MIT researchers develop SwaT to let AI models share internal memory directly

MIT researchers have developed a translator called SwaT that allows text-only AI models to directly access internal memory from models that have processed images or other data, bypassing text entirely. Tests across multiple model families showed that receiving models retained over 90 percent of their normal scores while operating significantly faster and more accurately than passing text summaries. However, each model pair requires a dedicated trained translator, and the passed memory can carry unreadable information that prevents text-based verification. Open article →