Announcement
ToM2Vec uses non-text tokens to model other AI agents’ reasoning
The researchers report gains across five multi-agent tasks, including when agents work with previously unseen partner models.
TLDR
The ToM2Vec researchers say their method trains AI agents to generate continuous tokens representing possible partner reasoning, rather than settling on one description in text. They report improved results across five multi-agent settings, including Overcooked and negotiation. They also say gains from training with a Qwen3-8B partner carried over when the agent was paired with previously unseen GPT-5.4 and Gemma-4 partners.
Combined views
4
10 Sources, first seen ago
