Extractive prompt compression is the cheapest win in the book: strip the low-information tokens out of your context before you send it, keep the meaning, pay for fewer tokens. The learned compressors people actually deploy were supervised on English. This audit runs them across ten languages and five scripts, scoring not raw accuracy but how much of the context's usable value survives — anchored between answering with the full context and answering with none. Keep a third of the tokens and English retains 57–62% of that value, Lithuanian 10–24%, and Chinese lands at or below the no-context floor: the compressed context is worth less than sending nothing at all. The cause is the supervision language, not the architecture — a multilingually trained compressor shows no gap, and neither do the dumb deterministic baselines.
Raw accuracy would be the wrong instrument here, because some questions can be answered without the context at all and some cannot be answered even with it. So the audit anchors: score the model with the full context, score it with no context, and express a compressed result as the fraction of the distance between those two that survives.
That gives a number with real meaning at both ends. One means the compressed context was as good as the whole thing. Zero means it was worth no more than sending nothing. And it can go negative — which is not a rounding artefact but a real outcome, and one that shows up here.
The keep-rate is what you ask the compressor for: three quarters of the tokens, half, a third. Everything below is measured at those three settings on parallel reading-comprehension items, 300 per language, so the questions are identical across every language tested.
At a gentle keep-rate of three quarters, everything looks fine. Most languages sit between 0.77 and 0.91, Chinese lowest at 0.62 to 0.73. This is the setting at which a team would test the feature, conclude it works, and ship it.
Halve the context and the languages separate. Push to a third and they come apart completely: English holds 0.57 to 0.62, Hindi 0.33 to 0.53, Lithuanian 0.10 to 0.24, Ukrainian 0.10 to 0.26, and Chinese reaches −0.03 on one model — below the no-context floor.
Read the columns as a pair, because the two evaluating models disagree in an instructive way. Chinese at a third of the tokens is −0.03 under one and 0.26 under the other. The compression damage is real in both; how much a given model can still salvage from the wreckage is not the same.
Three learned compressors were audited and all three show the gap, averaging between 0.23 and 0.31 at a half keep-rate. That is suggestive but not conclusive on its own — until you look at what does not show it.
The deterministic baselines have no comparable gap: lemmatisation with stopword removal holds 0.45 to 0.83 at its natural budget, and simple term-frequency extraction degrades smoothly with no cross-lingual cliff at all. Nor does the one compressor trained multilingually. And when a later version of that same compressor was trained on translated data, the gap came back on Chinese. Supervision language, then, not model family.
The linguistic account the authors offer is preliminary and worth understanding anyway. In inflected languages the thematic roles — who did what to whom — live in case endings and agreement morphology, so deleting "low-information" function morphology is not free. In Chinese the particles marking object, passive and aspect are precisely the high-frequency function tokens an English-trained classifier learns to throw away. A probe on Lithuanian minimal pairs found compression makes the model retreat to "not stated" rather than reverse the roles: the grammatical cue was destroyed, not the fact.
The operational problem is that none of this raises anything. A compressor handed Chinese may simply no-op — a whitespace-based method finds no whitespace to split on — and return your text unchanged, or return something worth less than nothing, and in both cases the call succeeds and the response looks like a response.
The one signal the authors point at is cheap and worth wiring up: the gap between the compression rate you requested and the rate you actually achieved. Every failure mode they describe is visible there, and nowhere else.
Compression is bought for cost, so the honest comparison is cost saved against context value lost. Set a keep-rate and a language and look at both sides at once.
The uncomfortable case is the middle: a keep-rate that looks like a reasonable trade in English is often past the cliff in a language whose tokens already cost you more, so you are paying a premium to destroy more.