Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.12453.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

The paper introduces mental-model-enabled agents, a framework that learns an amortized recursive Theory-of-Mind representation of a counterpart's beliefs, intentions, and likely reactions, combined with a belief-conditioned reward model to guide action selection, evaluated on language-only and multimodal benchmarks.

Frames partner-state as a reusable decision variable across tasks, with released code. Useful for designers building agents that must reason about other agents under partial observability.

Cited by

No citations on record.