From Connectome Wiring to Counted Memory and Recurrent Depth: Pre-Registered Tests of Drosophila Mushroom-Body Principles in Small Language Models
Year: 2026
Type: preprint
Abstract: Across 16 pre-registered stages (101 production training and evaluation runs, 25 registered predictions P1-P25, and a 218-mutant automated test audit), we systematically test where Drosophila mushroom-body principles help and where they fail in small language models (5M-29M parameters trained on 200M tokens of FineWeb-Edu). We test literal FlyWire v783 mushroom-body connectome wiring (Stages 1 & 2T), counted host-RAM n-gram memory (COUNT) with online leave-one-out subtraction and Kenyon similarity hash backoffs (Stages 2.0-2B, 2E-2K2), gradient-free one-shot associative count writes with context-gated overlays (GatedOverlay) and memory-on KL consistency consolidation (Stages 2C, 2C-2, 2C-3), and recurrent looped depth (LoopGPT) with learned per-pass residual gains and a 2x2 factorial isolating why unembedding head.weight at AdamW 64e-3 explodes 8x in L2 norm and kills recurrent depth (Stages 2D, 2H-2I, 2L-2N).
Author: Farrow, Nate
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234884313
