SwaT
A multi-agent communication method that transfers KV caches across model families, allowing frozen heterogeneous LLMs to share intermediate KV states directly without prefill on the receiving side. Lightweight alignment layers handle shape adaptation and semantic alignment.
Timeline (last 90 days)
SwaT (Swarm Telepathy) is a method that supports KV cache transfer across model families, enabling prefill-free communication in multi-family multi-agent systems.
SwaT training requires less than 50 million tokens and less than 1 GPU-day.
SwaT requires that both the sender and receiver models remain frozen, with no fine-tuning.
In agent-to-agent communication scenarios, SwaT combined with recent KV cache compression techniques achieves on average 40% higher accuracy than text-to-text communication and is on average 4 times faster in wall-clock time.
The SwaT experiments cover 12 LLMs, 6 model families, and 17 benchmarks.
If the sender is a vision-language model (VLM), SwaT can transfer visual capabilities to a pure-text receiver LLM, with effects surpassing text-to-text communication, even though SwaT itself was only trained on pure-text data.
The mechanism of SwaT is inferred to be doing projection or alignment at the cache level, rather than modifying the model itself.
All Facts (7)
SwaT (Swarm Telepathy) is a method that supports KV cache transfer across model families, enabling prefill-free communication in multi-family multi-agent systems.
50%UnverifiedSwaT training requires less than 50 million tokens and less than 1 GPU-day.
50%UnverifiedSwaT requires that both the sender and receiver models remain frozen, with no fine-tuning.
50%UnverifiedIn agent-to-agent communication scenarios, SwaT combined with recent KV cache compression techniques achieves on average 40% higher accuracy than text-to-text communication and is on average 4 times faster in wall-clock time.
50%UnverifiedThe SwaT experiments cover 12 LLMs, 6 model families, and 17 benchmarks.
50%UnverifiedIf the sender is a vision-language model (VLM), SwaT can transfer visual capabilities to a pure-text receiver LLM, with effects surpassing text-to-text communication, even though SwaT itself was only trained on pure-text data.
50%UnverifiedThe mechanism of SwaT is inferred to be doing projection or alignment at the cache level, rather than modifying the model itself.
50%