Lost in Compression: A Controlled Cross-Lingual Audit of Extractive Prompt Compressors
English-trained prompt compressors can sharply degrade non-English context instead of saving it.
The audit tests four learned compressors across ten languages and eleven target models, with budgets matched to each target tokenizer. At a 0.33 keep-rate, English retained 57-62% of normalized context utilization, while Lithuanian fell to 10-24% and Chinese nearly vanished. The paper ties the gap to English-only compression supervision, since deterministic baselines and the multilingual XProvence v1 did not show the same failure. In harder long-context tests, aggressive learned compression pushed three of five non-English languages down to or below no-context utility.
HF Daily Papers' note
The audit tests four learned compressors across ten languages and eleven target models, with budgets matched to each target tokenizer. At a 0.33 keep-rate, English retained 57-62% of normalized context utilization, while Lithuanian fell to 10-24% and Chinese nearly vanished. The paper ties the gap to English-only compression supervision, since deterministic baselines and the multilingual XProvence v1 did not show the same failure. In harder long-context tests, aggressive learned compression pushed three of five non-English languages down to or below no-context utility.
HF Daily Papers' note
score 5