Survey: Metacognition in LLMs — foundations, progress, and open problems for self-monitoring agents

arXiv·medium signal

A new arXiv survey (2607.11881) by Liu, Gani, Lu et al. maps metacognition in LLMs — the capacity to monitor, evaluate, and regulate one's own reasoning — as a foundation for reliable learning, planning, and decision-making. It's directly relevant to agent designs that rely on self-critique, confidence estimation, and adaptive tool use, consolidating a fragmented literature into a shared framework and research agenda. Useful as a grounding reference for teams building self-improving or self-verifying agent loops.