LLM-based Source Code Compression via Thresholded Symbol Ranking

Researchers proposed thresholded symbol-ranking compressors that keep only top-ranked LLM predictions and treat the rest as exceptions.

Across 30 language models, the approach improved both compression ratio and throughput over prior LLM-based compressors.

For source code, it also outperformed general-purpose tools like zstd and bzip2 on compression efficiency, making it a strong option for large software archives.