FrontierSeptember 12, 2026via MarkTechPost
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
Why it matters
Neuroscience-inspired architecture meets empirical rigor: researchers integrated the complete fruit fly connectome (166K neurons, 25.6M synapses) into a 1.2B model, but their own ablations reveal the biological wiring provides no benefit—a cautionary tale about copying nature without understanding why.
Key signals
- 166,700 neurons and 25.6 million edges from MaleCNS fruit fly connectome integrated
- Frozen LFM2.5-1.2B-Instruct backbone with 278,528 trainable parameters
- 0.0222 nat/token improvement over baseline, but parameter-matched control WITHOUT connectome performs better in all seeds
- Four controls tested; none favor the biological wiring
- 0.6 per token memory bound
- MIT-licensed code available for local reproduction
- Preprint released (date: September 2026)
- Fly Language Model integrates 166,700 neurons and 25.6M edges from fruit fly connectome
- Only 278,528 parameters trained on frozen 1.2B backbone
- 0.0222 nat/token improvement reported, but parameter-matched control without graph performed better across all seeds
- Four controls included to validate approach
- 0.6 nats/token memory bound
- Published as preprint with negative result for the proposed architecture
The hook
A team wired the entire fruit fly brain into an LLM. The controls showed it made things worse.
The Fly Language Model (FLM) drives all 166,700 retained neurons and 25.6 million edges of the MaleCNS fruit fly connectome with token embeddings, then adds a small learned correction to a frozen LFM2.5-1.2B-Instruct backbone. Only 278,528 parameters train. The accompanying preprint reports a 0.0222…