LatentMAS – agent collaboration from token space into the model's latent space

> without any additional training.

Could this be applied to existing models? (Sorry if in paper, just read abstract for now)