Fly Language Model (FLM) Wires the Full Fruit Fly Connectome Into a Frozen 1.2B LLM, and Its Own Controls Show the Wiring Does Not Help

Details

The Fly Language Model (FLM) couples the full MaleCNS v1.0 connectome — 166,700 retained nodes and 25,582,938 directed synaptic edges — to a frozen LiquidAI LFM2.5-1.2B-Instruct backbone. The architecture is a reservoir computer: the graph, backbone, and random projections are all fixed; only a tiny 278,528-parameter readout head (~0.024% of backbone parameters) is trained.\n\nAt each token, the backbone's 2,048-dim embedding is compressed to 128 channels via a fixed Gaussian projection, propagated through the biological graph with recurrence x = tanh(W(0.6x + 0.4Bc)), pooled into 128 bins, and injected as a bounded residual (RMS-capped at 0.25) back into the vocabulary head.\n\nThe headline result looks promising: NLL dropped from 1.38 (frozen backbone) to 1.36 with the fly readout on 32 held-out SmolTalk dialogues. But the controls are damning. A parameter-matched direct-input readout — identical architecture, no fly graph — scored 0.000488 nats/token better in all 3 seeds. The paired bootstrap interval (+0.000005 to +0.00104) offers no evidence for fly-specific gain.\n\nThe repo is MIT-licensed and deployable locally. The author explicitly disclaims being the first connectome language model and acknowledges the negative control result — a rare and honest admission for an attention-seeking release.