PaperlayerView exact revision on arXiv

Source-linked AI summary

When LLMs Sit Above Diagnostic Tools: Unrealized Complementarity in Industrial Fault Diagnosis

Donghwan Kim

arXiv:2610.05031v2cs.CLcs.SYeess.SY

TL;DR

This paper asks whether an LLM can integrate imperfect industrial diagnostic sources well enough to match the stronger standalone source. Across five datasets, it evaluates paired evidence-only and advised calls; integration did not consistently outperform the stronger source, despite specialist information improving the LLM in held-out TEP.

  • Problem

    Combining informative sources does not guarantee better decisions; the central question is whether an LLM can use complementary, imperfect predictions to match or exceed the stronger source.

    Source: §1 · p. 1

  • Method

    Across industrial diagnostic datasets, paired repeat calls separate advice effects from output instability while the LLM receives specialist labels and is compared with standalone sources.

    Source: §3.1 · p. 5, §3.4 · p. 7

  • Results

    Across four direct-comparison datasets, implicit integration showed no consistent advantage; in held-out TEP, specialist information improved the unaided LLM by 12.8 points but integration remained 5.9 points below the specialist.

    Source: §1 · p. 2, §8 · p. 18

  • Takeaways & Limitations

    Evaluate integration quality against the stronger standalone component, not only against the unaided LLM.

    Source: §1 · p. 2, §8 · p. 18

    The bearing and process results show poor case-level source selection but do not explain why the direction of reliance differs.

    Source: §6 · p. 17

Abstract

from arXiv · show

Large language models are increasingly used as integration layers above specialized tools, but a stronger component does not necessarily produce a stronger combined system. Across five diagnostic datasets (bearing vibration, process monitoring, semiconductor equipment), we study whether an LLM can reliably use external diagnostic information; paired repeat calls separate advice effects from output instability. In all five, conflicting external information overturned initially correct LLM judgments. Among the four datasets with direct integration comparisons, none showed a consistent advantage for implicit LLM integration over the stronger standalone source. On a Tennessee Eastman confirmation set whose protocol was fixed before evaluation, unaided accuracy was 64.67%, implicit LLM-specialist integration 77.43%, and the specialist alone 83.33%. Specialist information improved the LLM by 12.8 points (95% interval 9.7 to 15.9), yet the integrated output stayed 5.9 points below the specialist (95% interval -12.0 to -0.7). A two-source selector oracle reached 92.76%, indicating complementarity that the integrated output did not fully realize. The integrated output missed 140 of 295 specialist corrections (47.5%) but lost 15 of 99 initially correct LLM judgments (15.2%). The deficit remained under prompt and specialist sensitivity analyses. Among CWRU cases solved under both evidence presentations, task-aligned physical evidence yielded lower estimates of susceptibility to incorrect advice in six of seven models (five intervals excluding zero); higher reasoning effort gave no reliable reduction in five models, and a separate four-model TEP analysis gave no clear evidence that it resolves the integration problem. Source quality and integration quality should be evaluated separately: an integration layer should be compared with its stronger standalone component, not only with the unaided LLM.

1. Introduction

The paper asks whether an LLM can integrate two imperfect diagnostic sources well enough to match or exceed the stronger standalone source. Across five datasets, implicit integration did not consistently outperform that source, and the held-out Tennessee Eastman evaluation showed substantial unrealized complementarity.

1. Introduction

  • The study separates source quality from integration quality: a strong specialist and competent LLM can still yield a worse combined label if corrections are ignored or errors copied.Complementary sources require case-level arbitration to realize their potential.Source: §1 · pp. 1–2
  • Paired repeat calls distinguish advice effects from ordinary output instability when testing whether integration beats the better standalone source.The protocol repeats the evidence-only prompt before adding an external diagnosis.Source: §1 · p. 2
  • Across four datasets with direct comparisons, implicit LLM integration showed no consistent advantage over the stronger standalone source.Conflicting external information overturned some initially correct LLM judgments in all five datasets.Source: §1 · p. 2
  • 12.8 points of improvement over the unaided LLM still left the integrated output 5.9 points below the specialist in held-out TEP.A two-source selector oracle reached 92.76%, indicating substantial unrealized case-level complementarity.Source: §1 · p. 2
  • The deficit persisted with explicit arbitration, the stored unaided diagnosis, and an alternative specialist; the study also tests physical evidence and reasoning effort as interventions.These interventions assess whether improved diagnostic properties translate into better integration.Source: §1 · p. 2

