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

In three linesLLM-based multi-agent systems fail despite correct plan execution when agents misjudge their knowledge (epistemic miscalibration). Authors propose EPC-AW workflow that selects plans stable across agents and refines epistemic state over time using past discrepancies. 9.75% improvement in system-level success.

Summary generated by Claude — human-verified