EPC-AW: Addressing Epistemic Misalignment Planning in Multi-Agent Systems

This article introduces the EPC-AW (Epistemic Planning Calibration Agentic Workflow) framework, which aims to address the epistemic misalignment planning problem in multi-agent systems—where plans may still fail even if executed correctly—caused by agents' incorrect assessment of their own knowledge states. The framework effectively improves system success rates through two core components: information-consistent plan selection and consistency-guided epistemic state refinement. Experiments show that EPC-AW increases the system's success rate by an average of 9.75%. This content is based on the paper When Planning Fails Despite Correct Execution: On Epistemic Calibration for LLM-Based Multi-Agent Systems published on arXiv on May 22, 2026 (link: http://arxiv.org/abs/2605.23414v1).

Epistemic misalignment planning is a hidden failure mode in multi-agent systems: agents make incorrect judgments about their own knowledge states, leading to plans that are self-consistent and executable but actually fail. For example, Agent A believes the red item is in the East Area based on outdated inventory records, and instructs Agent B to check the East Area (execution is correct), but the item is actually in the West Area. Unlike execution, communication, or coordination errors, epistemic misalignment is difficult to detect (no obvious error signals) and dynamic—new information may mask the misalignment or cause it to recur.

Comparison of failure types:

Failure Type	Performance	Detection Difficulty
Execution Error	Action execution failure	Easy
Communication Error	Message delivery failure	Medium
Coordination Error	Conflict between agents	Medium
Epistemic Misalignment	Incorrect assessment of one's own knowledge	Difficult

The EPC-AW framework includes two core components:

1. Information-Consistent Plan Selection: Verify the stability of candidate plans across agents (whether they hold under different information conditions). For example, Plan P has large evaluation differences (unstable) among logistics, finance, and time agents, while Plan Q has consistent evaluations (more optimal).
2. Consistency-Guided Epistemic State Refinement: Record historical epistemic differences, analyze patterns, and update agents' cognition (mark uncertain knowledge). For example, Agent A learns from failures due to outdated inventory records and adds verification steps in subsequent plans.

The workflow consists of 5 stages:

• Plan Generation: Each agent generates candidate plans
• Cross-Agent Evaluation: Calculate the score variance (consistency score) of the plan among agents
• Plan Selection: Choose the plan with the highest consistency score
• Execution and Monitoring: Execute the plan and collect feedback
• Epistemic Refinement: Update the epistemic state based on feedback (reinforce successful assumptions or correct failed ones)

Epistemic state representation includes known facts, uncertain facts, assumptions, confidence levels, and history; consistency evaluation is measured using score variance; epistemic refinement is achieved by adjusting confidence levels and recording history.

Experiments validated the effectiveness of EPC-AW in three tasks: collaborative logistics planning, distributed resource allocation, and collaborative problem-solving.

Comparison methods and results:

Method	System Success Rate	Relative Improvement
Baseline	~65%	-
Simple Consistency Check	~70%	+5%
EPC-AW	~74.75%	+9.75%

In-depth analysis:

• Reduction in failure modes: Plans based on outdated information (-40%), ignoring resource constraints (-35%), failing to consider time limits (-30%)
• Learning effect: As the number of interactions increases, the frequency of epistemic misalignment decreases, agents' knowledge assessments become more accurate, and system performance continues to improve.

The application prospects of EPC-AW are broad:

1. Enterprise Workflow Automation: Coordinate sales, production, and logistics departments to avoid execution failures caused by information asymmetry.
2. Intelligent Customer Service Systems: Ensure the feasibility of transfer and escalation processes in multi-agent customer service.
3. Robot Collaboration: Reasonably allocate tasks, considering the capabilities and positions of each robot.
4. Distributed AI Systems: Optimize task scheduling and coordinate data and capabilities across different nodes.

Current limitations:

1. Cross-agent evaluation increases communication overhead
2. Agent evaluations may be subjective
3. Epistemic refinement requires multiple interactions to converge
4. Effectiveness in highly dynamic and uncertain environments needs to be verified

Future directions:

1. Develop efficient cross-agent evaluation protocols
2. More refined confidence modeling and uncertainty quantification
3. Agents actively seek information to reduce epistemic misalignment
4. In-depth theoretical analysis of epistemic calibration

EPC-AW touches on the metacognition problem in AI systems—how AI recognizes the boundaries of its own knowledge. Metacognition is a key part of human intelligence, and EPC-AW introduces it into multi-agent systems: agents not only perform tasks but also evaluate their own knowledge; not only make plans but also evaluate the knowledge basis of the plans; not only learn facts but also learn the boundaries of knowledge.

Traditional AI focuses on 'what is known', while EPC-AW emphasizes 'knowing what you know': if you know you don't know, you don't plan blindly; if you know you are uncertain, you seek information; if you know your limitations, you act cautiously. This is key to building reliable AI systems.