ToM2Vec learns continuous theory-of-mind tokens for agents
Original title: ToM2Vec: Continuous Theory-of-Mind Tokens for Multi-Agent Coordination
Latent tokens conditioned thoughts and actions and transferred to unseen partner models, including GPT-5.4.
ToM2Vec trains a fixed budget of continuous reasoning tokens to represent a partner's possible mental states for multi-agent coordination. During training, tokens are supervised from the partner's actual thoughts; at test time each agent generates them to condition its own thoughts and actions. Evaluations cover Overcooked-AI, multiparty negotiation, Decrypto, Hanabi, and Collab Escape/Capture. The authors report gains over the base model and trained or prompted baselines, a drop when reasoning is written in text, and transfer to partner families not seen in training, including GPT-5.4.
Unconfirmed: Improvements and theory-of-mind encoding are author-reported; no independent confirmation is stated.
- Tokens are trained to abstract and match the partner's actual thoughts.
- Text-form reasoning underperformed the latent-token approach.
- Training generalized to partner model families unseen at train time.
- Tokens improved answers about partner beliefs, desires, and next actions.
From the original article
1 min read at huggingface.co
Forming an accurate theory-of-mind (ToM) is key to successful multi-agent collaboration and requires reasoning under uncertainty to capture multiple possible partner states. This is hard to do in the text representation space native to large language model (LLM) agents, which more easily describe a single partner state.
By Nithin Sivakumaran, Ali Payani, Hyunji Lee · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.