2. Related work

Prior work studies advice following, tool use, human–AI teaming, and LLMs coordinating industrial tools or evidence. This paper instead evaluates whether an LLM integrator can benefit from two fallible predictions while matching or exceeding the stronger source.

2. Related work

  • Advice and source-conflict research examines compliance or source preference, while this study tests whether each source’s correct cases can improve the combined label.Prior findings include source-authority effects (Kumar and Chopra, 2026) and preferences among conflicting sources (Schuster, Gautam, and Markert, 2026; Li et al., 2025).Source: §2 · p. 3
  • Tool-use studies distinguish adopting bad returns from ignoring needed ones; here the specialist can err, so both failure directions matter.Xu et al. (2026b) define “Tool Ignored” as overriding a correct code-execution result with incorrect model reasoning.Source: §2 · p. 3
  • Human–AI teams and LLM ensembles can fall short of their better member, while learned deferral methods explicitly optimize case routing rather than using an off-the-shelf LLM integrator.The paper’s setting instead studies direct LLM integration without a separately trained routing objective.Source: §2 · pp. 3–4
  • Industrial LLM systems coordinate diagnostic tools or heterogeneous evidence, but often evaluate diagnosis, reporting, causal chains, or deployment rather than retention of the stronger source’s accuracy.Examples include FaultExplainer (Khan et al., 2025), Liang and Sin (2026), KDACL, and AgentRCA (Wei and Fink, 2026).Source: §2 · p. 4
  • Industrial benchmarks assess LLM capability, whereas this protocol fixes two imperfect predictions and measures integration relative to the stronger standalone source.AssetOpsBench (Patel et al., 2026), FactoryBench (Merzouki et al., 2026), and FailureSensorIQ (Constantinides et al., 2025) are not treated as drop-in substitutes.Source: §2 · p. 4

3. Experimental framework

The framework tests whether an LLM can arbitrate between imperfect diagnostic sources at the case level, separating source quality from the quality of the integration procedure. It spans five datasets with distinct roles and uses paired repeats, predefined comparisons, and dataset-matched bootstrap intervals.

3.1 Experimental objective

  • The two-source selector oracle is correct whenever either stored standalone prediction is correct, representing a maximum rather than an implemented integration method.The experiment evaluates whether the LLM can approach that case-level arbitration while comparing its final label with the better standalone source.Source: §3.1 · pp. 4–5

3.2 Datasets and experimental roles

  • Five datasets serve distinct roles: bearing data characterize integration and evidence presentation, TEP is the held-out non-bearing evaluation, and Lam 9600 is exploratory.These datasets are not matched replications.Source: Table 1 · p. 6, §3.2 · p. 6

3.3 Models

  • The fixed panel comprises seven models, with separate five-model CWRU reasoning-effort and four-model TEP sensitivity analyses.The four-model TEP panel was selected before confirmation outcomes at high reasoning effort were examined.Source: §3.3 · p. 6

3.4 Two kinds of external diagnosis

  • Controlled-advice experiments assign whether advice is correct, whereas integration experiments use actual predictions from fixed specialist models.Lam 9600 tests advice vulnerability, not source complementarity or performance against the stronger standalone source.Source: §3.4 · p. 7

3.5 Prompt and specialist sensitivity analyses

  • The post-primary TEP sensitivity analyses vary the prompt or specialist, and the prior-label condition was specified after earlier sensitivity results were known.The arbitration prompt was treated as a sensitivity condition, not an optimized or calibrated fusion procedure.Source: §3.5 · p. 7

