Agents in Multi agent systems

Preprint: ToM2Vec gives agents latent theory-of-mind tokens for teamwork

Researchers report that ToM2Vec trains continuous reasoning tokens to model a partner's state, beating baselines in games like Overcooked-AI and Hanabi, and that performance drops when the reasoning is in text.

Why it matters: Multi-agent setups often stumble on coordination. The authors say the training generalizes to partner model families unseen in training, including GPT-5.4.