MIT Researchers Unveil SwaT Method for Cross-Model KV-Cache Transfer

Researchers at MIT CSAIL and the MIT-IBM Computing Research Lab have published a paper on SwaT (Swarm Telepathy), a method that transfers KV caches directly between models from different families, and released the code on GitHub, according to aitimes.com.

SwaT converts one model's KV cache into a form another model can use, letting models from different developers exchange internal computation state without text. On average, it improved accuracy by 40% and cut total processing time by 4x compared with text-summary-based communication across dialogue and reading-comprehension benchmarks.

The method needed only about 44 million unique tokens and less than one GPU-day of training, and the models themselves were not fine-tuned. Experiments covered combinations including Qwen to Llama, and Gemma to Qwen and Granite.

In one test, a Gemma model analyzed images and passed its internal state via SwaT to a text-only model, which then answered image-related questions more accurately than when given only an image description. Passing Qwen 3 4B's internal state to Llama 3 8B improved Llama's scores on all 67 subjects of the Chinese benchmark CMMLU, closing about three quarters of the gap between Qwen and Llama.

Previously, multi-agent AI systems had one AI summarize its results as text for another AI to read, which could lose information and required extra computation.

The researchers noted limitations including the need for a separate converter trained per sender-receiver pair, instability when repeatedly combining or merging KV caches, and open questions on auditability and safety of transferred caches.