ToM2Vec: continuous theory-of-mind vectors help LLM agents coordinate with unseen partners
EliasEskin · x · 2026-10-22
ToM2Vec trains agents to produce continuous theory-of-mind thought vectors about their partners, letting them hold multiple hypotheses at once instead of committing to a single verbalized reasoning path.
Problem: LLM agents often plan only for themselves, causing redundant or conflicting actions with partners.
Method: agents learn to predict a continuous abstract representation of a partner's reasoning; those latent ToM tokens then guide the agent's own reasoning and actions, with test-time latent vectors estimating the partner's thoughts.
Results: improved coordination with no task rewards across cooperative, competitive and mixed-motive settings (Overcooked, negotiation, Decrypto, Hanabi, Collab Escape). Versus a matched Language ToM baseline that verbalizes partner reasoning, ToM2Vec scores 106.7 vs 97.6 on Overcooked, winning on all three layouts.
Related event: ToM2Vec: Continuous Theory-of-Mind vectors help LLM agents model teammates(5 posts)→
More from coding & agent
- Composio launches Agent Firewall, an SLM-powered runtime boundary for agent actions — arankomatsuzaki · 2026-10-22
- How many tokens does your MCP setup burn before you've typed anything? A DIY measurement — LitheShade1X7 · 2026-10-22
- Split debug runs from scored runs to make your internal evals trustworthy again — Designer_Bet_7673 · 2026-10-22
- Dev builds an agent org chart in Copilot where each specialist keeps its own second brain — PaulShellDev · 2026-10-22
- Email can poison your AI assistant's recommendations — give approvers one source the agent can't touch — YvesMulkers · 2026-10-22
- Steve Yegge's agent autonomously researches a cert vulnerability via TheHackersNews — Steve_Yegge · 2026-10-22