acceptodds
Withdrawn from ICLR 2027

EvoCause: LLM-Guided Evolution of Causal Graphs for Root Cause Analysis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modern telecommunication, cloud, and microservice systems emit correlated alarm cascades when components fail. Root cause analysis (RCA) aims to identify the small set of alarms that initiate each cascade. A common approach learns a causal graph from observational logs and predicts all zero-in-degree alarms in each incident-induced subgraph. However, the learned graph remains fixed and cannot benefit from expert diagnoses of historical incidents. We close this loop with EvoCause. Expert labels constrain which alarms should be source nodes but do not specify the edge edits needed to satisfy those constraints. EvoCause uses a large language model (LLM) to propose semantically plausible graph edits, while deterministic code validates node identities and acyclicity and retains the best graph on a labeled alignment set. At test time, the refined graph alone produces transparent predictions without an LLM call. We also release \dataset, an expert-annotated benchmark from a production telecommunication network containing alarm events spanning alarm types over resources. On synthetic data, EvoCause initialized with the PC causal discovery algorithm outperforms the unrefined PC baseline, raising Node F1, Case EM, and Graph F1 by , , and percentage points, respectively, while reducing nSHD by . On TeleRCA, replacing human-readable alarm titles with anonymous identifiers lowers Node F1 and Case EM by and percentage points, respectively, indicating that alarm-name information contributes to graph refinement.

Primary area:
causal reasoning
Keywords:
Root cause analysis, causal discovery, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_JrSYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes EvoCause, an LLM-guided framework for refining causal graphs using expert-labeled historical incidents to improve root cause analysis (RCA). The method first constructs an initial alarm propagation graph using existing causal discovery algorithms. It then iteratively compares graph-based root cause predictions with expert annotations and employs an LLM to propose edge additions, removals, or reversals. These modifications undergo deterministic structural validation, and the best-performing graph is selected based on RCA accuracy over a labeled alignment set. During inference, root causes are identified directly from the refined graph without additional LLM calls.

The paper also introduces TeleRCA, a production telecommunications dataset containing 10,922 expert-annotated incidents and 485,681 alarm events. Experiments on synthetic data and TeleRCA report improvements in RCA performance across different causal discovery backbones. Additional analyses examine graph reconstruction quality, LLM configurations, and the contribution of alarm-name semantics.

Critical strengths
  • Valuable real-world data: TeleRCA contains a large number of expert-annotated incidents from a production telecom network, addressing the scarcity of realistic RCA datasets.
  • Clear problem formulation: The paper explicitly characterizes root-cause labels as indirect constraints and acknowledges that they cannot uniquely identify the underlying causal graph.
  • Practical framework design: LLM-guided graph refinement is combined with deterministic validation and selection, enabling interpretable RCA without LLM calls at inference time.
  • Meaningful experimental analysis: Experiments across different causal discovery methods demonstrate promising improvements, while synthetic graph reconstruction and alarm-title anonymization studies provide additional insights into the method.
Critical weaknesses
  • Missing dataset and appendices: Although the paper claims to release TeleRCA, no accessible dataset link is provided. The repeatedly referenced appendices are also missing, limiting reproducibility and verification of experimental details.

  • Unclear necessity of LLMs: The paper lacks comparisons with simpler non-LLM graph refinement strategies using the same expert feedback, making it difficult to isolate the actual contribution of LLM-guided editing.

  • Restrictive graph assumptions: A single stationary alarm-type DAG may not capture resource-specific or context-dependent causal relationships. Moreover, disconnected incident subgraphs can produce multiple spurious root cause predictions, but graph connectivity is not analyzed.

  • Limited causal validity: Optimizing root cause prediction accuracy does not necessarily recover correct causal relationships. The real-world experiments lack ground-truth causal graphs to validate the learned structures.

  • Limited generalization evaluation: The real-world evaluation relies on random incident-level splits from a single telecom network, without testing temporal shifts, unseen resources, or cross-system generalization.

  • Limited evaluation on external datasets: Although the paper discusses existing RCA datasets, EvoCause is evaluated only on the authors' synthetic data and TeleRCA, without validation on independent real-world datasets. Combined with the lack of a publicly accessible TeleRCA release, this limits reproducibility and makes it difficult to assess the method's generalizability across different systems.

  • Insufficient ablation studies: Although the paper evaluates different LLM backbones and alarm-name anonymization, the contributions of key components, such as batch-based feedback, optimization history, and graph selection, are not sufficiently examined in the provided manuscript. The missing appendices further prevent verification of the claimed component ablations.

Questions for authors
  1. Dataset availability: When will TeleRCA and the missing appendices be made available? Could the authors provide details on dataset construction, expert annotation procedures, and preprocessing?

  2. Necessity of LLMs: How does EvoCause compare with simpler non-LLM graph refinement methods that optimize the same objective using expert labels? What additional benefit does LLM-guided editing provide?

  3. Graph connectivity: How does EvoCause handle disconnected incident subgraphs or isolated alarm nodes, which may introduce spurious root cause predictions? Could the authors report connectivity statistics for the learned graphs?

  4. Generalizability: Why is EvoCause not evaluated on existing external RCA datasets discussed in the paper? Could the authors provide additional experiments using independent datasets or temporal splits to demonstrate generalization?

  5. Causal validity: Since the refinement objective optimizes RCA accuracy rather than causal correctness, how can the authors assess whether the modified edges represent meaningful causal relationships rather than overfitting to expert annotations?

  6. Ablation studies: Could the authors provide component-level ablations to quantify the contributions of batch-based feedback, optimization history, and graph selection? How sensitive is the method to batch size and the number of refinement rounds?

Reviewer_ADsEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper addresses root cause analysis of cascading incidents and proposes EvoCause. It uses labels rather than discovering root cause from unlabeled data. At its core, the contribution is to repeatedly ask an LLM to propose graph edits that improve alignment with the labels. The proposed edits are checked for validity before being applied.

Critical strengths
  • The paper addresses an important problem: quickly understanding the causes of failures is highly relevant.
  • Investigating improvements enabled by LLMs is a valid research direction.
Critical weaknesses
  • Insufficient baselines and coverage of existing work. Comparing a pipeline that uses labeled data against unsupervised methods is fundamentally unfair. The absence of a non-LLM graph-learning approach that uses labels is a fundamental weakness. I am unfamiliar with methods such as Chain of Events and CCCM; an explanation of these methods and the rationale for choosing them as baselines would be welcome. The evaluation lacks simple baselines that use the labeled data

  • Missing ablations. Such as asking the LLM to edit a random graph or an empty graph.

  • Unclear writing. I found the paper surprisingly difficult to follow. Many concepts are unclear when first introduced and only become clear later in the paper. For example, the abstract states: “Expert labels constrain which alarms should be source nodes but do not specify the edge edits needed to satisfy those constraints.” What edge edits are being referred to, and with respect to which graph? Why are edge edits introduced at this point? I assume the intended meaning is that specifying source nodes does not uniquely determine the remaining causal graph.

  • Limited contribution. The elaborate framework boils down to a small contribution.

  • Limited comparison with causal discovery methods. