ToM2Vec uses continuous theory-of-mind vectors to boost multi-agent coordination across 5 tasks
EliasEskin · x · 2026-10-22
EliasEkin shares results for ToM2Vec, a method that lets LLM agents produce continuous theory-of-mind (ToM) thought vectors to model teammates, outperforming trained and untrained language ToM baselines across five multi-agent domains.
- Overcooked: average score rises from 94.7 to 106.7 across three layouts; baselines barely improve
- Multi-party negotiation: deal success jumps from 62% to 71%, with both own score and collective utility up
- Decrypto: decoder and interceptor win rates improve by 4.4 and 6.1 points
- Hanabi: score climbs from 8.9 to 11.3 (out of 25); gains also in Collab Capture and Escape
The author argues text-based ToM forces agents to commit to a single partner hypothesis, while continuous vectors preserve a superposition of hypotheses. Experiments use Qwen3-8B.
Related event: ToM2Vec: Continuous Theory-of-Mind vectors help LLM agents model teammates(5 posts)→
More from coding & agent
- 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
- Free "AI University" on GitHub: 7 weeks from Docker to agentic RAG — mdancho84 · 2026-10-22
- Entoli: open-source cross-platform agent platform with your own keys, planning and P2P sync — ProgrammingDev · 2026-10-22
- 1Password launches Agents platform: login without ever showing passwords — dotey · 2026-10-22