Beyond Final Answers: Auditing Trajectory-Level Hallucinations in Multi-Agent Industrial Workflows

Harshada Badave; Santosh Borse; Andrea Gomez; Harshitha Narahari; Sara Carter; Vishwa Bhatt; Aishani Rachakonda; Shuxin Lin; Dhaval Patel

Source-linked AI summary

Beyond Final Answers: Auditing Trajectory-Level Hallucinations in Multi-Agent Industrial Workflows

Harshada Badave, Santosh Borse, Andrea Gomez, Harshitha Narahari, Sara Carter, Vishwa Bhatt, Aishani Rachakonda, Shuxin Lin, Dhaval Patel

arXiv:2605.24219v2May 22, 2026cs.AI

TL;DR

Existing hallucination benchmarks often evaluate only final outputs, missing failures that emerge within multi-step industrial agent trajectories. Trajel introduces a taxonomy-grounded dataset and detection framework over audited traces, finding that trajectory-aware evaluation exposes multi-type and subtle hallucinations that binary checks obscure.

Problem
Static, final-output hallucination benchmarks miss intermediate failures in multi-step agent workflows, where procedural, referential, logical, and scope-related deviations can occur.
Method
Trajel constructs expert-annotated AssetOpsBench trajectories, formalizes five structural hallucination types, and benchmarks subtask-, trajectory-, and long-context detectors plus execution-quality signals.
Results
48.7% of hallucinated trajectories contain multiple types, while automated detectors with high binary performance still miss subtle categories; trajectory-aware evaluation therefore reveals failures hidden by binary checks.
Takeaways & Limitations
Taxonomy-grounded trajectory evaluation provides diagnostic information that final-output or binary evaluation cannot provide for safer agentic deployment.
Takeaways & Limitations
Trajel covers 225 trajectories from one industrial domain and six model configurations sharing AssetOpsBench, so behavior under other domains, orchestrators, and architectures remains untested.

Abstract

from arXiv · show

Large Language Models (LLMs) are increasingly deployed as autonomous agents that reason, use tools, and act over multiple steps. Yet most hallucination benchmarks still evaluate only the final output, missing failures that originate in intermediate Thought-Action-Observation steps. We present Trajel, a dataset and evaluation framework for auditing trajectory-level hallucinations in multi-agent industrial workflows. Trajel introduces a five-type hallucination taxonomy (factual, referential, logical, procedural, and scope-based) over expert-annotated agent traces from AssetOpsBench. We benchmark supervised detection models at the subtask, trajectory, and long-context levels. Our results show that the most common failure modes are missed by existing benchmarks, that nearly half of hallucinated trajectories involve multiple types at once, and that automated detectors with high binary accuracy still misclassify the subtlest types. Trajectory-aware detection significantly outperforms standard post-hoc verification, making taxonomy-grounded evaluation necessary for safer agentic deployment.

1 Introduction

Trajel addresses the mismatch between static, final-output hallucination benchmarks and multi-step industrial agent workflows, where deviations can propagate through tool-mediated trajectories. It introduces audited trajectory data, a five-type taxonomy, and modeling at multiple contextual levels.

Trajectory-level hallucinations can propagate through sequential Thought–Action–Observation cycles, causing cascading operational failures in high-stakes industrial workflows.
Static hallucination evaluations treat tasks as isolated input–output pairs and miss procedural skips, mis-referenced entities, and off-scope actions in agent loops.
Trajel audits where deviations begin in agent traces rather than relying only on post-hoc verification.
The dataset uses LLM-as-a-Judge refinement and blind human review, then benchmarks subtask-level, trajectory-level, and long-context modeling.
48.7% of hallucinated trajectories exhibit multiple taxonomy types simultaneously, supporting a multi-label evaluation of factual, referential, logical, procedural, and scope-based failures.
These datasets and frameworks aim to support safer agentic deployment in high-stakes industries.

2 Related Work

Prior work exposes intermediate agent failures and develops taxonomies or trajectory benchmarks, but does not combine structural hallucination categories with expert-annotated industrial multi-agent traces. Trajel positions itself as an integrated benchmark for this gap.

Large-scale human-labeled agent trajectory datasets remain scarce despite benchmarks covering long-horizon tasks, human-annotated environments, and multi-hop chains.
Existing studies show that hallucinations often arise during intermediate reasoning and tool use rather than only in final outputs.
Prior taxonomies distinguish factual, logical, contextual, instruction, history, and observation inconsistencies but do not separate procedural from scope-based violations.
Related multi-agent resources address general coordination failures or tool-use correctness without targeting hallucination detection in industrial multi-agent workflows.
Trajel combines industrial multi-agent trajectories, full trajectory-level evaluation, a structurally grounded taxonomy, expert human annotations, and LLM-as-a-Judge baselines.

3 Problem Formulation

The paper formalizes an agentic workflow as a structured, role-aware trajectory and defines hallucination as generated content inconsistent with prior evidence, task constraints, or agent mandate. It derives five violation types and corresponding detection tasks.