3.6 Evidence representations

  • CWRU’s task-aligned evidence adds envelope spectra, theoretical characteristic frequencies, and explicit fault mappings absent from the generic summary.The two evidence views therefore carry different information; Paderborn and XJTU use the same generic-versus-task-aligned design principle.Source: §3.6 · p. 8

3.7 Inclusion criteria for bearing comparisons

  • Twenty-two bearing comparisons met predefined inclusion rules; CWRU was excluded from that count because its specialists made too few errors.The rules required above-chance unaided accuracy, sufficient oracle advantage, and at least 15 specialist errors; the count is conditional on those criteria.Source: §3.7 · p. 9

3.8 Evaluation metrics; 3.9 Bootstrap confidence intervals

  • The better standalone source is selected per model, and adjusted harmful switching subtracts ordinary repeat-call errors from advised-call errors among initially correct cases.Intervals use 10,000 percentile-bootstrap resamples without multiple-comparison adjustment, with resampling units matched to each dataset.Source: §3.8 · p. 9, §3.9 · p. 10

4. Results

In the held-out Tennessee Eastman evaluation, implicit integration improved on the unaided LLM but remained below the specialist, with missed corrections dominating integration errors. Bearing and intervention analyses likewise found no consistent stronger-source gains, while task-aligned evidence reduced harmful switching in most CWRU models.

4.1 Implicit LLM integration remains below the specialist

  • 77.43% integrated accuracy remained 5.9 points below the 83.33% specialist, despite a 12.8-point gain over the unaided 64.67% model.The two-source selector oracle reached 92.76%.Source: Figure 2 · p. 10, §4.1 · p. 10

4.2 Most integration errors are missed corrections

  • 140 of 295 specialist corrections (47.5%) were missed, compared with 15 of 99 initially correct LLM judgments (15.2%) lost.The two sources disagreed in 394 of 1,050 model–case rows.Source: §4.2 · p. 11, Figure 3 · p. 11

4.3 Prompt and specialist sensitivity analyses of the TEP result

  • Prompt arbitration and an alternative specialist did not remove the integration deficit, while explicit arbitration traded fewer harmful acceptances for more missed corrections.The alternative random-forest specialist retained a 5.7-point deficit versus the better standalone source.Source: §4.3 · p. 13, Table 3 · p. 13

4.4 Bearing studies rarely show reliable gains over the stronger source

  • Across 22 eligible bearing comparisons, three intervals were below zero, eighteen included zero, and one was above zero; all 28 CWRU point estimates remained below the stronger source.Intervals were not adjusted for multiple comparisons.Source: §4.4 · p. 13

4.5 Task-aligned evidence reduces susceptibility to incorrect advice

  • Task-aligned evidence yielded lower harmful-switching estimates for six of seven CWRU models, with five model-specific intervals excluding zero.The analysis was restricted to cases each model correctly diagnosed under both evidence presentations.Source: §4.5 · pp. 13–14, Figure 5 · p. 14

4.6 Higher reasoning effort does not clearly resolve the integration deficit

  • Higher reasoning effort showed no reliable CWRU reduction across five models, and the separate four-model TEP analysis gave no clear evidence that it resolved the integration deficit.In TEP, unaided accuracy rose 5.2 points, while integrated accuracy rose 2.2 points with an interval including zero.Source: §4.6 · pp. 14–15

4.7 Exploratory evaluation on real-equipment data

  • In exploratory Lam 9600 data, incorrect external advice overturned six of 23 initially correct judgments across four wafers, while paired repeats did not flip.The analysis was a vulnerability check, not a benchmark of diagnostic capability.Source: §4.7 · p. 15

5. Discussion

Across the evaluated settings, source quality did not guarantee integration quality: implicit integration lacked a consistent advantage over the stronger standalone source. Missed corrections dominated the TEP error pattern, and intervention effects on susceptibility did not establish a general integration remedy.

