0.897/0.897). In a named-versus-anonymized test, the one-shot baseline drops from 1.000 to 0.532 Node F1 whereas planning remains at 1.000 in both settings. We interpret these results as a pilot benchmark sensitivity study rather than a definitive leaderboard, but they suggest that the digital-fly setting can reveal meaningful differences in iterative mechanistic reasoning, memory, and shortcut robustness.