Confidence Calibration and Reliability Assessment in Multi-Agent LLM Systems: A Focused Structured Review

Large Language Models (LLMs) have demonstrated strong performance in various natural language processing tasks, but their ability to accurately estimate and communicate confidence in their outputs remains a significant challenge. High task accuracy does not necessarily imply that a model’s confidence reflects the likelihood of correctness. This focused structured review examines the development of confidence estimation, uncertainty quantification, and reliability assessment from conventional neural-network calibration to modern LLM and multi-agent settings. The review first analyzes five foundational studies covering probabilistic calibration, model self-evaluation, reasoning consistency, evidence-based multi-hop reasoning, and broader LLM confidence estimation. It then incorporates recent work on multi-agent confidence calibration, trajectory-level uncertainty quantification, runtime calibration, communication-aware calibration, and epistemic calibration during multi-agent planning. The selected studies are analyzed according to their objectives, confidence or reliability signals, methodologies, evaluation strategies, and limitations. The synthesis highlights a transition from individual-model confidence toward system-level reliability and identifies five research gaps: local-to-system calibration and confidence propagation, correlated errors and agent disagreement, standardized multi-agent calibration benchmarks, explainable confidence traces, and runtime shift with computational cost. The review concludes by outlining corresponding future directions for reliable confidence calibration and system-level assessment in multi-agent LLM systems.