3 Problem Formulation: The formulation adapts compound AI system notation to multi-agent, tool-augmented AssetOpsBench workflows.
3.1 Trajectory Structure: The workflow comprises LLM-driven agents, an orchestrator, and tools; AssetOpsBench uses four domain agents covering perception, state modeling, forecasting, and execution.
3.1 Trajectory Structure: Each execution step is a Thought–Action–Observation triple, and a trajectory is the ordered sequence of such steps.
3.1 Trajectory Structure: Prior observations form the evidence set, while the task specification supplies constraints, goals, and allowed scope for interpreting the serialized trace.
3.1 Trajectory Structure: Correct workflow structure depends on the task specification because agent ordering is partly flexible, and an incorrect ordering can create downstream hallucinations.
3.2 Hallucination Taxonomy: A hallucination is generated thought or action inconsistent with prior evidence, task constraints, or the acting agent’s operational role.
3.2 Hallucination Taxonomy: The taxonomy contains factual, referential, logical, procedural, and scope violations, distinguishing evidence, inference, workflow, and role failures.
3.3 Detection Tasks: Detection includes per-step, per-category classification and trajectory-level flagging, evaluated against expert annotations alongside LLM-as-a-Judge and trained classifiers.

4 Methodology

The methodology builds and stress-tests labeled AssetOpsBench trajectories, then compares detection models matched to the contextual requirements of different hallucination types. It also evaluates execution-quality signals as potential monitors for real-time intervention.

The evaluation pipeline spans taxonomy prevalence, localization, detection modeling, and predictive execution-quality signals.
The pipeline generates and labels trajectories at both subtask and trajectory granularity before training supervised detectors.
Prompt variation changes instruction specificity, sub-goal ordering, and procedural detail to stress-test failure modes across trajectory variants.
ROC–AUC is the primary metric because correct executions typically outnumber hallucinated trajectories, creating class imbalance.
The taxonomy links factual detection to single-step context, referential and logical detection to trajectory history, and procedural and scope detection to workflow and role context.
BERT uses individual steps for efficient local detection, whereas trajectory-level NLI checks history-based consistency and Longformer models global serialized traces.
These evaluator paradigms are complementary: local classification favors efficiency, NLI tests pairwise consistency, and long-context modeling captures global structure at greater computational cost.
4.3 Signal Analysis: Execution-quality signals include task completion, retrieval accuracy, result verification, sequence correctness, clarity, and justification as potential lightweight runtime monitors.

5 Trajel Dataset

Trajel is built from complete multi-agent execution traces generated across industrial operations questions and model configurations, with annotations supporting trajectory-level analysis.

Dataset composition: 225 complete execution traces cover 42 industrial operations questions across four AssetOpsBench domain agents.The traces serialize Thought–Action–Observation steps for tasks such as sensor retrieval, anomaly detection, failure-mode identification, and work-order generation.
Dataset composition: Six model configurations generate a model × question matrix for controlled comparisons on identical tasks.The matrix enables comparison of hallucination behavior across architectures.
Dataset composition: The dataset overview summarizes Trajel’s trajectory collection and organization.
Dataset composition: Table 3 evaluates LLM-judge performance against human labels on 224 trajectories and analyzes disagreement among 141 jointly flagged trajectories.

6 Experiments

Experiments assess taxonomy-aware detection, model comparisons, localization, supervised classifiers, and execution-quality signals. Results show substantial variation across hallucination types and strong predictive value in runtime quality indicators.

Per-type detection: F1 reaches 0.784 for procedural and 0.719 for factual hallucinations, but only 0.258 for logical and 0.222 for referential types.Procedural detection over-flags, whereas referential and logical detection under-detects; human review remains essential for the latter types.
Taxonomy-aware evaluation: Binary agreement is 78.6% across 224 trajectories, but exact type-set agreement is only 58.2% among 141 jointly flagged trajectories.Binary evaluation would count all 141 jointly flagged trajectories as correctly detected despite systematic type-level failures.
Model comparison: Model 19 reaches the lowest judge precision, 0.703, while 18 of 27 hallucinated trajectories are scope-based.Scope-based failures involve factually correct content originating from the wrong agent and may confuse detectors without role-boundary awareness.
Step-component localization: Hallucinations localize differently: factual types concentrate in Responses at 40.3%, procedural types in Actions at 47.5%, and referential types in Thoughts at 44.0%.These signatures imply that action-validity and response-verification monitors cover different failure patterns.
Supervised detection: All supervised models exceed the majority-class AUC baseline of 0.500, while NLI achieves the highest ROC-AUC at 0.689.Fine-tuned models remain below the zero-shot judge’s F1 of 0.855, and overall AUC below 0.70 indicates that detection remains unsolved.
Predictive signals: CJ alone achieves AUC = 0.908, compared with 0.689 for NLI, while absent CJ and RV correspond to a 97.1% hallucination rate.The analysis uses five execution-quality dimensions on 213 trajectories with complete signal records.

