Event

ToM2Vec: Modeling Multi-Agent Theory of Mind with Continuous Reasoning Tokens

1 reports · 1 sources

The ToM2Vec paper's author team proposes a method that lets multiple agents form an abstract theory of mind (ToM) in latent space through a fixed budget of continuous reasoning tokens, and evaluates collaboration effectiveness in environments such as Overcooked-AI, multi-party negotiation, Decrypto, Hanabi, and Collab Escape/Capture. During training, these ToM tokens are abstracted and matched using partners' true thoughts as supervision signals; at test time, the ToM tokens generated by each agent serve as conditions for its thinking and actions. The authors state that changing reasoning to textual expression leads to performance degradation, indicating that latent-space reasoning is key; they also state that ToM2Vec training can generalize to partner model families unseen during training, including the proprietary model GPT-5.4.