SYNC: Training-Free Communication for Heterogeneous LLMs
October 14, 2026
SYNC enables multi-agent systems with different architectures to communicate by sharing word indices from a shared list rather than raw vectors. This training-free method allows a receiver to write facts into its hidden states using its own learned word directions.
HOW THIS AFFECTS YOU
●
builderYou can connect different LLMs in a multi-agent workflow without needing to fine-tune them for compatibility.
●
researcherThis provides a way to achieve neural communication between heterogeneous models using only symbolic indices.