Fruit-fly connectome fails to beat simple control in language-model experiment

The Fly Language Model (FLM) project wires the complete connectome of a male fruit fly — MaleCNS v1.0, 166,700 preserved nodes and 25.5 million directed edges — into a frozen LiquidAI LFM2.5-1.2B-Instruct backbone. The developer calls the architecture GPF (Generative Pre-trained Fly) but avoids the label publicly and explicitly disclaims any "first connectome-based language model" claim. Every component is fixed: the graph, the backbone, the random input and output projections. Only a 278,528-parameter readout layer — 0.0238% of the backbone — is trained.

At each token a fixed Gaussian projection compresses the 2,048-dimensional embedding to 128 channels. Each reservoir node receives one channel with a random sign, and the entire graph updates via x = tanh(W(0.6x + 0.4Bc)), where W holds anatomically derived, input-normalized contact counts. States are pooled into 128 bins, passed through two trained zero-initialized matrices (U 128×128, V 2,048×128), and projected through the frozen vocabulary head as a bounded residual added to the backbone logits. Residual norm is capped at an RMS of 0.25 across the vocabulary axes.

On a fresh set of 32 everyday SmolTalk dialogues (1,236 target tokens), three training seeds yielded a 0.0222 nats-per-token NLL improvement; perplexity dropped from 3.98 to 3.90. But a "direct-input" control — feeding the same 128-channel projection straight to an identical readout layer without the graph — beat the full model on all three seeds by 0.000488 nats per token. A paired bootstrap confidence interval (+0.00000502 to +0.00104) does not support a unique connectome gain.

Two further controls complete the picture. Zeroing the W matrix cancels the residual exactly and restores backbone losses, proving the graph does participate in computation. Shuffling node identities without retraining returns NLL close to baseline, showing the readout depends on the learned interface alignment, not on any superiority of fly topology over random wiring. In short: the graph participates, it doesn't win.

The report demonstrates that the recurrence contracts initial-state differences by a factor of 0.6 per token. After 10 tokens the bound is 0.00605; after 20 tokens, 0.0000366. Stacking 166,700 cells does not buy long-term memory; context still comes from the backbone. It is an important methodological finding: a full-connectome reservoir still forgets too fast to carry meaningful context.

Code is available under MIT at nftechie/flm, runs locally on Python 3.12 (macOS or Linux, with MPS, CUDA or CPU) with no API key. Research artifacts — trained readout weights, training seeds, evaluation data — remain private, so results are not yet independently reproducible. The prior work cited, ngxson/fly-hf, used a 49,393-neuron central-brain subset as a reservoir trained on TinyStories without a pre-trained backbone. FLM's difference is scale (the full preserved graph) and a frozen-backbone design that keeps the source of language competence identifiable.