Fruit Fly Connectome Wired Into 1.2B LLM Loses to No-Graph Control

A researcher has embedded the complete MaleCNS v1.0 fruit fly connectome — all 166,700 retained neurons and 25,582,938 directed edges — into a frozen LiquidAI LFM2.5-1.2B-Instruct backbone, producing the Fly Language Model (FLM). The project is MIT-licensed, runs locally on Python 3.12 across MPS, CUDA, or CPU, and requires no API key. The more consequential detail is in the researcher's own results: a parameter-matched control that feeds the same token projection directly into an identical readout — with no connectome graph at all — outperforms the fly-wired variant in every one of three random seeds. The connectome participates mathematically but does not help.

Architecture: reservoir computer meets frozen transformer

FLM's design is a reservoir computer grafted onto a language model. The backbone's 1,170,340,608 parameters are entirely frozen. At each token step, a fixed Gaussian projection compresses the LFM2.5's 2,048-dimensional token embedding down to 128 channels. Each of the 166,700 reservoir nodes then receives one of those channels with a random sign, and the graph updates according to x = tanh(W(0.6x + 0.4Bc)), where W holds incoming-normalized anatomical contact counts from the connectome. The only trainable components are two bias-free matrices — U at 128×128 and V at 2,048×128 — forming a readout of 278,528 parameters total, or 0.0238% of the backbone. Pooled graph states from 128 bins feed through U and V, producing a residual capped at an RMS of 0.25 across vocabulary coordinates before being added to the frozen backbone's logits.

The critical constraint is the forgetting rate. The recurrence contracts initial-state differences by at most 0.6 per token. After 10 tokens, the bound on any retained initial difference is 0.00605; after 20 tokens it collapses to 0.0000366. Packing 166,700 cells into the reservoir buys no long-range memory — context still originates entirely from the frozen transformer backbone.

Benchmark results and what the controls isolate

Testing used 32 held-out SmolTalk everyday-conversation dialogues comprising 1,236 target tokens, evaluated across three random fit seeds.

The fly readout improved on the raw backbone by 0.0222 nats/token, reducing perplexity from 3.98 to 3.90. The direct-input control beat the fly condition in all three seeds by 0.000488 nats/token; the paired bootstrap confidence interval runs from +0.00000502 to +0.00104 — entirely favoring the no-graph control. Zeroing W eliminates the readout residual exactly, confirming the graph does participate in the forward pass. Relabeling node identities without retraining returns NLL to near baseline, demonstrating that the readout depends on its learned alignment to specific graph structure — but that alignment still cannot beat a structureless projection of the same dimensionality.

AI Mastery analysis

FLM is a rigorous proof-of-concept that accidentally proves the opposite of its premise. The reservoir computing framework is theoretically sound: the connectome's fixed recurrent dynamics could in principle project the token embedding into a richer representational space than a linear mapping. The problem is the contraction rate. A per-token forgetting factor of 0.6 means the graph's internal state is almost entirely overwritten within a handful of tokens, leaving no window in which long-range structural properties of the connectome could express themselves. The 25.6 million edges encode connectivity, not sequence — and sequence is what language modeling requires.

The parameter asymmetry is also telling. The 278,528-parameter readout is trained to interface with a 1.17-billion-parameter frozen backbone. That readout has enough capacity to learn a useful correction regardless of what signal it receives — graph state or raw projection. The direct-input control demonstrates this cleanly: the readout is doing the work, not the connectome. This is a pattern production AI failures rooted in architecture rather than model intelligence illustrates well — when a component's contribution is not ablation-tested against a structurally equivalent alternative, capacity effects masquerade as structural insight.

The researcher explicitly disclaims priority, citing ngxson/fly-hf, an earlier prototype using 49,393 central-brain cells trained on TinyStories without a pretrained backbone. FLM's genuine contribution is methodological transparency: the ablation suite is exactly what this class of biostructure-integration experiment requires, even though the results are negative. Study artifacts remain private, so independent reproduction is not currently possible.

Biological plausibility does not confer computational advantage when the inductive bias of the biological structure mismatches the inductive demands of the task. Connectome-scale wiring encodes anatomical contact statistics optimized by evolution for embodied sensorimotor control in a 3-millimeter insect — not for predicting the next token in human conversation. The experiment also shows that scale alone — moving from 49,393 cells in prior work to 166,700 in the full retained MaleCNS graph — does not rescue a structurally mismatched integration strategy. Until researchers can identify a specific structural property of a biological graph that maps to a specific computational primitive a language model lacks, connectome integration will keep losing to direct projection baselines trained on the same token budget.

Primary source

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 — MarkTechPost