Fly Language Model Connects Full Fruit Fly Wiring to a Frozen LLM
The Fly Language Model (FLM) is a public chatbot that links the retained MaleCNS v1.0 fruit fly connectome to a frozen LiquidAI LFM2.5-1.2B-Instruct model. Its system uses all 166,700 nodes and 25,582,938 directed edges as a reservoir. Token embeddings enter the graph through fixed random projections, while the graph, backbone and input/output projections remain unchanged. Training updates only 278,528 readout parameters, about 0.0238% of the backbone’s 1,170,340,608 parameters.
On 32 SmolTalk dialogues containing 1,236 target tokens, the fly readout reduced loss from 1.381995 to 1.359816 nats per token across three seeds. However, a control that sent the same projected input directly to the readout performed slightly better in every seed, with a loss of 1.359328. The reported bootstrap interval did not establish a fly-specific advantage.
Setting the graph edges to zero removed the residual and restored the backbone result, confirming that the graph was active. Relabeling nodes without retraining returned performance close to baseline. The recurrence also rapidly contracts state differences, leaving no long-range memory beyond the backbone’s context. The MIT-licensed software runs locally on Python 3.12, though study artifacts remain private and independent reproduction is not yet possible.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.