Description
Agents collaborating through text burn tokens and lose information in translation. LatentMAS, an ICML 2026 Spotlight, moves multi-agent collaboration from token space into the model's latent space so agents exchange internal representations directly.
This continuous reasoning saves tokens and improves results.
Latent collaboration:Representations shared.
Fewer tokens:Less text back-and-forth.
Continuous reasoning:No information loss.
This continuous reasoning saves tokens and improves results.
Features
Latent collaboration:Representations shared.
Fewer tokens:Less text back-and-forth.
Continuous reasoning:No information loss.
