SNN-Synthesis v2: Scale-Invariant Stochastic Resonance—From Billion-Parameter LLMs to Micro-Brain CNNs

This preprint presents SNN-Synthesis v2, extending the investigation of stochastic resonance in neural networks from billion-parameter LLMs to extreme-edge CNNs. Building on v1 (Phases 1–7: trajectory distillation, multi-layer orchestration, task specificity on Qwen-0.5B and Mistral-7B), v2 adds Phase 8: a demonstration of scale-invariant stochastic resonance in ARC-AGI-3 sequential decision-making using Micro-Brain CNNs (63K–244K parameters). The key finding is that CNN agents trained via behavioral cloning achieve 0% clear rate without noise, rising to 20% at σ=0.2 (N=30), with both architectures showing identical peaks. The optimal noise σ*=0.2 precisely matches the σ*=0.15–0.30 range in 7B LLMs, establishing that optimal noise is determined by task structure, not model capacity—a universal property spanning five orders of magnitude in parameter count. Code and data: https://github.com/hafufu-stack/SNN-Synthesis AcknowledgmentsThis 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