SNN-Genesis v16: Stochastic Resonance in LLM Reasoning — Architecture-Specific Reasoning Manifolds and Causal Mechanisms

SNN-Genesis v16 establishes that stochastic resonance in LLM reasoning is architecture-specific and causally verified. Building on v15's cross-architecture divergence, v16 dissects Qwen2.5-7B's reasoning manifold through 7-band PCA decomposition and discovers that reasoning information is NOT in the top principal components — the exact opposite of Mistral-7B. Honest replication at N=100 corrects sampling optimism from N=30, and causal experiments separate the contributions of noise direction (~2/3) and stochastic randomness (~1/3). NEW in v16 (Season 16, Phases 84, 85, 86): Architecture-Specific Reasoning Manifolds (Phase 84): 7-band PCA decomposition of Qwen2.5-7B reveals that PCA-top-4 noise = baseline (46.7%), while Mistral's PCA-top-4 is catastrophic (3.3%). Reasoning representations are architecture-specific, not universal. Honest Correction at N=100 (Phase 85): Phase 82's PCA-bottom advantage (53.3% at N=30) does NOT replicate at N=100 (34.0% ≈ random 35.0%). Noise universally harms Qwen, confirming the architecture-specific nature of stochastic resonance. Causal Verification (Phase 86): Deterministic offset in PC 257+ achieves 30% vs. stochastic noise 40% (p=0.015 vs. baseline 10%). Direction provides ~2/3 of the benefit; randomness provides ~1/3. Deterministic mid-band (PC 65-256) collapses to 0%. 90 page paper. Full experimental code and data included. Code: https://github.com/hafufu-stack/snn-genesis