Latent Communication via Denoising States in Diffusion Agents

Diffusion language model agents can communicate through compact hidden messages consisting of signed dense vectors derived from their denoising states. Experiments show that passing these continuous states provides significant performance advantages in multi-agent coordination tasks.