AIS: Adaptive Importance Sampling for Quantized RL

Batch of arXiv ML/statistics papers (May 15, 2026) covering advances in efficient LLM inference, quantized RL, diffusion-model sampling, sampling and Bayesian quadrature, conformal/chain-of-thought calibration, fairness in conformal prediction, online multiple testing, training-free generative sampling, and new theoretical results for kernel methods and generalization bounds. Key contributions: (1) AIS — Adaptive Importance Sampling to correct non-stationary rollout/trainer bias when using low-precision (FP8) rollouts with BF16 trainers in RL, preserving exploration while preventing training-­