Where Root Cause Analysis Fails: A Retrieval-Reranking Decomposition

Hada Melino Muhammad; Luan Pham; Laure Barrière; Sachin Shetty; Leonardo Pulga; Flora D. Salim

cs.LGSep 29, 2026

Where Root Cause Analysis Fails: A Retrieval-Reranking Decomposition

Authors: Hada Melino Muhammad, Luan Pham, Laure Barrière, Sachin Shetty, Leonardo Pulga, Flora D. Salim

Organizations: University of New South Wales · Baker Hughes

Abstract

Identifying the root cause of an anomaly among hundreds of sensors is critical for preventing safety incidents and costly downtime in complex monitored systems. Existing studies evaluate root cause analysis (RCA) methods using top@k accuracy. We show that this metric has a fundamental blind spot: it conflates two failure modes, retrieval failure, where the true cause is never considered, and reranking failure, where it is considered but ranked too low. In this work, we introduce a retrieval-reranking decomposition and audit four well-known benchmarks to expose this blind spot. Our experiments show that, on benchmarks with complex faults, statistical baselines mis-rank the true cause 79-100% of the time, and graph-based methods never clearly beat the best statistical baseline, whether their causal graphs are learned on short fault windows, on retrieved candidate pools guaranteed to contain the cause, or on multi-day normal-operation data. Meanwhile, on simple benchmarks where faults manifest significantly at their origin, retrieval is nearly solved (98-100%). Guided by the decomposition, we build a two-stage pipeline combining a multi-signal retriever with an LLM reranker that, as one fixed configuration, matches or exceeds the best baseline's top@1 accuracy on all six benchmark suites (by up to +12 points), with no causal graph or labeled data required. When all methods rank the same retrieved candidates with the true cause guaranteed present, adding a short system-description document lets the reranker lead the best baseline by +7 to +18 points on every benchmark. Code is available at https://github.com/cruiseresearchgroup/DecompRCA.

Figures & tables

Figure 1: Our retrieval–reranking decomposition for RCA. Left: prior methods fail via propagation amplification (statistical) or graph quality bottlenecks (graph-based). Stage 1 retrieves a candidate set $\mathcal{C}$ using three complementary signals. Stage 2 reranks $\mathcal{C}$ with an LLM reasoner conditioned on per-candidate evidence and optional domain knowledge $\mathcal{D}$ , outputting permutation $\pi$ .

Industrial CPS benchmarks
	$K=5$			$K=10$			$K=15$
Dataset	mag	+ons	+stc	mag	+ons	+stc	mag	+ons	+stc
WADI ( $n{=}14$ )	0.36	0.36	0.50	0.43	0.50	0.57	0.64	0.57	0.64
SWaT ( $n{=}36$ )	0.33	0.25	0.31	0.44	0.44	0.61	0.53	0.61	0.67
HVAC ( $n{=}48$ )	0.21	0.06	0.25	0.27	0.56	0.54	0.35	0.65	0.63
Microservice benchmarks (RCAEval)

Table 1: Cumulative-budget Retrieval@

K

. At each

K

, the candidate pool is split across signals while the total pool size is held fixed: mag uses

K

magnitude items; +ons splits

K

between magnitude and onset; +stc splits it across magnitude, onset, and state-change (Appendix A ). Best per (dataset,

K

) in bold .

Industrial CPS benchmarks
	BARO	Retrieval@15	LLM (no DK)	Retrieval	Reranking
Dataset	top@1	(+stc, $K{=}15$ )	top@1	failure	failure	Rerank@1
WADI ( $n{=}14$ )	0.214	0.64	0.333 $\pm$ 0.082	0.36	0.31	0.52
SWaT ( $n{=}36$ )	0.194	0.67	0.213 $\pm$ 0.016	0.33	0.45	0.32
HVAC ( $n{=}48$ )	0.000	0.63	0.153 $\pm$ 0.043	0.38	0.47	0.24
Microservice benchmarks (RCAEval)

Table 2: Headroom decomposition for the multi-evidence LLM reranker without the domain document (no-DK; with-DK results in Table 3 ). Retrieval failure

=1-

Retrieval@15; reranking failure

=

Retrieval@15

-

top@1; Rerank@1

=

top@1

/

Retrieval@15, best per group in bold .

