Mental-Models for Multi-Agent Systems
A mental-model framework gives agents explicit Theory-of-Mind partner states that improve multi-agent decisions.
The paper introduces mental-model-enabled agents that maintain a latent model of a counterpart's beliefs, intentions, and likely reactions. The method learns an amortized recursive Theory-of-Mind representation with first- and second-order mental states, jointly with a belief-conditioned reward model used to train a policy. Evaluated on language-only and multimodal benchmarks, explicit partner modeling improves interaction quality and Theory-of-Mind performance over base agentic systems. Code is released in a public GitHub repository.
- Agents infer hidden beliefs, intentions, and likely reactions from observed history.
- Method learns first- and second-order amortized recursive Theory-of-Mind states.
- Belief-conditioned rewards train a policy that acts independently at inference.
- Explicit partner modeling beats base agents on language and multimodal benchmarks.
From the original article
1 min read at huggingface.co
Large foundation models have accelerated progress toward general-purpose agents that interact with humans and other agents through language and multimodal signals. However, robust multi-agent decision-making requires reasoning about what other agents know, intend, and are likely to do under partial observability.
By Hanan Gani, Lulu Shao, Manmohan Chandraker · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.