5.1 Source quality is not integration quality

  • Across four datasets with direct comparisons, implicit integration showed no consistent advantage over the stronger standalone source, and the TEP oracle gap reflects a setting selected for sufficient disagreement.The paper recommends validating an integration layer against the validated specialist itself.Source: §5.1 · p. 15

5.2 Missed corrections dominate harmful acceptance

  • Missed corrections dominated harmful acceptance, so suppressing tool-following alone can leave the main source-selection failure unaddressed.Explicit arbitration reduced harmful acceptance but increased missed corrections; the needed choice is which source to trust on each case.Source: §5.2 · p. 16

5.3 Interventions affecting susceptibility to incorrect advice

  • Task-aligned evidence reduced susceptibility for most CWRU models, but the same benefit did not appear for reorganized TEP measurements, and higher reasoning effort showed no reliable CWRU reduction.Targeted metric improvement need not mean uniformly better diagnostic behavior; Lam 9600 also lacked an integration-versus-source comparison.Source: §5.3 · p. 16

6. Limitations

The study’s conclusions are bounded by controlled, concise interfaces, a narrow TEP evaluation, and a limited set of specialist families and models. Several design choices also leave specific comparisons unresolved.

6. Limitations

  • The external diagnoses were fixed short statements, not live engineer notes or full tool logs, and fitted-specialist labels were framed as priors rather than physical evidence.Alternative presentations could shift the balance between missed corrections and harmful acceptance.Source: §6 · p. 17
  • The primary TEP test covers implicit integration; its fixed prompt order cannot separate prior-diagnosis anchoring from position or recency effects.The higher-reasoning analysis used four of seven primary models and tested provider-side effort with concise output, not explicitly elicited deliberation.Source: §6 · p. 17
  • The primary specialist was multinomial logistic regression, and the alternative random forest also showed an integration deficit, but both are classical diagnostic models.Other specialist families may have different error profiles and integration behavior.Source: §6 · p. 17
  • The TEP confirmation used a simulated process and a predefined six-class subset, so results need not extend to historical plant data or other fault modes.Specialist errors and copying of those errors were concentrated in a few classes and models.Source: §6 · p. 17
  • The Lam 9600 analysis used few wafers and three models, making it exploratory rather than a benchmark evaluation.Some rendered XJTU prompts were not retained, limiting exact prompt-level reproduction of that auxiliary analysis.Source: §6 · p. 17
  • The design does not establish whether baseline diagnostic accuracy causes advice reliance, nor why the direction of poor source selection differs between bearing and process results.Source: §6 · p. 17

7. Broader impacts

Industrial diagnostic decisions can affect equipment availability, maintenance, and safety, so the evaluated LLM layers should not be deployed autonomously. The results argue for validating each layer against the strongest available source.

7. Broader impacts

  • Industrial diagnostic decisions can affect equipment availability, maintenance, and safety; the evaluated LLM layers should not be deployed autonomously.The authors argue for validating such a layer against the strongest available source.Source: §7 · p. 17

8. Conclusion

Across four diagnostic datasets, implicit LLM integration did not consistently outperform the stronger standalone source. The held-out TEP evaluation shows that specialist information improved the LLM, but the integrated output remained below the specialist, with missed corrections dominating harmful acceptance.

8. Conclusion

  • Across four diagnostic datasets, prompt-level implicit LLM integration did not consistently outperform the stronger standalone source.In held-out TEP, specialist information improved the unaided LLM, yet integration remained below the specialist and missed corrections dominated harmful acceptance.Source: §8 · p. 18
  • Task-aligned physical evidence reduced susceptibility to incorrect advice in the CWRU setting, while higher reasoning effort did not clearly resolve the TEP integration deficit.Source: §8 · p. 18
  • The authors recommend evaluating an integration layer against its stronger standalone component, not only against the unaided LLM.This follows from the distinction between source quality and integration quality.Source: §8 · p. 18

Appendix A. Reproducibility details

The appendix records the fixed confirmation protocol and development-only specialist selection, alongside dataset, prompt, and decoding details intended to support reproducibility.

