BayesBeliefAgent replans only when a partner contradicts the inferred skill, cutting replans by an order of magnitude

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv:2608.18490 (submitted 19 Aug 2026) targets a failure mode familiar to anyone running multi-agent systems: agents keep executing stale plans after a teammate switches strategy. The method pairs hierarchical LLM planning with Bayesian partner tracking and interrupts only when a partner's actions directly contradict the inferred skill. On Overcooked it reduces what the authors call the belief-action gap, cases where an agent holds an accurate partner model but still acts against it, while using roughly an order of magnitude fewer replans than heuristic triggers.