Fly Language Model plugs a full fruit fly connectome into a 1.2B LLM, but tests show no performance gain

Researchers have introduced Fly Language Model, or FLM, a system that connects the full connectome of the male fruit fly central nervous system to a frozen LiquidAI LFM2.5-1.2B-Instruct language model backbone. The connectome used in the experiment contains 166,700 nodes and 25.58 million directed edges, while training was limited to a 278,000-parameter readout layer, or about 0.0238% of the backbone parameters. In tests across 32 everyday dialogue datasets, the connectome-augmented model reduced per-token negative log-likelihood by 0.0222 nats. Still, a direct-input control model without the connectivity graph performed better across all three random seeds by 0.000488 nats. The study said the connectome did take part in computation, but it did not produce a connectome-specific gain. Its state memory also decayed by 0.6 per token, which meant it failed to provide long-term memory. The developers said they do not claim FLM is the first connectome-based language model. They described its distinguishing features as the use of a complete connectome graph and a frozen-backbone design. The code has been released under the MIT license and can run locally in a Python 3.12 environment, according to MarkTechPost.