Skip to content

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

🕒 Published (v1): 2026-05-22 09:24 UTC · Source: Arxiv · link

Why this paper was selected

Epistemic miscalibration: novel failure mode where agents misjudge plan feasibility

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

LLM-based multi-agent systems can fail even when all actions execute correctly, because the planning agent may have miscalibrated confidence in plan feasibility — a failure mode distinct from execution errors. This paper formalizes epistemic miscalibration in planning and proposes EPC-AW, a workflow that selects plans by cross-agent evaluation consistency rather than direct feasibility checks, achieving a +9.75% average accuracy improvement across six benchmarks.

Problem

Prior multi-agent failure-repair work targets execution-level faults (wrong tool outputs, invalid returns) and relies on observable error signals. It misses a separate class of failures where plans are logically coherent and fully executable yet systematically fail to satisfy the task objective — because the planner assigns unwarranted confidence to a feasibility assessment formed under incomplete information. This miscalibration is both latent (no observable error signal) and dynamic (continual information updates shift and obscure past miscalibration signals, causing recurrence).

Method

EPC-AW is a three-agent (Planner \(A_P\), Executor \(A_E\), Diagnoser \(A_D\)) workflow with differentiated private memories that induce heterogeneous epistemic states.

Information-consistency-based Plan Selection (IPS) operates within each round. The Planner generates \(K=9\) candidate plans \(\Pi^{(t)}\). Each agent \(i\) both self-evaluates plans and predicts how other agents would evaluate them given approximated alternative information states \(\hat{\gamma}_{i \to j} \triangleq (U_j, M_i^{(t)})\). The plan-level consistency score is: $\(C_\text{IPS}(\pi_k^{(t)}) = \frac{1}{|A|}\sum_{i \in A} \left[\log e_i^{(t)}(k) - \log \bar{e}_{-i}^{(t)}(k)\right]\)$ and the selected plan maximizes this score.

Consistency-guided Epistemic State Refinement (CESR) operates across rounds. When IPS selects a plan differing from the Planner's self-ranked plan (\(\pi_P^{(t)} \neq \pi_*^{(t)}\)), the Diagnoser generates an epistemic constraint \(c^{(t)}\) capturing structural features of the miscalibration. This constraint is appended to the Planner's private memory, preventing the same pattern from recurring as the information context evolves.

Key Contributions

  • Formal definition of epistemic miscalibration in planning as a distinct system-level failure mode in LLM-based MAS, separate from execution errors.
  • IPS: a planning-time calibration signal based on cross-agent evaluation stability under heterogeneous information, requiring no external ground-truth supervision.
  • CESR: a memory-driven mechanism that accumulates cross-round consistency discrepancies as persistent epistemic constraints to prevent miscalibration recurrence.
  • Evaluation on six benchmarks (Bamboogle, 2Wiki, HotpotQA, Musique, GAIA, MedQA) demonstrating consistent improvement over three baselines.

Results

All numbers are accuracy (%) with Qwen3-Coder-30B on four NVIDIA RTX 4090s:

  • EPC-AW vs. No-Repair (AgentFlow): +9.75% average absolute improvement across all six benchmarks.
  • EPC-AW vs. Rollback (COCO-inspired): +4.64% average improvement.
  • EPC-AW vs. Retry (MAST-inspired): +6.86% average improvement.
  • Per-benchmark gains over No-Repair: Bamboogle +9.86%, 2Wiki +12.83%, HotpotQA +12.67%, Musique +8.00%, GAIA +8.13%, MedQA +7.00%.
  • All three repair-based methods beat No-Repair; EPC-AW achieves best accuracy on every benchmark.

Limitations

  • Evaluated with a single backbone (Qwen3-Coder-30B); generalization to other LLM families is untested.
  • IPS requires generating \(K\) candidate plans per round and evaluating them across all \(|A|\) agents, adding non-trivial inference overhead not quantified in the paper.
  • Cross-agent evaluation prediction uses the agent's own information to approximate another agent's state (\(\hat{\gamma}_{i \to j} \triangleq (U_j, M_i^{(t)})\)); this approximation quality degrades when private memories diverge strongly.
  • CESR's epistemic constraints are generated by the Diagnoser (itself an LLM), so the constraint quality inherits the same potential for miscalibration it is meant to correct.
  • The paper text is truncated; ablation details for IPS vs. CESR independently are not visible in the provided excerpt.

Relevance to Harnesses / Meta-Harnesses

EPC-AW is a meta-level workflow wrapped around an existing agent framework (AgentFlow), diagnosing and correcting plan quality without modifying the base agent internals — exactly the architectural pattern of a meta-harness. The IPS mechanism is a form of cross-agent consensus used as an intrinsic quality signal, directly relevant to harnesses that need to select or validate intermediate outputs from sub-agents without external ground truth. CESR's persistent epistemic constraint accumulation is a concrete mechanism for harness-level memory that shapes future sub-agent behavior across rounds, a pattern applicable to any iterative multi-agent pipeline. The formal separation of planning-time calibration from execution-time correction is a useful design principle for harness authors deciding where to inject monitoring and repair logic.