ToM2Vec: Continuous Theory-of-Mind Tokens for Multi-Agent Coordination

ToM2Vec. Continuous Theory-of-Mind Tokens for Multi-Agent Coordination. By Nithin Sivakumaran and four others, from the University of North Carolina at Chapel Hill, Cisco, and one other institution. Published on the twenty-first of October. Its code repository is public and real: about thirty source files.

Collaborating well with other agents depends on forming an accurate theory of mind, a working model of what a partner believes, wants and is about to do. That model has to be held under uncertainty, because a partner could be in several possible states at once. Here the native medium of language model agents gets in the way. Text more easily describes one partner state than a spread of them. Writing out a theory of mind in words tends to commit to a single guess when the honest answer is several.

ToM2Vec moves the theory of mind out of text and into a rich latent space. The agent learns to form abstract theories of mind through a fixed budget of continuous reasoning tokens. Rather than spelling out a belief about its partner in words, the agent carries that belief in tokens that never have to collapse into one sentence.

Training needs a target for those tokens, and the target is the partner's actual thoughts. The theory-of-mind tokens are trained to abstract and match them. At test time, the partner's thoughts are no longer available. Each agent generates its own theory-of-mind tokens, and those tokens act as conditioners for that agent's own thoughts and actions. What was learned from watching real partners becomes a prior the agent brings to every move.

The evaluation spans collaborative, competitive and mixed-motive settings: Overcooked-AI, multiparty negotiation, Decrypto, Hanabi, and Collab Escape and Capture. Across them, ToM2Vec improves consistently over the base model and over a variety of competitive baselines, both trained and prompted. Expressing the same reasoning in text makes performance drop, which the authors read as evidence that the latent form is critical.

Training also generalizes to partner model families not seen during training, including the proprietary GPT five point four. And in analysis, the tokens improve the agent's ability to answer questions directly about a partner's beliefs, desires and next actions. The authors take that as evidence the tokens encode a theory of mind rather than task-specific information.

The paper's case rests on a mismatch between medium and content. A theory of mind is a distribution over what a partner might be thinking, and text is better at stating one thing. Giving that reasoning a continuous home, and supervising it with what partners actually thought, is what lets agents coordinate better across games as different as cooking and negotiation.