EPC-AW: Epistemic Calibration in Multi-Agent Planning

• EPC-AW is a planning-centered framework that addresses epistemic miscalibration by evaluating candidate plans for stability under heterogeneous information.
• It implements dual calibration mechanisms—IPS for within-round evaluation and CESR for cross-round refinement—to mitigate latent planning failures.
• Empirical results demonstrate up to a 9.75% improvement over baseline methods across multiple benchmarks, underscoring its practical efficacy.

Epistemic Planning Calibration Agentic Workflow (EPC-AW) is a planning-centered framework for LLM-based multi-agent systems that addresses a specific failure mode: plans can be executable yet epistemically unsound. The term is introduced in the context of “epistemic miscalibration in planning,” where the planner assigns unjustified confidence to a plan’s feasibility under partial information even when tool calls succeed, outputs are valid, and the reasoning trace is coherent (Wang et al., 22 May 2026). In this formulation, calibration is not treated as confidence estimation alone, but as a workflow-level mechanism for assessing whether candidate plans remain supported under varying information conditions and for refining future planning through memory of prior discrepancies. Related work in epistemic planning, expectation-aware planning, and auditable multi-agent systems provides the formal and systems background against which EPC-AW is most naturally understood (Hu et al., 2019, Engesser et al., 2017, Sreedharan et al., 2019, Fabiano, 2021, Burigana et al., 2020, Stoffl et al., 13 Apr 2026).

1. Definition and problem setting

EPC-AW is proposed to address failures that occur one level earlier than execution: at the planning stage, where a planner misjudges whether a proposed plan is actually supportable by the information currently available (Wang et al., 22 May 2026). The framework distinguishes two failure types. Execution errors are observable during runtime and include incorrect tool outputs, invalid actions, broken reasoning traces, and unsupported execution results. Epistemic miscalibration in planning occurs when a planner chooses a plan whose goal or action structure is not actually sufficient to reach the objective given the available information, even though the plan may remain internally coherent and executable (Wang et al., 22 May 2026).

The EPC-AW formulation emphasizes two properties of this failure mode. It is latent because there may be no explicit error signal: no failed tool call or invalid output reveals the issue. It is also dynamic because feasibility judgments change as new evidence accumulates, and later information can hide earlier miscalibration, allowing similar failures to recur over time (Wang et al., 22 May 2026).

The formal setup treats the system as a multi-agent LLM workflow $M$ processing a user query $Q$. At round $t$, the system maintains an information context $I(t)$ and generates an intermediate plan

$Tr(t) = (g(t), a(t)),$

where $g(t)$ is the intermediate goal and $a(t)$ is the action intended to acquire relevant evidence. The action yields an observation $e(t)$ and updates the information context through

$I(t+1) = D(I(t), Tr(t), e(t)).$

Epistemic miscalibration is then characterized by a mismatch between a subjective feasibility judgment under partial information,

$J(Tr(t)\mid I(t)),$

and objective feasibility under the latent complete factual situation $Q$0,

$Q$1

This separation makes explicit that planning failure can arise even when execution is locally correct (Wang et al., 22 May 2026).

A plausible implication is that EPC-AW is best understood as a workflow for support-sensitive planning rather than feasibility certification in the strict sense: it does not directly prove that a plan is feasible, but evaluates whether plan assessments remain stable across heterogeneous information states.

2. Core architecture and calibration mechanisms

EPC-AW uses three role-specialized agents: a Planner $Q$2, an Executor $Q$3, and a Diagnoser $Q$4 (Wang et al., 22 May 2026). The workflow separates planning, execution, and diagnosis across rounds $Q$5. Its central design principle is that these roles do not share the same epistemic state. Instead, each agent has a different private memory, inducing heterogeneous information states.

The complete interaction history is

$Q$6

where $Q$7 is the plan, $Q$8 the execution outcome, and $Q$9 the diagnostic feedback. A binary failure indicator is defined for whether the diagnostic signal is UNSUPPORTED, and the system-level memory is specified as

$t$0

where $t$1 is the set of verifiable evidence accumulated so far and $t$2 is the abstract role description (Wang et al., 22 May 2026). Agent-level private memories are role-specific projections of history: the Planner stores plans that were not supported, the Diagnoser stores supported executions, and the Executor keeps the full trace.

