Translators Move Model Caches

Researchers demonstrate learned translators that map one model family’s internal key‑value caches into another model’s cache format, allowing a receiver model to consume context without reprocessing the original text. Experiments show substantial speedups and accuracy gains in benchmarked setups, suggesting mixed‑model agents could bypass text‑based interoperability and provenance controls. — If generalizable, cache translators change assumptions about model isolation, provenance and containment—complicating regulation and raising new security and accountability questions about mixed‑model systems.