Diffusion Language Models for Lossless Text Compression

A new arXiv paper (2608.11249) proposes using Diffusion Language Models (DLMs) for lossless text compression, a departure from traditional autoregressive LLM-based methods. The authors argue that DLMs can overcome the severe throughput limitations of current neural compression approaches, which, despite achieving compression ratios superior to general-purpose compressors like zstd, gzip, or bzip, are not yet practically usable due to speed constraints. The paper introduces DLMs as an alternative inference paradigm within the same compression framework, potentially enabling practical lossless compression for digital textual data, including plain text, source code, and structured formats like XML. The study is motivated by the rapid growth in digital data collection and storage, and recent advances in neural language model-based compression. The paper is available on arXiv and was announced as a cross-type submission.