Within this architecture, EPC-AW comprises two calibration components. The first is Information-consistency-based Plan Selection (IPS), the within-round mechanism. At each round, the Planner generates a set of candidate plans

$t$3

Each agent evaluates each candidate under its own information state, and each agent also predicts how other agents would score the same plans. The system then compares an agent’s own evaluation with the aggregated predicted evaluation from peers. The agent-level information-consistency score is

$t$4

and the plan-level IPS score is

$t$5

IPS selects

$t$6

High $t$7 indicates that a plan is stable under heterogeneous information, whereas low $t$8 indicates epistemic fragility or dependence on one agent’s private, potentially miscalibrated view (Wang et al., 22 May 2026).

The second component is Consistency-guided Epistemic State Refinement (CESR), the across-round mechanism. CESR compares the Planner’s own top choice,

$t$9

with the IPS-selected plan $I(t)$0. If they differ, the discrepancy is treated as a calibration signal. The Diagnoser converts this divergence into a lightweight epistemic constraint

$I(t)$1

which is then incorporated into planner memory through

$I(t)$2

CESR therefore turns persistent mismatches into memory constraints that shape future planning rather than replacing it (Wang et al., 22 May 2026).

The stepwise EPC-AW pipeline is: candidate plan generation, information-conditioned evaluation, cross-agent evaluation prediction, information-consistency aggregation, IPS-based plan selection, execution, diagnosis, cross-round refinement, and termination when enough evidence has been gathered to answer the query (Wang et al., 22 May 2026).

3. Epistemic planning foundations

EPC-AW belongs to a broader lineage of epistemic planning methods that reason not only over world states but also over knowledge and belief states. In this literature, Epistemic Planning concerns planning in the space of knowledge or belief states, while Multi-Agent Epistemic Planning extends the setting to multiple agents and information flows between them (Fabiano, 2021).

Several formal substrates are relevant. One major line uses Dynamic Epistemic Logic (DEL) with pointed Kripke structures. In “Cooperative Epistemic Multi-Agent Planning for Implicit Coordination,” an epistemic model is given as

$I(t)$3

with local states $I(t)$4 representing agent-specific perspectives (Engesser et al., 2017). Perspective shifts are central: the paper proves

$I(t)$5

and uses this to define implicit coordination, where agents do not negotiate a joint policy at plan time but must be able to justify their own steps from their own perspectives (Engesser et al., 2017). This perspective-sensitive notion of plan validity is closely aligned with EPC-AW’s emphasis on information-conditioned support.

Another line uses functional STRIPS and external functions instead of explicit Kripke accessibility relations. “What you get is what you see: Decomposing Epistemic Planning using Functional STRIPS” models each agent $I(t)$6 with a perspective function

$I(t)$7

subject to

$I(t)$8

The semantics define seeing and knowledge through visibility of variables: $I(t)$9 and

$Tr(t) = (g(t), a(t)),$0

This decomposes planning and epistemic evaluation: the planner handles search, while an external solver evaluates epistemic formulae lazily when needed for goals or preconditions (Hu et al., 2019). A plausible implication is that this decomposition anticipates EPC-AW’s treatment of epistemic assessment as a delegated calibration layer rather than as an intrinsic property of the search procedure.

A third line is expectation-aware planning, which treats calibration between the agent’s model and a human’s model as a single planning problem. “Expectation-Aware Planning” defines an EA problem as

$Tr(t) = (g(t), a(t)),$1

with solution

$Tr(t) = (g(t), a(t)),$2

where $Tr(t) = (g(t), a(t)),$3 is a set of model updates or explanations and $Tr(t) = (g(t), a(t)),$4 the executed plan (Sreedharan et al., 2019). Explanatory actions are compiled into classical planning through belief fluents and meta-fluents, and the approach unifies explanations, explicability, and self-explaining plans (Sreedharan et al., 2019). This suggests a direct conceptual connection: EPC-AW also treats calibration as part of planning rather than as a post hoc evaluation layer.

4. Workflow semantics, memory, and information heterogeneity

The distinctive feature of EPC-AW is that epistemic calibration is operationalized through heterogeneous private memories rather than through a single shared epistemic state (Wang et al., 22 May 2026). This stands in contrast to symbolic planners that operate over an explicitly shared structure, such as DEL models or possibilities-based semantics.