	Industrial CPS			Microservices (RCAEval)
Method	WADI	SWaT	HVAC	RE1-OB	RE1-SS	RE1-TT
BARO	0.214	0.194	0.000	0.784	0.856	0.560
RCD	0.071	0.083	0.083	0.304	0.248	0.144
$\epsilon$ -Diagnosis	0.000	0.028	0.146	0.072	0.232	0.008
CIRCA (PC)	0.214	0.083	0.062	0.536	0.632	0.320
CIRCA (FCI)	0.214	0.111	0.104	0.576	0.624	0.328

Table 3: End-to-end main results (top@1). LLM rows: mean

\pm

std over

n{=}3

runs at

T{=}1.0

, without and with the domain document; other rows single runs. HVAC LLM rows count the causes the retriever misses as misses (Appendix F ). Full top@1/3/5/Avg@5 grids in Appendix E .

Group	BB	LLM (no DK)	LLM (with DK)	$\Delta$	BB	LLM (no DK)	$\Delta$
	Retriever pool ( $K{=}15$ )				All candidates
WADI	0.308	0.410 $\pm$ 0.044	0.436 $\pm$ 0.089	$+$ 12.8	0.231	0.385 $\pm$ 0.077	$+$ 15.4
SWaT	0.200	0.229 $\pm$ 0.029	0.267 $\pm$ 0.017	$+$ 6.7	0.200	0.210 $\pm$ 0.017	$+$ 1.0
HVAC	0.188	0.188 $\pm$ 0.055	0.326 $\pm$ 0.012	$+$ 13.9	0.146	0.299 $\pm$ 0.032	$+$ 15.3
RE1-OB	0.784	0.883 $\pm$ 0.005	0.891 $\pm$ 0.012	$+$ 10.7	0.784	0.896 $\pm$ 0.014	$+$ 11.2
RE1-SS	0.856	0.859 $\pm$ 0.009	0.928 $\pm$ 0.000	$+$ 7.2	0.856	0.877 $\pm$ 0.005	$+$ 2.1

Table 4: Retrieval-controlled top@1 (true cause present in every pool). BB

=

best of nine baselines on identical pools;

\Delta

=

best LLM column minus BB, in pp. WADI/SWaT

n{=}13/35

(causes outside the evaluated sensor set cannot be pinned). Full grids in Appendix G .

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 2: Retrieval saturation per signal. Each signal is allocated its own full

K

. Grey dotted:

K{=}15

reference. The panels show three distinct saturation regimes (HVAC: onset-dominant amplitude-suppressed cause; WADI/SWaT: lag-dominated propagation; microservices: magnitude-dominant direct fault).

Statistical baselines
	WADI ( $n{=}14$ )				SWaT ( $n{=}36$ )				HVAC ( $n{=}48$ )
Method	top@1	top@3	top@5	Avg@5	top@1	top@3	top@5	Avg@5	top@1	top@3	top@5	Avg@5
BARO	0.214	0.357	0.357	0.314	0.194	0.306	0.417	0.306	0.000	0.021	0.062	0.025
RCD	0.071	0.214	0.214	0.186	0.083	0.111	0.111	0.106	0.083	0.146	0.146	0.133
$\epsilon$ -Diagnosis	0.000	0.071	0.071	0.043	0.028	0.056	0.083	0.061	0.146	0.146	0.146	0.146
Graph-based baselines (per-scenario PC/FCI graphs)

Appendix

Table 5: Full end-to-end grid, industrial CPS benchmarks (expands Table 3 ). LLM rows: mean

\pm

std over

n{=}3

runs at

T{=}1.0

; other rows single runs (RCD and

\epsilon

-Diagnosis seeded on HVAC, unseeded elsewhere). HVAC LLM rows are scored as in Appendix F .

