Fly Connectome Language Model Shows No Performance Gain

The Fly Language Model (FLM), developed by an individual using the moniker "the developer," has integrated the complete retained MaleCNS v1.0 fruit fly connectome into a frozen LiquidAI LFM2.5-1.2B-Instruct large language model (LLM). This project, which the developer claims is the world's first Fly Language Model built on a Generative Pre-trained Fly (GPF) architecture, does not explicitly claim to be the first connectome language model. Instead, it reports that a parameter-matched control model without the fly graph actually performs slightly better. The FLM is deployable locally, with its nftechie/flm repository available under an MIT license. It runs on Python 3.12 for macOS or Linux, supporting MPS, CUDA, or CPU processing without requiring an API key. The system functions as a reservoir computer connected to the LLM. It incorporates all 166,700 retained nodes and 25,582,938 directed edges of the MaleCNS graph. Both the graph and the LLM backbone are fixed, as are the random input and output projections. Only a 278,528-parameter readout layer is trained, representing approximately 0.0238% of the 1,170,340,608 parameters in the LLM backbone. At each token processing step, a fixed Gaussian projection reduces the 2,048-dimensional token embedding to 128 channels. Each reservoir node receives a single channel with a random sign. The entire graph then updates according to the equation x = tanh(W(0.6x + 0.4Bc)), where W represents incoming-normalized anatomical contact counts. States are aggregated into 128 bins, processed through two trained bias-free matrices (U at 128x128 and V at 2,048x128), and then projected through the frozen vocabulary head. This projection is applied as a bounded residual, added to the LLM's logits, and capped at an RMS of 0.25 across vocabulary coordinates. To evaluate its performance, the FLM was tested on a set of 32 SmolTalk everyday-conversation dialogues, totaling 1,236 target tokens. Three separate runs with different initial seeds for the "Fly readout" condition yielded an average negative log-likelihood (NLL) of 1.359816 ± 0.000110 nats per token. In comparison, the frozen LLM backbone alone achieved an NLL of 1.381995 nats per token, indicating a perplexity of 3.98. The "Direct-input readout" condition, which bypasses the fly graph and directly inputs data to the readout, achieved a slightly better NLL of 1.359328 ± 0.000108. A "Relabeled, no refit" condition, where the fly graph was relabeled but not refitted, resulted in an NLL of 1.381265 ± 0.000802. Critically, a condition with "No edges" in the fly graph, effectively disabling its structural influence while retaining the node processing, also resulted in an NLL of 1.381995, identical to the performance of the frozen backbone. This indicates that the complex wiring of the fruit fly connectome, when integrated in this manner, did not provide a performance benefit to the language model. The fly readout condition did show a marginal improvement of 0.0222 nats per token over the frozen backbone, but this was surpassed by the direct-input readout and was not demonstrably linked to the fly's neural architecture.