LoMC: Localized Multidirectional Correction for Refusal Suppression in Routed Foundation Models

Collection of recent arXiv ML/statistics papers (June 15 2026) covering: LoMC — a support-gated, layer-wise rank‑one intervention for refusal suppression in routed Mixture-of-Experts and hybrid-MoE foundation models that improves non-refusal behavior while preserving capability; theoretical and spectral analysis of recursive training for diffusion models showing convergence to a unique Gaussian-smoothed limit and a low-pass effect that explains collapse; ANTS — Adaptive Nucleus Truncation Sampling for long-form reasoning that adapts truncation width by entropy and improves multi‑budget decode/