Understanding Metacognition in Multi-Agent LLMs: Routing, Not Reasoning

Multi-agent reasoning and metacognitive strategies are widely employed to enhance large language model (LLM) performance; however, the functional role of metacognition remains unclear. Most existing approaches implicitly treat metacognition as a mechanism for generating improved or more diverse reasoning. In this work, we argue that metacognition primarily acts as an information routing and compression mechanism rather than a generator of new reasoning content. We introduce MC-MAS, an inference-time framework that separates problem solving from metacognitive arbitration, where independent solvers propose candidate answers and a metacognitive arbiter critiques and consolidates these outputs. Using routing-centric metrics such as semantic novelty, entropy reduction, and overconfidence errors, we analyze information flow across four reasoning benchmarks and multiple model settings. Results show that MC-MAS yields limited novelty but consistently reduces uncertainty and overconfident errors, with accuracy gains that are modest and dataset-dependent. We also find that reliable improvements arise from structured arbitration rather than reflection or increased sampling alone.