# Lost in Compression: A Controlled Cross-Lingual Audit of Extractive Prompt Compressors

arXiv: https://arxiv.org/abs/2608.26175 · HTML: https://mlsanity.com/paper/2608.26175/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Prompt and context compression, one of 12 directions in Efficient LLM inference and serving.
- Builds on 22 of its 23 references that are in the map.
- First posted on arXiv 27 Jul 2026.

## Placed in

- Efficient LLM inference and serving / [Prompt and context compression](https://mlsanity.com/inference/prompt-and-context-compression/)

## Builds on (most cited first, 22 of 22)

- [Unsupervised Cross-lingual Representation Learning at Scale](https://mlsanity.com/paper/1911.02116/) (2019)
- [No Language Left Behind: Scaling Human-Centered Machine Translation](https://mlsanity.com/paper/2207.04672/) (2022)
- [LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression](https://mlsanity.com/paper/2310.06839/) (2023)
- [Learning to Compress Prompts with Gist Tokens](https://mlsanity.com/paper/2304.08467/) (2023)
- [LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression](https://mlsanity.com/paper/2403.12968/) (2024)
- [The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants](https://mlsanity.com/paper/2308.16884/) (2023)
- [LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models](https://mlsanity.com/paper/2310.05736/) (2023)
- [RECOMP: Improving Retrieval-Augmented LMs with Compression and Selective Augmentation](https://mlsanity.com/paper/2310.04408/) (2023)
- [Language Model Tokenizers Introduce Unfairness Between Languages](https://mlsanity.com/paper/2305.15425/) (2023)
- [In-context Autoencoder for Context Compression in a Large Language Model](https://mlsanity.com/paper/2307.06945/) (2023)
- [Do All Languages Cost the Same? Tokenization in the Era of Commercial Language Models](https://mlsanity.com/paper/2305.13707/) (2023)
- [Compressing Context to Enhance Inference Efficiency of Large Language Models](https://mlsanity.com/paper/2310.06201/) (2023)
- [MultiEURLEX -- A multi-lingual and multi-label legal document classification dataset for zero-shot cross-lingual transfer](https://mlsanity.com/paper/2109.00904/) (2021)
- [Prompt Compression for Large Language Models: A Survey](https://mlsanity.com/paper/2410.12388/) (2024)
- [Prompt Compression with Context-Aware Sentence Encoding for Fast and Improved LLM Inference](https://mlsanity.com/paper/2409.01227/) (2024)
- [Provence: efficient and robust context pruning for retrieval-augmented generation](https://mlsanity.com/paper/2501.16214/) (2025)
- [EXIT: Context-Aware Extractive Compression for Enhancing Retrieval-Augmented Generation](https://mlsanity.com/paper/2412.12559/) (2024)
- [Prompt Compression in the Wild: Measuring Latency, Rate Adherence, and Quality for Faster LLM Inference](https://mlsanity.com/paper/2604.02985/) (2026)
- [Sentinel: Decoding Context Utilization via Attention Probing for Efficient LLM Context Compression](https://mlsanity.com/paper/2505.23277/) (2025)
- [XProvence: Zero-Cost Multilingual Context Pruning for Retrieval-Augmented Generation](https://mlsanity.com/paper/2601.18886/) (2026)
- [Prompt Compression in Diffusion Large Language Models: Evaluating LLMLingua-2 on LLaDA](https://mlsanity.com/paper/2605.17932/) (2026)
- [Cross-Lingual Token Arbitrage: Optimizing Code Agent Context Windows via Local LLM Preprocessing](https://mlsanity.com/paper/2606.03618/) (2026)

## Built on by (most cited first, 1 of 1)

- [The Invisible Language Tax: Token Premiums of French and Regional Languages in 2026 LLM Tokenizers, and a French-Optimized Prototype](https://mlsanity.com/paper/2609.39001/) (2026)

## Abstract (arXiv)

Extractive prompt compression promises to cut LLM inference costs by removing low-information tokens, and learned compressors such as LLMLingua-2 report strong results on English benchmarks. Most other languages already pay a token premium: the same content costs 1.3-1.8x more tokens than in English. We ask whether compression closes or widens this gap. Using fully parallel data in ten languages spanning five scripts, with controls budget-matched in the target model's tokenizer, we audit four learned compressors against four deterministic baselines, on eleven target models from ten vendors (over 250,000 evaluation calls). Three of the compressors are trained with English supervision (LLMLingua-2 XLM-R/mBERT; Kompress-v2 from the production Headroom stack); the fourth, XProvence, is trained multilingually. First, the transfer gap is real, replicates across target models and compressor backbones, and is strongly rate-dependent: at a 0.33 keep-rate English retains 57-62% of normalized context utilization while Lithuanian retains 10-24% and Chinese essentially none, despite Chinese having the smallest token premium. Second, the gap tracks compression supervision data, not architecture. All three English-trained compressors show it, deterministic methods show no comparable gap, and the multilingually trained XProvence v1 shows none. Its v2 release, retrained on translated data, empties 92% of Chinese contexts at its aggressive threshold without any warning. Third, in a harder long-context setting, aggressive learned compression drives compressed contexts to or below no-context utility in three of five non-English languages. A translate-then-compress pipeline matches or beats native compression at roughly half the token cost in three of five tested languages. We release all code, compressions, and model outputs. Safe compression budgets are much smaller outside English.