Fly Language Model Ties Full Fruit Fly Connectome to Frozen 1.2B LLM; Controls Show No Gain From Wiring

The system is a reservoir computer bolted onto a language model. All 166,700 retained nodes and 25,582,938 directed edges of the MaleCNS graph participate. The graph, the backbone, and the random input and output projections are all fixed. Only a 278,528-parameter readout is trained, about 0.0238% of the 1,170,340,608 backbone parameters. At each token, a fixed Gaussian projection compresses the 2,048-dimensional token embedding to 128 channels. Each reservoir node receives one channel with a random sign. The whole graph then updates with x = tanh(W(0.6x + 0.4Bc)), where W holds incoming-normalized anatomical contact counts. States are pooled into 128 bins, passed through two trained bias-free matrices, U at 128 by 128 and V at 2,048 by 128, and projected through the frozen vocabulary head as a bounded residual added to the backbone logits. The residual is capped at an RMS of 0.25 across vocabulary coordinates.

On a freshly frozen set of 32 SmolTalk everyday-conversation dialogues with 1,236 target tokens, three fit seeds gave the following results. The frozen backbone scored 1.381995 nats per token. The fly readout scored 1.359816 plus or minus 0.000110. The direct-input readout scored 1.359328 plus or minus 0.000108. The relabeled, no-refit condition scored 1.381265 plus or minus 0.000802. The no-edges condition scored 1.381995. The fly readout improved on the backbone by 0.0222 nats per token, with perplexity moving from 3.98 to 3.90. But a direct-input control, which feeds the same 128-channel token projection straight into an identical readout with no graph, did better in all three seeds by 0.000488 nats per token. The paired bootstrap interval of +0.00000502 to +0.00104 does not support a fly-specific gain.

Two other controls matter. Setting W to zero removes the residual exactly, reproducing the backbone's per-token losses, so the graph verifiably participates. Relabeling node identities without retraining returns NLL near baseline, which shows the readout depends on its learned interface alignment, not that fly topology beats random wiring. The research report also proves the recurrence contracts initial-state differences by at most 0.6 per token. After 10 tokens that bound is 0.00605; after 20 it is 0.0000366. Piling in 166,700 cells does not buy long memory, and context still comes from the backbone.

The research report cites ngxson/fly-hf, an earlier prototype that used a 49,393-cell central-brain subset of MaleCNS as a reservoir trained on TinyStories without a pretrained backbone, and states plainly that it makes no claim to be the first connectome-based language model. FLM's distinction is scale, using the full retained graph, and the frozen-backbone design that keeps the source of language competence identifiable. The nftechie/flm repository is MIT-licensed and runs on Python 3.12 on macOS or Linux with MPS, CUDA, or CPU, and no API key is required. Study artifacts stay private, so results are not independently reproducible yet.