arXiv cs.AIMonday · May 25, 2026FREE

When Planning Fails Despite Correct Execution: On Epistemic Calibration for LLM-Based Multi-Agent Systems

A new paper on arXiv (2605.23414v1) identifies epistemic miscalibration as a failure mode in LLM-based multi-agent systems, where agents correctly execute planned actions but misjudge their knowledge when evaluating plan feasibility. Unlike execution errors, this miscalibration is latent during planning (plans remain self-consistent and executable) and dynamic (new information can alter feasibility assessments, obscuring past signals). To address this, the authors propose the Epistemic Planning Calibration Agentic Workflow (EPC-AW), which assesses whether plans remain supported under varying information conditions rather than directly verifying feasibility. EPC-AW includes Information-consistency-based Plan Selection (selecting plans with stable evaluations across agents) and Consistency-guided Epistemic State Refinement (adapting calibration over time using past discrepancies). Experiments demonstrate an average 9.75% improvement in system-level success. The paper was published on May 25, 2026.