ToM2Vec: Latent Theory-of-Mind Vectors Boost LLM Agent Cooperation by 9 Points
EliasEskin · x · 2026-10-23
New paper ToM2Vec trains LLM agents to latently model their partner's reasoning via continuous theory-of-mind vectors capturing multiple hypotheses, instead of explicitly verbalizing them.
- On Overcooked, ToM2Vec beats the Language ToM baseline by 9.0 points on average.
- Direct ToM probes show gains of 5.6 / 2.0 / 2.0 percentage points on predicting partners' beliefs, desires, and next actions over the strongest baselines.
- Trained only with a Qwen3-8B partner, the improvements transfer to unseen partner models, including frontier ones like GPT-5.4 and Gemma 4.
- Gains hold across 5 multi-agent tasks (cooperative and competitive).
More from coding & agent
- AgentGUI at EMNLP 2026: a pixel-art interface for observing and steering AI agents — Michael_D_Moor · 2026-10-24
- Cisco's Antares: a security analyst agent that never leaves your laptop — aminkarbasi · 2026-10-24
- Chromium contributor asks why Claude Code and Codex don't use a headless browser — Kuprel · 2026-10-24
- 5 YC startups building infra for coding agents: Boxes, Caveman, Rystic, Simulithic — KlausCodes · 2026-10-24
- OpenAI DevEx lead to keynote BoxWorks Dev Summit on tools, memory and agentic work — OpenAIDevs · 2026-10-24
- Developer slams Kimi coding plan: 5-hour quota gone in under 10 minutes — jasonkneen · 2026-10-24