Statistical baselines
	RE1-OB ( $n{=}125$ )				RE1-SS ( $n{=}125$ )				RE1-TT ( $n{=}125$ )
Method	top@1	top@3	top@5	Avg@5	top@1	top@3	top@5	Avg@5	top@1	top@3	top@5	Avg@5
BARO	0.784	0.936	0.976	0.912	0.856	0.992	1.000	0.962	0.560	0.856	0.952	0.810
RCD	0.304	0.432	0.432	0.403	0.248	0.512	0.512	0.446	0.144	0.184	0.184	0.176
$\epsilon$ -Diagnosis	0.072	0.224	0.336	0.221	0.232	0.496	0.616	0.459	0.008	0.072	0.112	0.062
Graph-based baselines (per-scenario PC/FCI graphs)

Appendix

Table 6: Full end-to-end grid, RCAEval microservice suites (expands Table 3 ). LLM rows report mean

\pm

std across

n{=}3

independent runs at temperature

1.0

. Other rows are single runs (RCD and

\epsilon

-Diagnosis unseeded).

Ranker (same candidate set $\mathcal{C}$ )	WADI	SWaT
mag-on- $\mathcal{C}$ ( $\approx$ BARO)	0.286 / 0.357 / 0.357 / 0.329	0.167 / 0.222 / 0.333 / 0.244
ons-on- $\mathcal{C}$	0.214 / 0.286 / 0.500 / 0.329	0.194 / 0.278 / 0.417 / 0.311
stc-on- $\mathcal{C}$	0.286 / 0.357 / 0.429 / 0.357	0.139 / 0.389 / 0.389 / 0.339
fusion-on- $\mathcal{C}$ (Borda)	0.286 / 0.500 / 0.571 / 0.443	0.194 / 0.306 / 0.361 / 0.294
LLM (no DK, $n{=}3$ )	0.333 / 0.381 / 0.381 / 0.362	0.213 / 0.472 / 0.481 / 0.402

Appendix

Table 7: Same-candidate controls, industrial CPS (the candidate lists the LLM was shown; top@1 / top@3 / top@5 / Avg@5). LLM rows are the

n{=}3

runs on the identical lists.

Ranker (same candidate set $\mathcal{C}$ )	RE1-OB	RE1-SS	RE1-TT
mag-on- $\mathcal{C}$ ( $\approx$ BARO)	0.784 / 0.936 / 0.976 / 0.912	0.856 / 0.992 / 1.000 / 0.962	0.560 / 0.856 / 0.912 / 0.802
ons-on- $\mathcal{C}$	0.240 / 0.336 / 0.696 / 0.403	0.368 / 0.576 / 0.864 / 0.594	0.176 / 0.224 / 0.232 / 0.214
stc-on- $\mathcal{C}$	0.552 / 0.920 / 0.976 / 0.837	0.360 / 0.968 / 1.000 / 0.810	0.000 / 0.128 / 0.360 / 0.157
fusion-on- $\mathcal{C}$ (Borda)	0.464 / 0.912 / 0.960 / 0.818	0.512 / 0.936 / 0.992 / 0.846	0.224 / 0.592 / 0.760 / 0.542
LLM (no DK, $n{=}3$ )	0.875 / 0.952 / 0.968 / 0.937	0.872 / 0.989 / 1.000 / 0.964	0.653 / 0.821 / 0.912 / 0.809

Appendix

Table 8: Same-candidate controls, RCAEval microservices (service-level; top@1 / top@3 / top@5 / Avg@5).

Ranker	top@1
mag-on- $\mathcal{C}$	0.000
ons-on- $\mathcal{C}$	0.104
stc-on- $\mathcal{C}$	0.250
fusion-on- $\mathcal{C}$ (Borda)	0.104
LLM (no DK, $n{=}3$ )	0.188 $\pm$ 0.055
LLM (with DK, $n{=}3$ )	0.326 $\pm$ 0.012

Appendix

Table 9: Same-candidate controls on HVAC (retrieval-controlled retriever pools, top@1). State-change wins at no-DK and domain knowledge flips the comparison; the rule that wins HVAC scores 0.000 on RE1-TT, so no fixed rule deploys uniformly.

