ToM2Vec: Continuous Theory-of-Mind Tokens for Multi-Agent Coordination
Agents performed better when their models of other players stayed in latent tokens; expressing that reasoning in text reduced performance.
ToM2Vec learns a fixed budget of continuous tokens, supervised during training by a partner’s actual thoughts. At test time, each agent generates those tokens to guide its own reasoning and actions. The authors report consistent gains over the base model and trained and prompted baselines across five collaborative, competitive and mixed-motive environments. They also report generalization to unseen partner model families, including GPT-5.4, and evidence that the tokens capture partners’ beliefs, desires and next actions.
HF Daily Papers' note
ToM2Vec learns a fixed budget of continuous tokens, supervised during training by a partner’s actual thoughts. At test time, each agent generates those tokens to guide its own reasoning and actions. The authors report consistent gains over the base model and trained and prompted baselines across five collaborative, competitive and mixed-motive environments. They also report generalization to unseen partner model families, including GPT-5.4, and evidence that the tokens capture partners’ beliefs, desires and next actions.
HF Daily Papers' note
score 4