Fly Language Model (FLM) Wires the Full Fruit Fly Connectome Into a Frozen 1.2B LLM, and Its Own Controls S…
Quick take
The Fly Language Model (FLM) attempts a radical fusion: it encodes the entire MaleCNS fruit fly connectome into a frozen 1.2 billion parameter language model. All 166,700 neurons and 25.6 million synaptic connections become token embeddings, powering a small learned correction on top of a frozen LFM2.5-1.2B-Instruct model. The fine-tuned component trains just 278,528 parameters and claims a modest 0.0222 nat per token improvement over the backbone. But four carefully matched controls—notably one without the connectome graph—outperform the FLM in every seed, showing the wiring data offers no real benefit.
Why it matters
This experiment puts biology-inspired neural wiring into a standard large language model framework and finds it does not improve performance. That challenges some assumptions about the direct value of detailed biological wiring maps for scaling or enhancing language models. For builders chasing novel architectures or new data sources to boost model efficiency or accuracy, the Fly Language Model signals that just adding connectome structure is not a reliable shortcut. It also illustrates the importance of solid controls in model evaluation, as small improvements may be due to factors unrelated to the biological graph.
AI buyers and investors should temper expectations about connectome-derived models delivering out-of-the-box boosts in language understanding or generation. The FLM approach is clever but ultimately pressure-tests the idea that wiring the full connectome into a large model architecture automatically helps. So far, it does not.