💻 computer science

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

This paper introduces Trajel, a dataset and evaluation framework that audits trajectory-level hallucinations in multi-agent industrial workflows by categorizing five distinct failure types, demonstrating that existing benchmarks miss critical intermediate errors and that taxonomy-grounded, trajectory-aware detection is essential for safer agentic deployment.

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

Published 2026-05-26

📖 4 min read☕ Coffee break read

CC BY 4.0

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

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

Imagine you hire a team of expert robots to fix a complex machine in a factory. You don't just ask them for the final repair report; you watch their entire process: what they think, what tools they pick up, what they see, and how they talk to each other.

This paper, called Trajel, is about catching these robots when they start "dreaming" (hallucinating) in the middle of their work, not just when they get the final answer wrong.

Here is the breakdown of the paper using simple analogies:

1. The Problem: The "Final Answer" Trap

Most current tests for AI are like grading a student's essay only by reading the last sentence. If the conclusion looks good, the student gets an A. But what if the student made up facts in paragraph two, skipped a whole step in paragraph three, or talked to the wrong teacher in paragraph four?

In the real world, industrial robots (agents) don't just write essays; they perform multi-step tasks like checking sensors or ordering parts. If a robot skips a safety check or talks to the wrong system, it might still give a "correct-looking" final answer, but the machine could still break. The paper argues that we need to audit the entire journey (the trajectory), not just the destination.

2. The New Tool: A "Hallucination Taxonomy"

The authors created a new way to categorize how robots get confused. They call this a Taxonomy (a fancy word for a classification system). They found five specific ways robots "dream" while working:

Factual (The Liar): The robot says, "The temperature is 100 degrees," but the sensor actually said 20. It's making up data.
Referential (The Ghost): The robot says, "I just checked the Blue valve," but it never checked a Blue valve. It's remembering things that never happened.
Logical (The Confused): The robot says, "The machine is hot, therefore I should turn on the heater." The logic is broken, even if the facts are right.
Procedural (The Skipper): The robot says, "I finished the job," but it actually skipped the step where it had to verify the oil pressure. It's pretending to do work it didn't do.
Scope (The Imposter): The robot says, "I fixed the engine," but it's actually a robot designed only to check the tires. It's doing a job it wasn't hired for.

Key Finding: The paper found that nearly half of the time robots mess up, they do multiple of these things at once. It's not just one mistake; it's a chain reaction.

3. The Dataset: "Trajel"

To study this, the researchers built a massive library called Trajel.

What's inside? 225 detailed recordings of robots trying to fix industrial machines (like cooling systems).
Who checked it? They didn't just use another AI to check the work. They had human experts (like senior engineers) read every single step of the robot's thought process to see where it went wrong.
The Result: They found that robots fail much more often than we thought (between 52% and 81% of the time, depending on the robot model), and many of these failures are invisible if you only look at the final result.

4. The Detective Work: Can AI Catch AI?

The researchers tried to build computer programs (detectors) to automatically spot these mistakes.

The Good News: Simple checks work well for obvious lies (Factual) or skipped steps (Procedural).
The Bad News: The detectors are terrible at catching the "Ghost" (Referential) and "Confused" (Logical) mistakes. Even advanced AI struggles to tell if a robot is remembering something that didn't happen or if its logic is twisted.
The Surprise Winner: The best "detector" wasn't a complex AI model. It was a simple clarity check. If the robot's reasoning sounded clear and justified, it was usually right. If the robot sounded vague or couldn't explain why it did something, it was almost certainly hallucinating (94% of the time).

5. Why This Matters

The paper concludes that we can't just trust the final answer from an industrial robot. We need to build "safety nets" that watch the robot's entire thought process in real-time.

If a robot starts skipping steps (Procedural) or talking about tools it doesn't have (Referential), the system needs to stop it immediately, not wait until the machine breaks. The paper provides the first set of tools and rules to help us build these safety nets for the future of industrial automation.

In short: We used to only check if the robot finished the job. Now, thanks to Trajel, we know we need to watch how it thinks, what it remembers, and who it talks to, because that's where the real danger lies.

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

1. Problem Statement

The transition from static Large Language Models (LLMs) to autonomous agentic systems introduces a critical failure mode: trajectory-level hallucination. While existing benchmarks (e.g., TruthfulQA, AgentBench) evaluate hallucinations primarily as "one-shot" factual errors in final outputs, they fail to capture deviations that occur within the sequential, tool-mediated Thought-Action-Observation (TAO) cycles of multi-agent workflows.

In industrial settings like data center monitoring (specifically within the AssetOpsBench framework), a hallucination is not merely a wrong answer; it is a structural deviation from evidence that propagates through an agent loop. These deviations can manifest as procedural skips, mis-referenced entities from previous steps, or scope violations where an agent acts outside its mandate. Current evaluation paradigms, which treat instances as isolated input-output pairs, cannot diagnose whether a system failed due to environmental misunderstanding or the invention of non-existent states. Furthermore, the definition of hallucination in interactive loops is often conflated with logic errors or tool-execution failures, lacking the granularity required for diagnostic analysis.

2. Methodology

The authors introduce Trajel, a dataset and evaluation framework designed to audit trajectory-level hallucinations in multi-agent industrial workflows. The methodology is structured around four core components:

A. Trajectory Structure and Formalism

The workflow is modeled as a compound AI system $\Phi = (M, C, T_{tool})$ , where $M$ represents LLM-driven agents, $C$ is the orchestrator, and $T_{tool}$ is the set of tools. Execution proceeds as a sequence of steps $s_t = (a_t, \tau_t, \alpha_t, \omega_t)$ , where $\tau_t$ is the thought, $\alpha_t$ is the action, and $\omega_t$ is the observation. A trajectory $T$ is the ordered trace of these steps.

