Learning Selective Merge Policies for Deadline-Constrained Coded Caching via Deep Reinforcement Learning

This document is an arXiv math/IT feed (18 May 2026) summarizing multiple new submissions across coding, communications, information theory, and learning for communications. Highlights include: (1) "Learning Selective Merge Policies" — a DRL approach (graph-attention policy, PPO) for deadline-constrained coded caching that learns selective merging and reduces broadcast-packet expiration by ~40.9% versus the best baseline; (2) "PrismQuant" — a constructive rate--distortion theory and practical EM-driven implementation for Gaussian-mixture sources achieving near-optimal vector quantization with低