ToM2Vec learns abstract Theory‑of‑Mind (ToM) representations in a continuous latent space. It uses a fixed budget of continuous reasoning tokens to capture the uncertainty over multiple possible partner states. During training, supervision for these tokens comes from the partner's actual thoughts, forcing the model to abstract and align with them. At test time each agent generates its own ToM tokens, which condition the agent's internal reasoning and actions.
We evaluate ToM2Vec across several environments—collaborative Overcooked‑AI, multi‑party negotiation, Decrypto, Hanabi, and collaborative Escape/Capture scenarios that mix cooperative and competitive motives. Results show consistent improvements over the base model and a range of competitive baselines. An ablation study reveals that expressing reasoning in text dramatically hurts performance, highlighting the importance of latent continuous reasoning.
Moreover, the training transfers to partner model families unseen during training, including a proprietary GPT‑5.4, while still outperforming baselines. Analysis indicates that the learned tokens encode a ToM—capturing partner beliefs, desires, and next actions—rather than merely task‑specific cues; agents equipped with ToM2Vec answer direct partner‑state queries more accurately.
Review