Adaptive Theory of Mind for LLM Multi-Agent Coordination: ToM Order Misalignment Impairs Collaboration (AAAI 2026)

arXiv 2603.16264·medium signal

This AAAI 2026 paper identifies a previously uncharacterized failure mode in multi-agent LLM systems: when agents reason at different depths of theory-of-mind (one agent models the other shallowly, the other deeply), coordination degrades from both insufficient and excessive meta-reasoning. The proposed A-ToM agent dynamically estimates its partner's ToM order from prior interactions and aligns its own reasoning depth accordingly. Evaluated on four tasks (matrix games, grid navigation, Overcooked) with generalizability to non-LLM agents.