Cross-Lingual Compression Gap in Extractive Prompt Compressors

An audit of four learned and four deterministic compressors across ten languages reveals that English-supervised compressors like LLMLingua-2 increase the token cost gap for non-English languages. The transfer gap persists across 250,000 evaluation calls, showing that compression benefits are highly rate-dependent and language-sensitive.