B. Trajectory-Aware Hallucination Taxonomy

The paper defines hallucination as a structural predicate over the TAO trace, categorized into five distinct types:

Factual ( $h_F$ ): Claims contradicted by ground-truth data at the current step.
Referential ( $h_R$ ): References to entities or prior results absent from the trajectory history (the model "remembers" something that never happened).
Logical ( $h_L$ ): Reasoning that does not follow from its premises, even if the premises are correct.
Procedural ( $h_P$ ): Skipping, reordering, or fabricating steps required by the task specification, or claiming completion of absent steps.
Scope ( $h_S$ ): An agent acting or claiming outside its specific operational mandate (unique to multi-agent settings).

C. Dataset Construction (Trajel)

The dataset comprises 225 expert-annotated agent trajectories generated across 6 model configurations and 42 industrial AssetOps tasks.

Annotation Protocol: A hybrid approach mitigates bias. Trajectories are first evaluated by an LLM-as-a-Judge, followed by blind human review from two institutions (IBM and Columbia University) without access to the LLM's judgments.
Labels: Annotations include hallucination presence, type (multi-label), localization within the trace (Thought, Action, Observation, Response), and free-text rationale.
Statistics: The human-identified hallucination rate is 68.3%, with a Cohen's $\kappa$ of 0.456 between automated and human judgments. Notably, 48.7% of hallucinated trajectories exhibit multiple types simultaneously.

D. Detection Modeling Framework

The authors benchmark three supervised detection paradigms to address different context requirements:

Subtask-level Classification (BERT): Operates on individual steps ( $\tau_t, \alpha_t, \omega_t$ ) to detect local cues (effective for Factual hallucinations).
Trajectory-level NLI: Treats detection as an entailment problem where trajectory history is the premise and the current step is the hypothesis (targets Referential and Logical hallucinations).
Long-context Modeling (Longformer): Ingests the full serialized trajectory to detect global structural violations like Procedural and Scope-based hallucinations.

E. Signal Analysis

The study investigates execution-quality signals observable during or immediately after execution (Task Completion, Data Retrieval Accuracy, Result Verification, Agent Sequence Correctness, and Clarity & Justification) to determine their predictive power for hallucination.

3. Key Results

A. Prevalence and Taxonomy

Procedural hallucinations are the most common failure mode, accounting for 38.5% of identified failures. These are invisible to output-only evaluation.
Multi-label complexity: Nearly half (48.7%) of hallucinated trajectories involve multiple types simultaneously, validating the need for multi-label formulations over single-label classifications.
Model Variance: Hallucination rates vary drastically across models, ranging from 52.4% to 81.0%, indicating significant discrepancies in how different architectures maintain grounding over long horizons.

B. Detection Performance

LLM-as-a-Judge: While achieving high binary accuracy (F1 = 0.855), the judge systematically misclassifies subtle types. Agreement with humans drops significantly for Logical ( $\kappa = 0.211$ ) and Referential ( $\kappa = 0.176$ ) hallucinations.
Supervised Models: Fine-tuned models (BERT, NLI, Longformer) improve over majority-class baselines but fail to match the zero-shot LLM judge on F1 scores. The best supervised model (NLI) achieves an ROC-AUC of 0.689, confirming that trajectory-level hallucination detection remains an unsolved challenge for current classifiers.
Localization: Hallucinations are most frequently localized to Actions (47.5% of procedural) and Responses (40.3% of factual), suggesting failures often occur in the translation of reasoning to tool invocation or output claims.

C. Predictive Signals

The study identifies Clarity and Justification (CJ) as the strongest univariate predictor of hallucination.

Trajectories lacking clear, grounded reasoning hallucinate at a rate of 94.3%, compared to 9.1% for high-clarity trajectories.
The CJ signal alone achieves an AUC of 0.908, substantially outperforming the best trained classifier (NLI, AUC = 0.689).
When both Clarity and Result Verification are absent, the hallucination rate reaches 97.1%, suggesting these signals can serve as effective "kill-switch" conditions for agent orchestrators.

4. Significance and Claims

The paper claims that trajectory-aware detection significantly outperforms standard post-hoc verification, making taxonomy-grounded evaluation necessary for safer agentic deployment.

Shift in Evaluation: The work argues that moving beyond binary "final answer" verification to structural analysis of the agent trace is essential. The current paradigm of static benchmarks fails to capture the temporal and interactive dynamics of agent loops.
Diagnostic Value: The five-type taxonomy disentangles grounding failures from reasoning errors and control-flow violations, providing the granularity needed to diagnose where in the trajectory a deviation began.
Practical Pathway: The findings suggest that lightweight runtime monitors based on execution-quality signals (specifically Clarity & Justification) offer a more practical and effective path to real-time hallucination detection than training complex supervised classifiers on limited trajectory data.
Limitations: The authors acknowledge that the dataset is limited to a single industrial domain (AssetOpsBench) and six model configurations. Inter-annotator agreement remains moderate for the most subtle types (Logical and Referential), indicating a need for richer annotation protocols. The gap between supervised classifier performance and LLM-judge performance motivates future work on hybrid architectures.

In conclusion, Trajel provides the first systematic framework for auditing trajectory-level hallucinations in multi-agent industrial workflows, demonstrating that procedural and scope-based failures are prevalent, often co-occur, and require evaluation methods that respect the structural integrity of the agent's execution trace.