Fly Connectome Model Fails to Beat Simpler Control

The Fly Language Model (FLM) integrates the full 166,700-node fruit fly connectome into a frozen 1.2B-parameter language model, training only a 278,528-parameter readout. While the fly graph reduced perplexity, a direct-input control without the connectome performed better in all test seeds. The study concludes the biological wiring does not provide a performance advantage over a simpler architecture.