🤖 AI

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

This paper introduces the Epistemic Planning Calibration Agentic Workflow (EPC-AW) to address epistemic miscalibration in LLM-based multi-agent systems—where agents misjudge knowledge feasibility despite correct execution—by employing information-consistent plan selection and dynamic state refinement to improve system-level success by an average of 9.75%.

Original authors: Zehao Wang, Shilong Jin, Zhao Cao, Lanjun Wang

Published 2026-05-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zehao Wang, Shilong Jin, Zhao Cao, Lanjun Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Core Problem: "Confidently Wrong" Plans

Imagine you are leading a team of three experts (a Planner, a Doer, and a Checker) to solve a complex mystery.

Usually, when these teams fail, it's because someone made a mistake while doing the work. Maybe the Doer used the wrong tool, or the Checker missed a typo. We know how to fix those errors: just tell them, "Hey, that didn't work, try again."

But this paper discovered a sneaky new type of failure. Sometimes, the team follows the plan perfectly. The Doer uses the right tools, the Checker sees no errors, and the actions are executed exactly as written. Yet, the team still fails to solve the mystery.

Why? Because the Planner was overconfident.

The Planner looked at the information they had and said, "I'm 100% sure this plan will work!" But they were wrong. They didn't realize they were missing a crucial piece of the puzzle. The plan looked good on paper, but it was actually impossible to complete with the information available.

The authors call this "Epistemic Miscalibration." In plain English: The team is confidently wrong about what they know.

The Analogy: The Blind Hikers

Think of the team as hikers trying to cross a mountain range in the fog.

The Planner draws a map based on what they can see right now.
The Doer walks the path.
The Checker makes sure the Doer doesn't trip.

The Old Way (Execution Errors):
If the Doer falls into a hole, the Checker says, "Stop! You fell." The team fixes the path and keeps going.

The New Problem (Epistemic Miscalibration):
The Planner draws a path that looks clear. The Doer walks it perfectly. The Checker sees no holes. But the path leads to a cliff edge that the Planner couldn't see in the fog. The team walks all the way to the edge and falls off, even though they walked perfectly.

The problem isn't that they walked badly; it's that the map was flawed because the Planner was overconfident about what they could see.

The Solution: EPC-AW (The "Reality Check" Team)

The authors created a new workflow called EPC-AW to fix this. Instead of just checking if the Doer walked correctly, they check if the Plan makes sense to everyone, even when they have different information.

They use two main tricks:

1. The "Group Consensus" Check (Information-Consistency-based Plan Selection)

Instead of letting the Planner pick the best path alone, the system asks: "If we showed this plan to three different people who know slightly different things, would they all still think it's a good idea?"

The Trick: The system simulates different versions of the team, each with slightly different "memories" or information.
The Test: If the Planner thinks a path is great, but the "other versions" of the team think it's a bad idea because they know something the Planner doesn't, the system rejects that plan.
The Result: They only pick plans that everyone agrees on, regardless of who has what information. This filters out the "confidently wrong" plans.

2. The "Lessons Learned" Notebook (Consistency-guided Epistemic State Refinement)

Sometimes, the team makes the same mistake over and over. Maybe the Planner keeps trying to use a tool that doesn't work for a specific type of question.

The Trick: The system keeps a special notebook. Every time the Planner picks a plan that the "Group Consensus" check rejects, the system writes down why it was a bad idea.
The Result: Next time, the Planner looks at the notebook and says, "Oh, I remember I tried that before and it failed because I was missing info." This stops the team from making the same overconfident mistake twice.

What Did They Find?

The researchers tested this new method on six different "mystery-solving" challenges (like answering hard questions that require searching the internet or doing math).

The Result: By using this "Group Consensus" and "Lessons Learned" approach, the team solved 9.75% more problems than before.
The Takeaway: Even if you have the smartest AI and the best tools, you will still fail if the AI is too confident about what it knows. You need a system that constantly checks, "Are we sure about this?" before taking action.

Summary

This paper teaches us that in AI teams, being wrong isn't always a mistake in execution; sometimes it's a mistake in confidence. By forcing the AI to check its plans against different perspectives and learn from past overconfidence, we can build systems that are much harder to fool.

Technical Summary: Epistemic Calibration for LLM-Based Multi-Agent Systems

1. Problem Definition: Epistemic Miscalibration in Planning

The paper identifies a distinct failure mode in Large Language Model (LLM)-based multi-agent systems termed epistemic miscalibration in planning. Unlike traditional execution errors (e.g., invalid tool outputs or logical reasoning flaws), this failure occurs even when all actions are executed correctly according to the plan.

The core issue arises when a planning agent assigns unwarranted confidence to the feasibility of a plan based on partial information. The agent fails to recognize the limits of its knowledge in tasks requiring iterative information acquisition. Consequently, the system may generate a plan that appears self-consistent and executable but is fundamentally infeasible with respect to the objective. This miscalibration is characterized by two properties:

Latency: It leaves no explicit error signals during the planning phase; the plan appears coherent, making it difficult to detect before execution.
Dynamics: As new information is acquired, feasibility assessments evolve. Past miscalibration signals can be obscured by new data, causing the same planning errors to recur over time.