Appendix A. Reproducibility details

  • The TEP confirmation used fixed stored outputs and dataset-appropriate cluster bootstrap procedures, with a fixed seed and pinned provider configuration.Provider fallback was disabled; model identifiers and prompt-family decoding settings are reported.Source: Appendix A · pp. 20–21

Appendix Table A1. Development-only TEP specialist selection.

  • Development-only eligibility and ranking criteria selected multinomial logistic regression from two specialist candidates without using confirmation outcomes.The candidates were a 400-tree random forest and standardized multinomial logistic regression; both passed eligibility, and logistic regression had more pooled LLM-only-correct rows.Source: Appendix A · p. 21, Appendix Table A1. Development-only… · p. 21

Appendix Table A2. Descriptive model-level reliance on the TEP confirmation set.

Appendix B. Dataset details

  • The datasets include 72 CWRU cases, 87 Paderborn segments, 88 XJTU-SY recordings, and 150 TEP simulation runs across six specified classes.The Lam analysis uses publicly available Eigenvector Lam 9600 machine-state data.Source: Appendix B · pp. 21–22

Appendix Table B1. Fields displayed in the generic and task-aligned bearing presentations.

  • The bearing-presentation comparison contrasts generic summaries with task-aligned geometry, characteristic frequencies, envelope-analysis measures, and deterministic frequency contrasts.The listed generic and aligned fields vary by bearing dataset.Source: Appendix Table B1. Fields displayed in… · p. 22
  • Controlled-advice analyses assigned correct or incorrect labels by design, whereas integration analyses used naturally occurring fitted-specialist errors; the settings were not pooled.Source: Appendix Table B1. Fields displayed in… · p. 22

C.1 Shared instructions and output format

  • Prompt families specify JSON-only outputs and distinguish controlled incorrect-label wording from fitted-specialist labels framed as case-specific priors.The shared output instruction prohibits written chain-of-thought; provider-side reasoning effort was controlled separately.Source: §C.1 · p. 22, §C.2 · pp. 22–23

C.3 Tennessee Eastman

  • The TEP and CWRU prompt examples provide case evidence and static references, while the appendix also reports Paderborn, XJTU-SY, Lam, and decoding-setting details.The TEP reference is described as identical across cases and not evidence about the current run.Source: §C.3 · pp. 23–24, §C.4 · p. 25, §C.5 · p. 26

Appendix D. CWRU common-correct analysis

The common-correct analysis compares susceptibility under generic and task-aligned evidence; the reported row shows lower susceptibility with task-aligned evidence.

Appendix Table D1. CWRU common-correct susceptibility analysis.

  • 0.341 generic susceptibility and 0.095 task-aligned susceptibility yielded a difference of 0.246 (95% CI [0.105, 0.405]).The analysis retains cases diagnosed correctly before advice under both presentations; the table reports the difference as low minus high susceptibility.Source: Appendix D · p. 26, Appendix Table D1. CWRU common-correct… · p. 27
  • The final row is an unweighted summary across the fixed evaluated model panel, with an interval jointly resampling all 28 recording files.It supplements rather than replaces model-specific results.Source: Appendix Table D1. CWRU common-correct… · p. 27

Appendix E. Evidence-strength analysis

Neither dataset’s evidence-strength score reliably predicted susceptibility to incorrect advice, and the score was not shown to the LLM.

Appendix E. Evidence-strength analysis

  • CWRU OR 1.478 (95% CI [0.789, 2.597]) and TEP OR 0.923 (95% CI [0.746, 1.142]) were not statistically reliable predictors.Both intervals included 1; the score was separately computed and never shown to the LLM.Source: Appendix E · p. 27

F.1 CWRU reasoning-effort analysis

Across these secondary analyses, higher reasoning effort did not reliably reduce the integration deficit, while bearing and presentation analyses show mixed, setting-dependent patterns.

F.1 CWRU reasoning-effort analysis

  • Every CWRU reasoning-effort susceptibility interval included zero, so the analysis found no reliable reduction at higher effort.Conditional Hadj is computed on initially correct unaided rows.Source: §F.1 · p. 27, Appendix Table F1. CWRU… · p. 28

Appendix Table F2. Four-model TEP accuracies at low and high reasoning.