7 Conclusion and Limitations

Trajel formalizes trajectory-aware hallucination detection and finds that procedural and multi-type failures are common, while runtime execution signals outperform supervised classifiers. Its evaluation remains bounded by one industrial domain, a shared orchestrator, limited model configurations, and annotation ambiguity for subtle types.

Trajel formalizes five hallucination types as structural predicates over Thought, Action, and Observation traces, enabling diagnostic analysis beyond binary labels.
Procedural hallucinations dominate at 38.5% of occurrences, while 48.7% of hallucinated trajectories contain multiple types simultaneously.
The clarity-and-justification execution signal reaches AUC = 0.908, exceeding the best supervised classifier at AUC = 0.689.
Trajel covers 225 trajectories from one industrial domain, six model configurations, and a shared AssetOpsBench orchestrator; other settings remain untested.
Overall inter-annotator agreement is moderate at κ = 0.456, but referential and logical types show only slight agreement at κ ≤0.211.

A Sample Trajectory

The sample trajectory illustrates how a silent tool failure can cascade into procedural and factual hallucinations, while the annotation protocol records presence, type, location, rationale, and agreement against human review.

Sample Trajectory: Each AssetOpsBench step records the responsible agent, Thought, Action, and Observation for an anomaly-detection task.
Sample Trajectory: An empty TSFM observation is followed by a fabricated positive anomaly finding, combining procedural and factual hallucination.
Annotation Protocol: The LLM-as-a-Judge returns structured annotations for hallucination presence, type, location, and rationale, using the paper’s five-type taxonomy.
Annotation Protocol: Human reviewers independently assess trajectories and record binary presence, one or more types, the originating component, and a rationale.
Annotation Protocol: Reviewers distinguish explicit fabrication from inefficiency, repetition, self-correction, and partially correct model judgments.
Annotation Quality: The LLM judge reaches 92.2% recall and 79.7% precision, while 32 cases contain disagreement about hallucination type or location.
Annotation Quality: Per-type agreement is moderate-to-substantial for local-evidence types but only slight for referential and logical hallucinations requiring cross-step reasoning.

C.3 Type Distribution

Human evaluation finds procedural hallucinations most prevalent and shows that nearly half of hallucinated trajectories combine multiple types, supporting multi-label trajectory analysis.

Type Distribution: Procedural hallucinations account for 38.5% of all type occurrences, making them the most prevalent category.
Type Distribution: Procedural failures such as skipped diagnostic steps or fabricated workflow completion are invisible to factual-accuracy-only evaluation.
Type Distribution: Factual hallucinations comprise 26.3%, scope-based violations 19.8%, logical hallucinations 8.5%, and referential hallucinations 6.9%.
Type Distribution: 48.7% of hallucinated trajectories exhibit multiple types simultaneously, most commonly procedural with factual or scope-based failures.

C.4 Cross-Model Analysis

Hallucination rates vary substantially across model configurations on identical tasks, and their type profiles differ qualitatively, demonstrating the diagnostic value of taxonomy-based analysis.

Cross-Model Analysis: Hallucination rates range from 52.4% for Model 6 to 81.0% for Model 17 on the same 42-question task suite.
Cross-Model Analysis: Model 19 has 18 of 27 hallucinated trajectories classified as scope-based, whereas Model 6 has only 2 of 22.
Cross-Model Analysis: Model 17 has the highest referential hallucination count, with 7 cases.
Cross-Model Analysis: These qualitatively different type profiles would be hidden by a binary hallucination label.

C.5 Localization Analysis

Hallucinations cluster in externally visible actions and responses, with risk peaking mid-trajectory before declining as tasks narrow. The taxonomy and scope emphasize why trajectory-level analysis matters for diagnosing control-flow, coordination, and downstream cascade risks.

Step-component localization: 70 hallucinations localized to Actions and 58 to Responses, compared with 40 in Thoughts and 11 in Observations.The distribution suggests failures concentrate in tool invocations and output claims rather than internal reasoning alone.
Task-position localization: 63 hallucinations occurred at Task 3, after which occurrences declined monotonically through Tasks 6 to 8.This mid-trajectory concentration is consistent with accumulated context increasing risk while the remaining task narrows.
Taxonomy implications: The five hallucination types are modeled as multi-label vectors because one step may exhibit several types simultaneously.The taxonomy distinguishes grounding, control-flow, and coordination problems rather than collapsing them into a binary label.
Trajectory-level scope: Trajectory analysis captures failure cascades in which an earlier hallucination corrupts state consumed by later steps.This downstream compounding risk is outside the scope of single-step evaluation.
Scope and setting: The analysis covers benign reasoning, grounding, and control-flow limitations in deterministic-tool environments with faithful observations.The industrial scenarios are simulated rather than tied to specific real-world facilities or individuals.