Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT-IBM Watson AI Lab publicly released their ‘SwaT’ (Swarm Telepathy) study on the preprint platform ‘arXiv’ on October 10, describing technology that transmits AI’s internal computations without relying on textual explanations. The technology leverages ‘KV caches’—computational information stored by AI during data processing—and connects a program that converts information between two AI models so the receiving AI can utilize the prior model’s computations without re-reading data from scratch. The team tested 12 models from six families, including Meta’s ‘Llama’, Alibaba’s ‘Qwen’, and Google’s ‘Gemini’; in experiments combining information compression with text summarization, average accuracy was approximately 40% higher, and computation time for the sending model and the receiving model’s response initiation time each decreased to about a quarter of the original. Anthropic revealed in June 2025 that its research system, where multiple AI agents collaborate, uses approximately 15 times more tokens than general chat. In experiments, Google’s ‘Gemini 3 4B’ transmitted internal computations from processing a photo to IBM’s text-only model ‘Granite 3.3 8B’, and when internal information was passed instead of a photo description, Granite answered photo-related questions more accurately. The technology remains in the research phase, requiring separate converters for each model combination.