Task-Oriented Key-Layer KV Communication for Efficient Latent Multi-Agent Collaboration
KITE proposes a training-free method for latent communication between AI agents by focusing on task sufficiency rather than state fidelity. It selects a specific key layer using receiver trajectory distortion to identify the most relevant information. By communicating only the latent memory from this effective layer, KITE significantly reduces communication volume and boosts performance across various benchmarks.