In the possibilities-based ASP planner PLATO, for example, a planning domain is defined as

$Tr(t) = (g(t), a(t)),$5

and planning proceeds by searching for a sequence of agent-action instances whose transition function

$Tr(t) = (g(t), a(t)),$6

yields a state satisfying the epistemic goal formula (Burigana et al., 2020). PLATO represents epistemic states with possibilities whose entailment relation supports individual belief, group belief, and common belief, and implements planning in multi-shot ASP using base, step(t), and check(t) subprograms (Burigana et al., 2020). The declarative organization separates entailment checking, initial-state generation, transition computation, and goal checking.

“Comprehensive Multi-Agent Epistemic Planning” similarly emphasizes explicit representations of world states plus multi-agent knowledge or belief states, using epistemic logic and Dynamic Epistemic Logic with pointed Kripke structures, and distinguishes S5 knowledge from KD45 belief (Fabiano, 2021). The manuscript also describes a modular C++ epistemic planner and an ASP version supporting multiple transition and update functions, custom update models, observability-based updates, and attitude-aware updates for trust, lies, and misconception (Fabiano, 2021).

Against this background, EPC-AW can be seen as relocating calibration from an explicit symbolic entailment engine to a cross-agent consistency process over role-specific memories. This is not a claim of equivalence. Rather, it suggests that EPC-AW uses information heterogeneity itself as the signal for calibration, whereas earlier epistemic planning systems more commonly encode the epistemic state directly and then test formula entailment or plan applicability relative to that encoded structure.

The workflow’s use of past discrepancies as persistent memory constraints also has a close analogue in iterative epistemic updating. However, EPC-AW’s constraints are described as lightweight epistemic constraints generated from mismatches between planner-preferred and IPS-selected plans, rather than as full structural updates of an explicit Kripke or possibilities model (Wang et al., 22 May 2026).

5. Empirical performance and demonstrated gains

EPC-AW is evaluated on six benchmarks: Bamboogle, 2Wiki, HotpotQA, Musique, GAIA, and MedQA (Wang et al., 22 May 2026). The main backbone is Qwen3-Coder-30B, with additional tests on Qwen3-14B and DeepSeek-R1-32B. The Planner samples $Tr(t) = (g(t), a(t)),$7 candidate plans with temperature $Tr(t) = (g(t), a(t)),$8, other steps use temperature $Tr(t) = (g(t), a(t)),$9, the maximum number of rounds is $g(t)$0, and the tool set includes a base generator, Python coder, Google Search, Wikipedia Search, and Web Search (Wang et al., 22 May 2026). GPT-4o is used as the automatic judge, each experiment is repeated three times, and the evaluation metric is accuracy or task success rate (Wang et al., 22 May 2026).

The framework is compared against three baselines implemented in AgentFlow: No-Repair, Retry, and Rollback (Wang et al., 22 May 2026). The reported average accuracies are 36.52% for No-Repair, 39.41% for Retry, and 41.63% for Rollback, while EPC-AW achieves the best performance on all six benchmarks, with an average absolute improvement of 9.75% over No-Repair and a 4.64% gain over Rollback (Wang et al., 22 May 2026). The paper notes especially strong gains in search-heavy and multi-step settings where epistemic miscalibration is most damaging (Wang et al., 22 May 2026).

The ablations isolate the contributions of the two calibration mechanisms. EPC-AW* removes both IPS and CESR, approximating planning without epistemic repair, while EPC-AW+ keeps IPS but disables CESR (Wang et al., 22 May 2026). The findings are that IPS alone can help in some reasoning-dominant settings but may become overly conservative on search-intensive tasks, and that CESR is critical because it prevents this conservatism from reducing exploration by turning cross-round discrepancies into structured guidance (Wang et al., 22 May 2026). The paper also compares IPS with Mean-Score Aggregation (MSA) and reports that IPS is more effective on multi-step reasoning and exploration tasks because it favors agreement under heterogeneous information rather than simple averaging (Wang et al., 22 May 2026).

Sensitivity analysis over plan-sample sizes $g(t)$1 shows that $g(t)$2 is weakest, performance improves monotonically as $g(t)$3 increases, the largest gain is from $g(t)$4 to $g(t)$5, and subsequent gains taper off (Wang et al., 22 May 2026). The method is therefore described as benefiting from epistemic diversity in candidate generation. The paper further argues that EPC-AW preserves the same asymptotic complexity as the base workflow,

$g(t)$6

with IPS introducing mostly a constant-factor increase that can be parallelized, and CESR adding only a small amount of extra context (Wang et al., 22 May 2026).

