Back to AI Research

AI Research

ToM2Vec learns continuous partner-state representations for multi-agent tasks

Key Takeaways

  • The method trains latent reasoning tokens using partner thoughts, then tests coordination across collaborative and competitive environments.
  • ToM2Vec studies how an AI agent can reason about another participant’s possible beliefs, intentions and next actions.
  • The paper argues that a single textual description can struggle to represent uncertainty over several possible partner states.
  • It proposes a fixed budget of continuous reasoning tokens instead.
  • ## Train a latent representation of the partner

ToM2Vec studies how an AI agent can reason about another participant’s possible beliefs, intentions and next actions. The paper argues that a single textual description can struggle to represent uncertainty over several possible partner states. It proposes a fixed budget of continuous reasoning tokens instead.

Train a latent representation of the partner

The ToM2Vec paper uses a partner’s actual thoughts as training supervision for these tokens. At test time, each agent generates its own tokens and uses them to condition its reasoning and actions.
That training setup is an important part of the method. The study is not claiming to read another participant’s private thoughts during deployment. It learns from supervised partner information during training, then evaluates how the resulting representation helps when that information is not directly supplied.

Evaluate more than one kind of interaction

The experiments cover collaborative, competitive and mixed-motive environments, including Overcooked-AI, negotiation, Decrypto, Hanabi and Collab Escape/Capture. The authors report consistent improvements over a base model and several trained or prompted baselines.
They also report weaker performance when the reasoning is expressed in text, and transfer to partner model families not seen in training. Those results support the latent representation within the tested settings; they do not establish universal understanding of human intentions or success in arbitrary real-world negotiations.

Probe what the tokens contain

The analysis examines whether the representation helps answer questions about partner beliefs, desires and next actions. That is evidence aimed at separating partner-state information from a purely task-specific shortcut.
For researchers, the useful evaluation question is whether the representation remains helpful when the partner’s behavior changes and when uncertainty is high. Performance on an interaction task and accuracy on partner-state questions provide different kinds of evidence. ToM2Vec connects those tests, while leaving the broader question of transfer beyond its experimental environments open.

Comments