ToM2Vec from Cisco and UNC gives LLM agents latent theory-of-mind

On a multi-party negotiation task, teaching an LLM agent to think about its partner in a latent vector space, not in words, raised the deal-success rate across 100 games from 62.0% to 71.0%. On Overcooked, the same technique added 12.0 points to the Qwen3-8B base model's mean episode score across three kitchen layouts and 30 seeds. The method is called ToM2Vec, and comes from Cisco, UNC Chapel Hill and UT Austin, in a preprint by Ali Payani, Hyunji Lee, Mohit Bansal and Elias Stengel-Eskin.

The pitch is that natural language is a bad representation for modeling another agent. "A general sketch of the partner's beliefs and intentions is often sufficient, and excessive fidelity can be inefficient and in fact harmful to the task," the paper argues. Committing a partner's state to words tends to collapse several possible hypotheses into one. ToM2Vec instead trains the main agent to predict a fixed budget of continuous "thought tokens" that abstract over the partner's reasoning, supervised in self-play by the partner's own hidden states under a cosine similarity objective.

The ablation makes the latent part load-bearing. Replacing the ToM tokens with a trained natural-language summary of the partner, a baseline the authors call Language ToM, recovers only part of the gains. In Collab Escape, "Language ToM even leads to a 10 percentage point (pp) drop in performance, whereas ToM2Vec improves by 16pp."

Generalization to partners the agent was not trained against holds. With a GPT-5.4 partner that was not seen during training, ToM2Vec improves over the base agent by 17.3 points on average on Overcooked across 20 seeds. Across Decrypto, Hanabi and Collab Capture, the authors report consistent gains over both prompting-based ToM baselines and trained verbalized ToM. On a Bayesian probe asking what the partner believes, wants and will do next, ToM2Vec beats the next-best baseline by 5.6, 2.0 and 2.7 accuracy points.

The method still has in-paper caveats: training needs the partner's chain-of-thought during self-play, both agents must share a representation space (so the training partner is the same base model), and the cross-family results come from a 20-seed evaluation on three Overcooked layouts. Code is public at github.com/nsivaku/tom2vec.

The arrival sits in a busy week for multi-agent coordination work in our tracker, alongside FrontierCS 2 on end-to-end CS research tasks.