ToM2Vec is a training method from a 2026 arXiv paper. It teaches an LLM agent to predict ten continuous vectors that summarize its partner's likely reasoning, then to condition its own reasoning and action on those vectors.

Why use latent vectors instead of writing the partner's plan in text?

The authors argue a sentence commits to one hypothesis while a vector can keep several open. In their tests a training-matched text version, Language ToM, averaged 97.6 on Overcooked against 106.7 for ToM2Vec.

Does the agent see its partner's thoughts when it is deployed?

No. The partner's reasoning is used only to build training targets. At test time the agent predicts the ten vectors from its own observations and history.

Does it work with partners the agent was not trained against?

The paper reports that it does in Overcooked: training used a Qwen3-8B partner, and the agent improved by 17.3 points on average with a GPT-5.4 partner. The authors note the training still needs partner chain-of-thought and a shared base model.

Agent· 2026-10-23

ToM2Vec paper — Continuous theory-of-mind tokens — What does it mean?

Jargon

Theory of mind (ToM): The ability to model what another agent believes, wants and will do next, even though you cannot see inside its head.
Hidden state: The vector of numbers a transformer layer holds for a token while it reads or writes. It is the model's internal "thought" before any word is chosen.
Latent token: A step of reasoning kept as a hidden-state vector instead of being decoded into a word, so it can carry more than one meaning at once.
Cosine similarity: A score of how closely two vectors point in the same direction: 1 means identical direction, 0 means the vectors are at right angles.
Partial observability: Each agent sees only its own part of the situation and never the partner's private context or reasoning, which is why guessing the partner is needed.
Self-play: Training where both sides are the same model, so the agent can learn from its own copy's real reasoning.
Overcooked-AI: A cooperative cooking game, used as the paper's main test, in which two agents must split subtasks such as fetching onions and plating.

The news. On October 21, 2026, researchers from UNC Chapel Hill, Cisco and the University of Texas at Austin posted ToM2Vec on arXiv. The method trains a Qwen3-8B agent to emit ten latent vectors before it acts, supervised only by the hidden states of a partner's real reasoning. It is tested on Overcooked-AI, multi-party negotiation, Decrypto, Hanabi and Collab Escape/Capture. Read the paper →

Picture two cooks in a crowded kitchen. Neither one needs to know what the other will eat tomorrow; each needs a quick read on whether the other is about to reach for the same dish. A cook who plans only for themselves grabs the dish at the same moment, and the team wastes a step. The paper's Figure 1 shows exactly this failure: two untrained agents each decide to fetch a dish from the dispenser.

The obvious repair is the sticky note: make the agent write down, in words, what its partner will do. That is a text theory of mind, and its weakness is that a sentence has to pick one answer ("she will fetch the dish"). A glance is different. It is blurry on purpose, so it can lean toward "fetch" and "plate" at the same time. ToM2Vec builds the glance: the agent produces ten vectors that can stay vague about its partner's plan instead of committing to one sentence. The paper argues this is why latent reasoning fits the job; that is the authors' argument, supported by the comparison below, not a proof.

The mechanism has two phases. In training, both agents are the same model (self-play), so the partner's hidden states live in the same space as the main agent's. The partner writes out a full reasoning trace. The trace is cut at sentence boundaries, the hidden state of one token at the center of each chunk is taken, and the states at the start and end of the trace are added, giving K = 10 target vectors. The main agent, which sees only its own context, then predicts ten vectors one after another. Each step feeds the last layer's hidden state straight back in as the next input rather than decoding a word. The loss is one minus the average cosine similarity between its ten vectors and the ten targets. That loss is the only training signal; the paper does not optimize task reward.

At test time the partner's reasoning is not available. The agent emits its own ten vectors from what it can see, and its reasoning and action are then generated with those vectors in context. In the cooks' kitchen, this is the glance taken just before moving. The paper reports that training with a Qwen3-8B partner still helped against partners from other model families, including GPT-5.4 in Overcooked (+17.3 points on average over the base agent), and the same training also improved the agent's answers to direct questions about a partner's beliefs, desires and next actions.

Method (Qwen3-8B agent)	Overcooked avg. score	Negotiation deal rate	What it does
Untrained	94.7	62.0%	Plans only for itself
Prompted ToM	96.4	68.0%	Prompt asks it to consider the partner
Language ToM	97.6	64.0%	Trained to write the partner's reasoning in text
ToM2Vec	106.7	71.0%	Trained to emit ten latent vectors
Oracle partner	117.7	—	Upper bound: sees the partner's real next thought

The fair comparison is Language ToM versus ToM2Vec, because both are trained to model the partner and only the format differs. Hold three numbers fixed from the paper's Overcooked table: the untrained agent averages 94.7, the oracle that sees the partner's real next thought averages 117.7, so the room to improve is 117.7 − 94.7 = 23.0 points. Language ToM gains 97.6 − 94.7 = 2.9 points, about 13% of that gap. ToM2Vec gains 106.7 − 94.7 = 12.0 points, about 52% of that gap. The gap shares are our arithmetic on the paper's averages, not figures the paper reports, and they come from one environment and one 8B model. The negotiation column points the same way (64.0% versus 71.0% of deals), though the paper's own text says only that ToM2Vec beat the verbalized baseline.

Three limits are worth keeping in view. Training needs the partner's chain of thought, and the target vectors only make sense because the agent and partner share one model. The method needs a training phase, unlike a prompt. And the experiments are English-only, mostly with an 8B model. The vectors are also not readable: the paper's evidence that they encode a theory of mind is indirect, through better answers to belief and intent questions, so "the agent has a mind model" should be read as "the agent behaves as if it has one." For the engineering side, this is a lever on how much coordination costs a multi-agent team: the motivation is that an agent which predicts its partner's next move may need fewer messages and fewer handoffs to avoid duplicating work. The paper tests no message counts, so that is a hypothesis, not a result.

Goes deeper in: Agent Engineering → Agent Teams → Coordination Costs

Frequently Asked Questions

Check what you knowMap your AI & GPU knowledge across every track — free, role-based