What question did this study set out to answer?

The research aims to enhance LLM reasoning capabilities through advanced stochastic resonance techniques.

March 24, 2026Open Access

SNN-Genesis v17: Stochastic Resonance in LLM Reasoning — Differential PCA, Subspace Correlation, and Temporal Noise Dynamics

Key Points

The research aims to enhance LLM reasoning capabilities through advanced stochastic resonance techniques.
Utilized differential PCA to identify outcome-discriminant axes in LLM reasoning.
Implemented inter-layer correlation to combine noise for performance improvement.
Refined the direction/randomness ratio through a 5-point mixing ratio sweep.
Analyzed the impact of temporal noise correlation on performance.
Achieved a 50% solve-rate on Qwen with targeted noise applied to discriminant PCs.
Broke the 40% ceiling with a new record of 46.7% using subspace-targeted inter-layer correlated noise.
Refined random interaction contributions to stochastic resonance at a 31:69 ratio.
Demonstrated that correlated temporal noise significantly degrades LLM performance.

Abstract

SNN-Genesis v17 discovers architecture-specific reasoning directions via Differential PCA, breaks the 40% solve-rate ceiling with subspace-targeted inter-layer correlation, refines the direction/randomness decomposition to 31:69, and proves that temporal noise correlation is harmful to stochastic resonance. NEW in v17 (Season 18, Phases 87, 88, 89, 90): Differential PCA Discovers Qwen's Reasoning Direction (Phase 87): Standard PCA captures variance, not reasoning. Using labeled hidden states (solved vs. failed), Differential PCA identifies the outcome-discriminant axis. Noise in the top-10 discriminant PCs achieves 50.0% on Qwen — the first-ever demonstration that Qwen benefits from targeted noise. 46.7% New All-Time Record (Phase 88): Combining inter-layer correlated noise (L17+L18, ρ=+1) with the safe subspace (PC 257+) at σ=0.075 breaks the 40% ceiling. σ=0.106 collapses to 0%, confirming the cliff-edge is subspace-independent. Direction/Randomness Ratio Refined to 31:69 (Phase 89): 5-point mixing ratio sweep reveals randomness contributes 69% of the stochastic resonance effect, reversing v16's estimate of ~2/3:1/3. Pure stochastic noise (46.7%) dominates pure deterministic offset (16.7%). Temporal Noise Correlation is Harmful (Phase 90): AR(1) correlated noise monotonically degrades performance: IID (ρ=0.0) achieves 20.0%, ρ=0.95 drops to 3.3% (= baseline). Stochastic resonance requires fresh, independent noise at every step. 96 page paper. Full experimental code and data included. Code: https://github.com/hafufu-stack/snn-genesis

SNN-Genesis v17: Stochastic Resonance in LLM Reasoning — Differential PCA, Subspace Correlation, and Temporal Noise Dynamics

Key Points

Abstract

Cite This Study