Existing repair methods (post-hoc debugging or online execution monitoring) primarily address observable execution errors and rely on explicit failure signals. They fail to address the latent, plan-level epistemic miscalibration that persists even under correct execution.

2. Methodology: EPC-AW

To mitigate this issue, the authors propose the Epistemic Planning Calibration Agentic Workflow (EPC-AW). Rather than attempting to directly verify plan feasibility (which is often impossible without full information), EPC-AW assesses whether a plan remains supported by agents under varying information conditions. The workflow operates through three specialized agents: a Planner, an Executor, and a Diagnoser, each maintaining heterogeneous private memories.

EPC-AW consists of two complementary components:

A. Information-Consistency-based Plan Selection (IPS)

Operating within a single planning round, IPS serves as a diagnostic tool to select robust plans without explicit error signals.

Mechanism: The Planner generates a set of candidate plans. Each agent (Planner, Executor, Diagnoser) evaluates these plans based on its own private information state.
Cross-Agent Prediction: Agents predict how their peers would evaluate the same plans given the peers' distinct information states.
Selection Criteria: Plans are scored based on the consistency of their evaluations across these heterogeneous information states. Plans whose evaluations vary significantly across agents are deemed "epistemically fragile." The system selects the plan with the highest cross-agent consistency score, assuming that a plan robust to information variations is less likely to suffer from miscalibration.

B. Consistency-guided Epistemic State Refinement (CESR)

Operating across multiple rounds, CESR addresses the dynamic nature of miscalibration by learning from past discrepancies.

Mechanism: The system compares the plan selected by the Planner (based on its own confidence) with the plan selected by IPS (based on cross-agent consistency).
Constraint Induction: When a divergence occurs (i.e., the Planner's choice differs from the IPS choice), the Diagnoser generates a lightweight epistemic constraint. This constraint abstracts the structural features or information gaps that led to the Planner's misjudgment.
Memory Integration: These constraints are accumulated in the Planner's persistent memory. Future planning is guided by these constraints, preventing the recurrence of specific miscalibration patterns as the system accumulates new evidence.

3. Key Contributions

The paper makes the following specific contributions:

Formalization of a New Failure Mode: It defines epistemic miscalibration in planning as a distinct repair target, highlighting failures that arise from miscalibrated feasibility assessments even when execution is flawless.
IPS Mechanism: It introduces a method to select plans that remain stable across agents with heterogeneous information, addressing the lack of observable error signals in planning.
CESR Mechanism: It proposes a memory-driven approach to adapt epistemic calibration over time, using past discrepancies to constrain future planning and suppress persistent misjudgments.
Empirical Validation: Extensive experiments on six benchmarks demonstrate the efficacy of the approach.

4. Experimental Results

The authors evaluated EPC-AW on six diverse benchmarks: Bamboogle, 2Wiki, HotpotQA, Musique, GAIA, and MedQA. These datasets span reasoning-intensive inference to search-driven open-world problem solving.

Performance Gains: EPC-AW achieved an average absolute improvement of 9.75% in system-level success rate compared to the "No-Repair" baseline (standard AgentFlow).
Comparison with Baselines: EPC-AW outperformed existing repair strategies, including Retry (local metacognitive re-generation) and Rollback (system-level state recovery). For instance, on the 2Wiki benchmark, EPC-AW improved accuracy by 12.83% over the baseline, significantly outperforming Retry (+9.00%) and Rollback (+7.00%).
Ablation Studies: Removing either IPS or CESR resulted in performance degradation, confirming that both intra-round consistency selection and cross-round constraint feedback are essential.
Generalization: The method demonstrated robustness across different LLM backbones (Qwen3-14B, Qwen3-30B, DeepSeek-R1-32B), suggesting the approach is not dependent on specific model architectures.
Cost Analysis: The additional computational overhead introduced by IPS (parallel evaluation of candidate plans) and CESR (constraint injection) is minimal. The asymptotic time complexity remains $O(T \cdot (\tau_{llm} + \tau_{tool}))$ , with token costs scaling primarily with existing memory and retrieval contexts rather than introducing new bottlenecks.

5. Significance and Claims

The paper claims that EPC-AW shifts the paradigm of failure mitigation from execution-time correction to planning-time epistemic calibration. By treating epistemic calibration as a system-level consideration, the approach enables multi-agent systems to better identify unreliable plans and revise reasoning trajectories before errors propagate.

The authors position this work as a step toward more reliable deployment of LLM-based multi-agent systems in environments with incomplete and evolving information. They emphasize that their method improves task success by reducing unsupported feasibility assessments without introducing new capabilities or ethical concerns beyond those already studied in LLM deployment. The work underscores the importance of addressing latent planning failures that are invisible to traditional execution monitoring.