ToM2Vec trained via self-play with Qwen3-8B transfers to GPT-5.4 teams, adding 17.3 points
EliasEskin · x · 2026-10-22
Generalization experiments for ToM2Vec: the agent is trained purely via self-play with a Qwen3-8B partner, then paired with unseen models on Overcooked.
- With GPT-5.4, ToM2Vec improves team score by 17.3 points on average over pairing GPT-5.4 with an untrained Qwen3-8B agent — surprising given the much stronger partner.
- With Gemma-4, scores improve across all three layouts (e.g., AA 90→105).
- Against a matched Language ToM baseline, the latent approach averages 106.7 vs 97.6, winning all layouts.
Implication: partner modeling can be trained with open-weight models and deployed inside teams of stronger closed models.
Related event: ToM2Vec latent vectors beat language baselines and generalize across models(3 posts)→
More from Research
- Brain-inspired semiconductor chips aim for smarter, greener AI data centers — striketheviol · 2026-10-22
- How machine cognition research bridges to human CogSci: a Yale linguistics talk — najoungkim · 2026-10-22
- SARAL AI turns Indian public research into reels and posters for farmers — ponguru · 2026-10-22
- ETH Zurich maps economic theory papers into claim graphs with AI to weigh which errors matter — Afinetheorem · 2026-10-22
- Sheffield advertises funded PhD on reliability and bias of LLM-based evaluation — nikaletras · 2026-10-22
- New arXiv paper extracts the cruxes of LLM forecasts by contrasting 30 predictions — universeinanegg · 2026-10-22