Individual Synaptic Morphogenesis as a Replacement for Low-Rank Adaptation

Propose of individual synaptic morphogenesis ("SyMa"), a method for specializing language models by growing new MLP tissue instead of training a low-rank adapter. On a frozen host, Hebbian co-activation of neuron pairs (co ∈ [0, 1]) sets three individual growth laws: birth 2 + 14·co, dendrite 128 + 896·co, write 0.10 + 0.25·co. New neurons are born with zero output so the host cannot become worse; finished tissue is merged into the layer matrices and inherited by later generations. Night result on Intel Core i3-1115G4, no GPU, 7-hour budget:• SyMa — 31/40 in 32.4 minutes, 89,088 parameters, 8.99 Wh, 2644 MB peak RAM; stopped itself (reason: control)• LoRA r=16 — 30/40 in 7.01 hours, 4,399,104 parameters, 116.08 Wh, 3038 MB; control loss unchanged (1.0032 → 1.0032) v1 preprint, English. Author: Rauf Pashabayli. Host: Qwen2.5-0.5B-Instruct. Run: 20261102-234611-s20261024.Sites: syma-lab · syma-fleet · syma-lab-archive