Method	WADI	SWaT	RE1-OB	RE1-SS	RE1-TT	HVAC
BARO	0.308	0.200	0.784	0.856	0.560	0.000
RCD	0.154	0.114	0.440	0.384	0.280	0.188
$\epsilon$ -Diagnosis	0.077	0.057	0.176	0.368	0.104	0.146
PageRank (PC)	0.077	0.086	0.304	0.472	0.192	0.062
PageRank (FCI)	0.154	0.114	0.296	0.352	0.176	0.083
RandomWalk (PC)	0.077	0.114	0.264	0.480	0.208	0.042

Appendix

Table 10: Retrieval-controlled grid, retriever pool (

K{=}15

, reserved ground-truth spot): top@1 for every baseline on the identical pools. LLM row: no-DK,

n{=}3

Method	WADI	SWaT	RE1-OB	RE1-SS	RE1-TT	HVAC
BARO	0.231	0.200	0.784	0.856	0.560	0.000
RCD	0.231	0.086	0.304	0.248	0.144	0.083
$\epsilon$ -Diagnosis	0.000	0.029	0.072	0.232	0.008	0.146
PageRank (PC)	0.000	0.029	0.112	0.136	0.008	0.021
PageRank (FCI)	0.077	0.000	0.096	0.184	0.048	0.000
RandomWalk (PC)	0.077	0.000	0.080	0.112	0.064	0.000

Appendix

Table 11: Retrieval-controlled grid, all candidates (every sensor in the pool): top@1 for every baseline. LLM row: no-DK,

n{=}3

Group	gpt-oss-120b	Llama-3.3-70B	Best baseline
WADI	0.410 $\pm$ 0.044	0.308 $\pm$ 0.000	0.308 (BARO)
SWaT	0.229 $\pm$ 0.029	0.171 $\pm$ 0.029	0.200 (BARO, CIRCA)
RE1-OB	0.883 $\pm$ 0.005	0.856 $\pm$ 0.008	0.784 (BARO)
RE1-SS	0.859 $\pm$ 0.009	0.893 $\pm$ 0.005	0.856 (BARO)
RE1-TT	0.688 $\pm$ 0.024	0.597 $\pm$ 0.012	0.560 (BARO)

Appendix

Table 12: Backbone control: top@1 on identical retrieval-controlled retriever pools (no-DK,

n{=}3

Temperature	top@1	top@3	top@5
$T{=}0$	0.359 $\pm$ 0.044	0.538 $\pm$ 0.000	0.590 $\pm$ 0.044
$T{=}1.0$ (default)	0.410 $\pm$ 0.044	0.538 $\pm$ 0.077	0.615 $\pm$ 0.077
$T{=}2.0$ (provider max)	0.308 $\pm$ 0.000	0.513 $\pm$ 0.044	0.590 $\pm$ 0.044

Appendix

Table 13: Decoding-temperature control (WADI, retrieval-controlled retriever pool, no-DK,

n{=}3

per temperature).

Group	real / no-DK	anon / no-DK	real / with-DK	anon / anon-DK
WADI	0.410	0.410	0.436	0.487
SWaT	0.229	0.286	0.267	0.286
HVAC	0.188	0.139	0.326	0.375
RE1-OB	0.883	0.848	0.891	0.880
RE1-SS	0.859	0.877	0.928	0.933
RE1-TT	0.688	0.645	0.741	0.731

Appendix

Table 14: De-identification control: top@1 on retriever pools (

n{=}3

means) with real vs. anonymized sensor names and DK documents.

Benchmark	$-15$ min	$-5$ min	$+5$ min	$+15$ min
WADI	$-14.3$ pp	$0.0$ pp	$-12.5$ pp	$0.0$ pp
SWaT	$-22.2$ pp	$-22.2$ pp	$-16.7$ pp	$-25.0$ pp
HVAC	$-25.0$ pp	$-27.1$ pp	$-12.5$ pp	$-16.7$ pp
RCAEval $\times 3$ ( $\pm 1$ / $\pm 2$ min)	$\leq 3.2$ pp at every tested offset

Appendix

Table 15: Detection-timestamp sensitivity: change in Retrieval@15 of the +stc candidate pool under

t_{\mathrm{det}}

perturbation, matched scenario subsets on WADI/SWaT. RCAEval suites are perturbed at

\pm 1

and

\pm 2

minutes (windows of about 6 minutes per side); the entry reports the maximum absolute change across the three suites. No HVAC scenarios are skipped at any offset.