  • 5.2 points of unaided accuracy gain (95% interval 2.2 to 8.2) accompanied 2.2 points of integrated accuracy gain (−0.2 to 4.8), a −3.0-point difference (−6.9 to 0.8).Integrated output minus the better standalone source was −6.3 points (−13.5 to 1.0) at low reasoning and −4.2 points (−10.5 to 0.8) at high reasoning.Source: Appendix Table F2. Four-model TEP… · p. 28
  • At high reasoning, integration missed 69 of 134 specialist corrections (51.5%); on low-setting rows it missed 84 of 165 (50.9%).At high setting it recorded 11 harmful acceptances in 58 LLM-only-correct rows (19.0%); on low-setting rows, 6 of 57 (10.5%).Source: Appendix Table F2. Four-model TEP… · pp. 28–29

Appendix Table G2. Included bearing integration comparisons.

  • Across 28 bearing-comparison point estimates, all were below the better standalone source; three intervals were below zero, eighteen included zero, and one was above.CWRU was excluded from the 22-comparison inferential count because its specialists made fewer than 15 errors.Source: Appendix Table G2. Included bearing… · p. 29

Appendix H. Paderborn and XJTU supporting analyses

  • On XJTU-SY, task-aligned evidence lowered harmful switching for all seven models while balanced diagnostic accuracy increased; the two changes could not be separated.Paderborn and XJTU used related but nonidentical interfaces; Paderborn’s task-aligned presentation also raised accuracy and reduced harmful switching.Source: Appendix H · p. 30

Appendix I. TEP development presentation contrast; Appendix J. Real-equipment advice-vulnerability check

  • On the TEP development set, retest-adjusted harmful switching was 0.107 under generic presentation and 0.124 under subsystem presentation, a −0.017 difference (95% CI [−0.036, 0.003]).The separate Lam five-class check found six of 60 unaided-correct judgments changed to the incorrect external label while paired repeats did not flip; this result was descriptive.Source: Appendix I · p. 30, Appendix J · p. 30

Appendix K. TEP secondary sensitivity analyses

Appendix M. Post-hoc TEP NORMAL-class source comparison

  • On true-NORMAL TEP runs, integrated output differed from the unaided LLM by +5.1 percentage points (95% interval −1.7 to 11.4), which did not establish increased errors.The integrated output differed from the specialist by −46.3 points (−63.4 to −29.7); these conditional errors are not operational alarm frequencies.Source: Appendix Table M1. NORMAL-to-fault… · p. 33, Appendix M · p. 33

Appendix N. Post-hoc normal-versus-fault error profiles

The post-hoc bearing analysis finds dataset-dependent shifts in error profiles, not a uniform task-aligned benefit or cost; analysis artifacts were not yet public.

Appendix N. Post-hoc normal-versus-fault error profiles

  • The three bearing datasets’ generic and task-aligned presentations contained different information, with task-aligned evidence adding fault-relevant physical structure.The analysis characterizes the evaluated interfaces rather than isolating a formatting effect; its conditional errors are not operational plant alarm frequencies.Source: Appendix N · p. 33
  • The post-hoc results do not support a general claim that task-aligned evidence reduces normal/healthy-to-fault errors.XJTU-SY showed a large increase alongside an almost complete disappearance of fault-to-NORMAL misses; the authors interpret the patterns as dataset-dependent operating-profile shifts.Source: Appendix N · p. 34

Appendix O. Computational and API budget

  • Approximate API expenditure was $426 in total, including retries associated with transient API or provider errors.The reported total includes bearing, Tennessee Eastman, and Lam 9600 studies.Source: Appendix O · p. 34

Appendix P. Data and code availability

  • Analysis code and reproduction artifacts were not yet publicly available.The public datasets remain available from their original maintainers; the Lam source files are not redistributed.Source: Appendix P · p. 34
Loading 2610.05031v2…