Fruit fly connectome wiring fails to improve a language model

single source · 1 articles · research · confidence: medium · first seen 2026-09-12 18:51 UTC

A preprint describes a Fly Language Model (FLM) that 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 preprint reports a 0.0222 nat per token improvement over the backbone, but a parameter-matched control without the graph does slightly better in every seed. Four controls are used. The work is MIT-licensed and runs locally with a reported 0.6 per token memory bound. This suggests the connectome does not help.

What this means for you

Nothing to act on today. This is a preprint, not a product—no API and no benchmark beyond its own controls. If you work on biologically inspired architectures, treat the reported 0.0222 nat improvement as not meaningful: the parameter-matched control without the graph does slightly better in every seed.

Key facts

  • ·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. source
  • ·It adds a small learned correction to a frozen LFM2.5-1.2B-Instruct backbone. source
  • ·Only 278,528 parameters are trained. source
  • ·The preprint reports a 0.0222 nat per token improvement over the backbone. source
  • ·A parameter-matched control without the graph does slightly better in every seed. source
  • ·The code is MIT-licensed. source

What the sources say

  • MarkTechPostBreaks down the FLM architecture, its parameter-matched control outcomes, and the MIT-licensed local code.

Sources

The original reporting. Follow these — they did the work.

← the wire