Probing ToM2Vec: latent vectors beat baselines on belief, desire and action prediction
EliasEskin · x · 2026-10-22
To understand what ToM2Vec's vectors encode, the authors directly probe agents about a partner's beliefs, desires and next actions (Bayesian ToM QA) in Overcooked.
- ToM2Vec scores highest in all three categories, beating the strongest per-category baselines by 5.6, 2.0 and 2.7 points (52.3% / 62.3% / 64.0% absolute).
- Evidence of belief revision: as the partner moves toward an onion or dish dispenser, the model shifts belief toward reasoning consistent with that trajectory.
The analyses support the interpretation that ToM2Vec learns partner representations that update with new evidence.
Related event: ToM2Vec latent vectors beat language baselines and generalize across models(3 posts)→
More from Research
- Sheffield advertises funded PhD on reliability and bias of LLM-based evaluation — nikaletras · 2026-10-22
- Scientific bottlenecks: reinvent assays, don't just automate them — owl_posting · 2026-10-22
- AI academic writing's vague-language trap: it reads as consistent to you but confuses everyone else — anshulkundaje · 2026-10-22
- YC hosts AI for Biology Paper Club as 81% of US doctors adopt AI — ycombinator · 2026-10-22
- CONCERT: Generative Modeling of Spatially Resolved Perturbation Responses — anshulkundaje · 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