An unusual experiment has produced a clean negative result. The Fly Language Model couples the full MaleCNS fruit fly connectome, 166,700 nodes of it, to a frozen LiquidAI LFM2.5-1.2B backbone.
Only 278,528 parameters were trained. The fly readout trimmed next-token loss by 0.0222 nats per token, a small gain over the frozen model. The trouble is the controls: a version with no graph at all beat the connectome build in every seed.
Other checks reinforced the point. Disconnecting the graph zeroed the residual exactly, and relabeling broke it, so the wiring really does sit in the compute path. It simply does not help, and its state forgets at about 0.6 per token, adding no long-range memory.
The developer makes no priority claim. The write-up explicitly declines the label of first connectome language model and keeps the language backbone frozen so its own contribution stays identifiable.
Code runs locally on Python 3.12, though the study artifacts stay private, so the finding is not yet independently reproducible.