Method	WADI	SWaT	HVAC	RE1-OB	RE1-SS
PageRank (global PC)	0.071	0.000	0.000	0.104	0.184
RandomWalk (global PC)	0.000	0.000	0.000	0.176	0.208
CIRCA (global PC)	0.214	0.083	0.083	0.568	0.632
PageRank (global FCI)	0.071	0.000	0.000	0.000	0.216
RandomWalk (global FCI)	0.000	0.000	0.000	0.000	0.192
CIRCA (global FCI)	0.214	0.028	0.083	0.632	0.648

Appendix

Table 16: Ranking heads on global PC/FCI graphs learned from multi-day normal-operation corpora (top@1).

Root Cause Analysis

Root cause analysis identifies the initiating faults behind observed failures by tracing symptoms through causal and system dependencies.

See all 80 papers

StableRCA: Robust Graph-Agnostic Mechanism-Level Root Cause Analysis
Jun 4, 2026Xiaoyu Lin, Nicholas Tagliapietra, Kehan Li +2Root Cause Analysis Causal Intervention
Computer Science Department TU Darmstadt, Germany
Root Cause Analysis with Latent Confounders using Partial Ancestral Graphs
Jun 18, 2026Henrique O. Caetano, Rafael Arone, Carlos Dias MacielCausal Inference with Unobserved Confounding Partial Identification
Signal Processing Laboratory, São Carlos School of Engineering, University of São Paulo, São Paulo, Brazil · Guaratinguetá School of Engineering, São Paulo State University, Guaratinguetá, Brazil
Beyond Conditional Independence: Root Cause Analysis with Deep Causal Models
Sep 29, 2026Md Musfiqur Rahman, Kenneth Lee, Ziwei Jiang +3Structural Causal Models Causal Inference
Electrical and Computer Engineering, Purdue University · Computer Science, Johns Hopkins University · Intuit AI Research +1

Jun 4, 2026J/K move · Enter open · S save

StableRCA: Robust Graph-Agnostic Mechanism-Level Root Cause Analysis

Xiaoyu Lin, Nicholas Tagliapietra, Kehan Li, Lavdim Halilaj+1

Computer Science Department TU Darmstadt, Germany

Root Cause Analysis Causal Intervention

Open paper AskPDF arXiv

Root-Cause Analysis (RCA) seeks to identify the variables responsible for abnormal system behavior in complex domains such as manufacturing, cloud computing, and healthcare. Existing approaches face a critical bottleneck: graph-based causal methods can identify intervention targets but typically require a known or accurately estimated causal graph, while graph-free statistical methods either localize marginal anomalies rather than structural causes, or rely on restrictive assumptions about graph structure or functional form. We propose StableRCA, a local mechanism-level RCA framework that avoids global graph discovery by estimating local Markov boundaries and detecting conditional distribution shifts within them. Leveraging the Independent Causal Mechanism principle, we show that intervention targets can be identified with probability converging exponentially in sample size under faithful Markov boundary recovery and non-degenerate mechanism shifts. Experiments on synthetic benchmarks and five real-world datasets demonstrate that StableRCA is robust to graph misspecification, effective under multiple intervention targets, scalable to large systems, and reliable across diverse application domains. Code is available at: https://anonymous.4open.science/r/StableRCA-E362

Abstract

Root Cause Analysis

Explore similar work

StableRCA: Robust Graph-Agnostic Mechanism-Level Root Cause Analysis

Root Cause Analysis with Latent Confounders using Partial Ancestral Graphs

Beyond Conditional Independence: Root Cause Analysis with Deep Causal Models