I present SNN-Synthesis v11, a comprehensive investigation of stochastic resonance and emergent intelligence in neural networks spanning 2.8K-parameter Neural Cellular Automata to 7B-parameter LLMs across 137 experimental phases. Building upon v1–v10 (Phases 1–100), v11 adds 37 new experiments (Phases 101–137) that deploy L-NCA to real ARC-AGI tasks for the first time, establishing the principle of "Continuous Thought, Discrete Action". Eight New Findings (v11) (XXXI) Latent-NCA Breakthrough — Operating NCA dynamics in a learned latent space achieves 22.7× improvement in IoU over raw-pixel NCA (0.027 → 0.613), enabling real ARC-AGI reasoning for the first time. (Phase 120) (XXXII) Context-NCA with TTCT — In-context meta-learning with gradient-based Test-Time Context Tuning achieves 78.0% pixel accuracy on real ARC, establishing the strongest zero-shot + adaptation paradigm for parameter-efficient reasoning. (Phases 122–124) (XXXIII) VQ Breaks the Exact-Match Barrier — VQ-NCA achieves the first non-zero exact match (1.9%, 3/160) on toy tasks. Gumbel NBS boosts this by +150% via discrete attractor tunneling. (Phases 128, 130) (XXXIV) The VQ Paradox — The most important null result: full-loop VQ degrades real-ARC pixel accuracy by −12.5pp (72.6% → 60.1%) and kills TTCT gradient flow (gap = +0.00%). The Straight-Through Estimator collapses over iterative NCA steps. (Phases 132–134) (XXXV) Readout-Only VQ Fails — Even single-layer VQ at the output degrades TTCT adaptation. STE gradient noise accumulates over NCA steps, making any VQ counterproductive with gradient-based adaptation. (Phase 135) (XXXVI) D8 TTA Cannot Cancel Semantic Errors — Geometric averaging via 8-fold symmetry produces −0.03% pixel gap, confirming NCA errors are systematic, not stochastic. (Phase 136) (XXXVII) Continuous Thought, Discrete Action — Removing VQ from the NCA loop achieves the highest TTCT gain (+5.05%). Intelligence requires continuous, differentiable thought with discrete output crystallization. (Phase 135) (XXXVIII) v23 Chimera Agent — Achieves 83.53% pixel accuracy and 1/50 exact match on real ARC—the first time any L-NCA system has produced a pixel-perfect solution on an unseen real ARC task. +10.93pp over v21. (Phase 137) v1–v10 Foundations (Phases 1–100) All 30 previous findings remain validated, including Noisy Beam Search, SNN-ExIt, SR-Quantization, the space-time-precision triad, Liquid MoE, attractor regularization, and the v20 agent (88% solve rate, 14K params). 38 contributions spanning 2.8K–7B parameters, NCAs to Transformers, 9 task domains, 4 model families, and 27 honest null results. Code and data: https://github.com/hafufu-stack/SNN-Synthesis Acknowledgments This research was conducted entirely independently, without institutional affiliation or corporate funding. The author currently faces financial constraints that make it increasingly difficult to maintain subscriptions to AI services essential for this line of research. To sustain and improve the quality of future work, the author is actively seeking community sponsorship. Details are available at https://github.com/sponsors/hafufu-stack.
Hiroto Funasaki (Sun,) studied this question.