Covert Multi-bit LLM Watermarking: An Information Theory and Coding Approach

This feed contains multiple new arXiv submissions (May 19, 2026) across information theory, coding, wireless communications, and signal processing. The most notable paper introduces a covert multi-bit watermarking framework for large language models: a block-autoregressive embedding model with an information-theoretic capacity characterization (Gelfand–Pinsker + channel synthesis), CMDP-based block optimization, and an explicit polar-code-based algorithm achieving ~0.375 bits/token and <10% bit-error rate on short blocks with negligible perplexity/distortion. Other papers cover: alpha-fairness