Latent beats verbal: ToM2Vec outscores language ToM baseline by 9 points on Overcooked
EliasEskin · x · 2026-10-22
This post compares ToM2Vec (latent vectors) against a matched model trained to explicitly verbalize its partner's reasoning (Language ToM), isolating the effect of latent vs. discrete tokens.
- On Overcooked, Language ToM averages 97.6 vs. 106.7 for ToM2Vec, with the latent approach winning on all three layouts
- On the Coordination Ring layout, verbalized ToM slightly underperforms the untrained model, while ToM2Vec gains 11.6 points
Takeaway: learning abstract partner representations in latent space is more effective than requiring agents to express their partner model fully in text.
Related event: ToM2Vec latent vectors beat language baselines and generalize across models(3 posts)→
More from Research
- MIT launches course on self-evolving AI spanning context, harness, model and task evolution — pliang279 · 2026-10-23
- Daily Experiment Planner agent: one prompt to save scientists hours of lab prep — nlarusstone · 2026-10-23
- Phylo partners with Bayer to bring Biomni Lab AI co-scientist platform to its research teams — KexinHuang5 · 2026-10-23
- Study: 96% of ChatGPT memories are created unilaterally by the system — alex_verem · 2026-10-22
- AI2's OlmoEarth embeddings beat AlphaEarth and Tessera v2 on 10 of 11 evals — allen_ai · 2026-10-22
- Testing decision models like Jev and Drex Max for research triage: Drex Max heads to production — omarsar0 · 2026-10-22