Speculating Experts Accelerates Inference for Mixture-of-Experts

This collection of recent ML/LLM research papers covers performance, robustness, privacy, and safety advances: (1) Speculating Experts Accelerates Inference for Mixture-of-Experts (MoE) proposes expert prefetching/speculative execution to overlap CPU-GPU transfers and reduce time-per-token by up to 14% for offloaded experts. (2) A Visualization for Comparative Analysis of Regression Models introduces a 2D residual/Mahalanobis-distance-based visualization to reveal error structure beyond aggregate metrics. (3) MIPO (Mutual Information Preference Optimization) is a contrastive self-improvement/p