Epistemic Miscalibration in LLM Multi-Agent Systems Addressed by New Workflow
A new arXiv paper (2605.23414) identifies a failure mode in LLM-based multi-agent systems called epistemic miscalibration in planning, where agents misjudge their knowledge when evaluating plan feasibility. Unlike execution errors, this miscalibration is latent and dynamic, making it hard to detect. The authors propose the Epistemic Planning Calibration Agentic Workflow (EPC-AW), which uses Information-consistency-based Plan Selection and Consistency-guided Epistemic Stat to assess plan stability under varying information conditions.
Key facts
- arXiv paper 2605.23414 introduces epistemic miscalibration in planning for LLM-based multi-agent systems.
- Epistemic miscalibration occurs when agents misjudge their knowledge during plan feasibility evaluation.
- This failure mode is latent and dynamic, unlike execution errors.
- The proposed solution is the Epistemic Planning Calibration Agentic Workflow (EPC-AW).
- EPC-AW uses Information-consistency-based Plan Selection.
- EPC-AW also employs Consistency-guided Epistemic Stat.
- The paper was announced on arXiv with type 'new'.
- The approach assesses plan support under varying information conditions.
Entities
Institutions
- arXiv