6. Auditable instantiations, scope, and limitations

A concrete instantiation of an EPC-AW-like pattern appears in VERITAS, a multi-agent system for image-derived hypothesis testing on multimodal clinical datasets (Stoffl et al., 13 Apr 2026). VERITAS decomposes the end-to-end study cycle into four sequential phases handled by role-specialized agents: Principal Investigator (PI), Medical Imaging Specialist, Statistician, and a phase-aware Critic (Stoffl et al., 13 Apr 2026). The phases are Analysis planning, Segmentation, Statistical analysis, and Interpretation (Stoffl et al., 13 Apr 2026).

Phase 1 produces a structured analysis plan specifying target cohorts, required structures, observations or timepoints, derived measurements, intended statistical test, feasibility status as TESTABLE or UNTESTABLE, missing requirements if untestable, and a priori power based on queried sample sizes (Stoffl et al., 13 Apr 2026). If the plan is marked untestable, the run terminates as Invalid. Phase 2A uses the promptable segmentation backend SAT to produce binary masks stored in a shared database. Phase 2B generates executable Python code that computes quantitative metrics, runs the planned statistical test, outputs sample sizes, effect size, confidence interval, and p-value, and saves results in a structured JSON artifact. Phase 3 yields a final verdict YES, NO, or INCONCLUSIVE without access to the evaluator’s final evidence label (Stoffl et al., 13 Apr 2026).

VERITAS introduces a deterministic evidence-labeling layer, the Evidence Classification Operator (ECO), that maps outcomes to Supported, Refuted, Underpowered, or Invalid based on significance, effect direction, and study power (Stoffl et al., 13 Apr 2026). The core rule is:

• Supported: $g(t)$7 direction matches
• Refuted: $g(t)$8 opposite direction$g(t)$9
• Underpowered: $a(t)$0
• Invalid: Untestable or execution failure with $a(t)$1 and $a(t)$2 denoting power at a smallest effect size of interest (Stoffl et al., 13 Apr 2026).

The benchmark contains 64 hypotheses spanning six complexity levels across ACDC and UCSF-PDGM MRI datasets, and the system reaches 81.4% verdict accuracy with frontier models and 71.2% with locally-hosted open-weight models, while producing independently verifiable statistical outputs at rates of 86.6% and 77.3% respectively (Stoffl et al., 13 Apr 2026). VERITAS is therefore significant not because it defines EPC-AW formally, but because it instantiates an auditable, planning-oriented, epistemic calibration workflow in a concrete scientific domain.

The limitations of EPC-AW and related systems are explicit. EPC-AW assumes that the system can generate multiple candidate plans, that agents have heterogeneous private memories, that the Diagnoser can detect support mismatches and convert them into useful constraints, and that cross-agent consistency is a useful proxy for epistemic support (Wang et al., 22 May 2026). IPS is a heuristic rather than a proof of feasibility, and a plan can remain consistent across agents yet still fail (Wang et al., 22 May 2026). In highly search-oriented tasks, overly conservative plan selection can reduce exploration if CESR does not counterbalance it (Wang et al., 22 May 2026).

The broader epistemic planning literature identifies additional limits. The perspective-function approach in functional STRIPS is naturally suited to knowledge, not arbitrary false belief, because $a(t)$3 collapses uncertainty into “not seen” and cannot fully preserve disjunctive uncertainty the way Kripke models can (Hu et al., 2019). DEL-based decentralized planning is computationally expensive, assumes exact epistemic reasoning and explicit event models, and leaves issues such as livelocks and replanning open (Engesser et al., 2017). PLATO is competitive but incomplete because epistemic plan existence is undecidable in the MEP setting (Burigana et al., 2020). “Comprehensive Multi-Agent Epistemic Planning” also emphasizes the difficulty of scalability and the need for flexible update models, observability handling, and support for trust, dishonesty, deception, incomplete knowledge, and misconceptions (Fabiano, 2021).

Taken together, these results situate EPC-AW as a recent planning-time calibration framework within a larger research area concerned with action selection under epistemic uncertainty, information flow, observability, explanation, and auditable reasoning. A plausible implication is that EPC-AW’s main novelty lies not in replacing epistemic planning formalisms, but in reframing calibration as a workflow property: plans are selected for stability under heterogeneous information and refined through persistent memory of past support failures.