ToM2Vec: Continuous Theory-of-Mind Vectors Boost LLM Multi-Agent Collaboration
A UNC Chapel Hill team, together with Cisco and UT Austin, published ToM2Vec, a paper tackling LLM agents' difficulty inferring teammates' latent states in multi-agent collaboration. It encodes a teammate's Theory of Mind (ToM) as a continuous latent reasoning vector within a fixed budget—rather than locking in a single reasoning path like verbalized ToM—allowing multiple hypotheses about the teammate's thoughts to be held simultaneously.
Confirmed
- Training mechanism: the main agent predicts continuous ToM thinking vectors from its own observations and interaction history to anticipate teammates' reasoning; during training, teammates generate chain-of-thought (CoT) traces, from which 10 hidden-state vectors are uniformly sampled as latent supervision targets.
- Results: outperforms both trained and untrained verbalized-ToM baselines across five multi-agent tasks, with Overcooked winning on average across three layouts.
- In Overcooked, researchers directly queried agents about teammates' beliefs, desires, and next actions (a comparison against Bayesian ToM); the method transfers to unseen partner models.
Why It Matters
- Verbalized ToM tends to converge prematurely to a single reasoning path; ToM2Vec's continuous vector representation preserves multiple hypotheses, offering a more flexible and transferable approach to teammate modeling for multi-agent human-AI collaboration.
2026-10-22 ~ 2026-10-23 · 9 related posts
Primary sources
- ToM2Vec: continuous theory-of-mind tokens improve multi-agent coordination — UNC-ChapelHill ·
- ToM2Vec: continuous theory-of-mind vectors help LLM agents coordinate with unseen partners — EliasEskin ·
- How ToM2Vec works: 10 hidden-state vectors from a partner's CoT supervise latent ToM tokens — EliasEskin ·
- [source] ToM2Vec: continuous theory-of-mind tokens improve multi-agent coordination — UNC-ChapelHill · 2026-10-22
- [source] ToM2Vec: continuous theory-of-mind vectors help LLM agents coordinate with unseen partners — EliasEskin · 2026-10-22
- [source] How ToM2Vec works: 10 hidden-state vectors from a partner's CoT supervise latent ToM tokens — EliasEskin · 2026-10-22
- Latent beats verbal: ToM2Vec outscores language ToM baseline by 9 points on Overcooked — EliasEskin · 2026-10-22
- ToM2Vec trained via self-play with Qwen3-8B transfers to GPT-5.4 teams, adding 17.3 points — EliasEskin · 2026-10-22
- Probing ToM2Vec: latent vectors beat baselines on belief, desire and action prediction — EliasEskin · 2026-10-22
- ToM2Vec encodes Theory of Mind as vectors, beating baselines in Overcooked agent QA — EliasEskin · 2026-10-22
2 near-duplicate retellings: EliasEskin · EliasEskin