SNN-Synthesis v8: Quantization Noise as Stochastic Resonance, Multi-Model Beam Ensembles, and ARC-AGI-3 Field Validation from 63K to 7B Parameters

Abstract I present SNN-Synthesis v8, a comprehensive investigation of stochastic resonance in neural networks spanning 63K-parameter CNNs to 7B-parameter LLMs across 63 experimental phases. Building upon v1–v7 (Phases 1–60), this version adds 3 new experiments (Phases 61–63) and the first ARC-AGI-3 Kaggle field results. Three New Findings (v8) (XIII) Quantization Noise as Stochastic Resonance — 4-bit Qwen-1.5B achieves 58% at K=1 without beam search, surpassing both FP16 (32%) and Mistral-7B baseline (42%). Quantization noise acts as free stochastic resonance. At K=51, all precisions converge to 84%. A "double noise" non-monotonicity (8-bit K=21: 82% > K=51: 78%) reveals destructive interference. The space-time duality extends to a space-time-precision triad. (Phase 61) (XIV) Multi-Model Beam Ensemble — Mixing beams from Mistral-7B (×6) + Qwen-1.5B (×5) achieves 86.7%, surpassing all single-model ensembles (Mistral ×11: 70%, Qwen ×11: 80%). Diversity analysis confirms architectural diversity is orthogonal to noise diversity. (Phase 63) (XV) ARC-AGI-3 Kaggle Field Validation — Five agents submitted to the live competition. The simplest (v5: macro-stats UCB, score 0.13) beats all "intelligent" agents (v14 CfC: 0.10, v12 LLM: 0.07, v13 SimHash: 0.02). Root cause: three "death traps" — timeout starvation, pixel-noise sensitivity, action-space explosion. The winning strategy uses thermodynamic coarse-graining: macroscopic statistics invariant to microscopic noise. v1–v7 Foundations (Phases 1–60) (I) Noisy Beam Search: K=11 parallel noisy trajectories achieve 78% on CNN (from 12%) and 100% on Mistral-7B Modified Hanoi (from 16%). (II) SNN-ExIt: Oracle-free self-evolution reaches 99% on ARC-AGI-3 LS20, surpassing Oracle CNN (78%). (III) Knowledge Multiplexing via ID Gating (Phase 35c): A single 115K parameter CNN stores distinct knowledge for multiple games without interference, using discrete condition-ID gating (h ← h ⊙ σ(Embed(id))). Knowledge separation score reaches +0.572. Four alternative approaches (noise modulation, SNN chaotic noise, continuous-wave gating, pink noise) all fail—only discrete gating succeeds, mirroring biological neurotransmitter-based mode switching. (IV) σ-Diverse NBS (Phase 37a): Assigning different σ values to each of K=11 beams eliminates the need for task-specific σ* tuning. Performance matches the best individually-tuned fixed σ across all tested difficulty levels, providing a hyperparameter-free exploration strategy. (V) Capacity Scaling (Phase 38a): ID gating requires ≥2.7K parameters for effective knowledge separation. At 115K parameters, gated models surpass ungated models (0.706 > 0.625), demonstrating that gating acts as positive regularization. (VI) Multi-Model NBS (Phase 38): Qwen2.5-7B-Instruct achieves 100% solve rate at K=11 on Modified Hanoi, matching Mistral-7B and confirming cross-model universality. (VII) GSM8K LLM-ExIt (Phase 33): Extending LLM ExIt to math reasoning; Mistral-7B K1 accuracy improves from 56.5% to 58.0% over 3 iterations. The modest gain confirms ExIt functions on open-ended tasks but reveals that high-baseline tasks limit self-improvement headroom. (VIII) σ* Prediction (Phase 34): TruthfulQA MC1 achieves 100% accuracy at σ*=0.2 with K=11, extending the σ* map to four tasks: GSM8K (0.01), TruthfulQA (0.2), Hanoi (0.15), ARC-AGI (0.2). v7 Key Findings (Phases 39–60) RND Curiosity (Phase 39): 63.5% solve rate vs. Random 2.5% at difficulty 6 (+61pp), but 0.5ms/action overhead is fatal under time budgets σ-Diverse NBS on LLMs (Phase 40): Mistral-7B GSM8K achieves 70% with σ-diverse K=11, confirming LLM-scale superiority The Crossover Law (Phases 44–46): Overhead >0.5ms/action → intelligence loses to random exploration SimHash O(1) Curiosity (Phase 51): Locality-sensitive hashing matches RND at ~100× less overhead (0.005ms) Grand Simulation (Phase 52): Final benchmark validates SimHash + σ-diverse NBS as optimal ARC-AGI-3 agent Asymptotic Scaling (Phase 56): At 10M actions, all agents converge to 100%—no fundamental wall exists 6 Null Results (Phases 53–55, 57–58): Associative SimHash, temporal SimHash, macro chunks, bag-of-patches, and curiosity-diverse swarms all fail, confirming design convergence SR-Quantization (Phase 59): Qwen-1.5B + NBS (80%) > Mistral-7B baseline (42%) — small + noise > large + greedy TTC Scaling Law (Phase 60): Logarithmic accuracy scaling with K, optimal cost-performance at K=5–11 38 contributions spanning 63K–7B parameters, CNNs to Transformers, 9 task domains, 2 model families (Mistral, Qwen), 4 model scales (1B–7B), and 20 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.