What is new in this controlled cross-lingual audit is isolating the effect: learned extractive prompt compressors trained with English supervision retain far less context outside English. At a 0.33 keep-rate, English retains 57–62% of normalized context utilization, while Lithuanian retains 10–24% and Chinese essentially none. The transfer gap replicates across eleven target models and compressor backbones.1
Context
The study compares four learned compressors—three English-supervised (LLMLingua-2 variants and Kompress-v2) and one multilingual (XProvence v1)—against four deterministic baselines. It uses fully parallel data in ten languages and five scripts, budget-matched in the target model's tokenizer, with over 250,000 evaluation calls. Deterministic methods show no comparable gap, and the multilingually trained XProvence v1 shows none, pointing to supervision data rather than architecture as the driver.1
Assessment
The new constraint is that safe compression budgets are language-dependent. Because English-trained compressors degrade sharply regardless of backbone while a multilingual model does not, language of supervision data, not computational architecture, is the binding factor. The authors' conclusion that safe budgets are much smaller outside English follows from the reported utilization numbers, though the abstract does not define how utilization is normalized.1
Caveats
What would change the headline is if re-analysis with defined confidence intervals, or independent replication, showed that the Chinese and Lithuanian losses were artifacts of a specific threshold or tokenizer. The XProvence v2 result—92% of Chinese contexts emptied after retraining on translated data—is striking but reported without the aggressive threshold's exact value; its mechanism and whether it is a release bug or a deliberate trade-off are not specified.1
