A bio-inspired neural architecture that replaces static transformer layers with a dynamic population of reasoning cells that split, merge, and die based on computational demand.
Head-to-head comparison — parameter-matched at ~34K params
| Benchmark Task | CRF Perplexity ↓ | Transformer Perplexity ↓ | CRF Advantage |
|---|---|---|---|
| Synthetic Sequences | 6.69 | 21.14 | 3.2× |
| ARC Reasoning | 17.36 | 53.04 | 3.1× |
| Arithmetic | 22.37 | 62.75 | 2.8× |
| Chain-of-Thought | 21.96 | 60.10 | 2.7× |
| Code Generation | 23.95 | 60.29 | 2.5× |
Cells split when overloaded, merge when redundant, and die when inactive — dynamically allocating compute where it's needed.
Each cell communicates with k nearest neighbors instead of all-to-all attention, reducing complexity from O(N²) to O(N·k).
Cells maintain energy levels that govern their lifecycle — high-energy cells split, low-energy cells die, similar cells merge.
Dynamic halting lets the model use more reasoning steps for harder inputs and fewer for easy ones, saving FLOPs automatically.
Observed during CUDA training runs