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

The Fly Language Model (FLM) project integrates the complete male fruit fly connectome—166,700 neurons and 25.6 million synaptic edges—into a frozen LFM2.5-1.2B-Instruct language model using token embeddings.

Rather than training the entire model, researchers kept the LLM backbone frozen and trained only a small learned correction layer comprising roughly 278,528 parameters. This approach let them test whether injecting biological wiring patterns into an AI model would enhance its capabilities.

The study is notable for its rigorous ablation testing. The researchers' own control experiments indicated that the fly connectome wiring did not actually improve the model's performance, offering a sobering data point on the direct transferability of biological neural architectures to artificial systems.

This work contributes to the broader neuromorphic AI field, where scientists explore whether insights from natural nervous systems can inform better machine learning designs. The negative result suggests that simply porting biological connectivity into existing transformer architectures may not yield straightforward gains.

The outcome underscores the complexity of translating neurobiology into AI advances, pointing researchers toward more nuanced methods rather than direct structural mimicry when bridging these two domains.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.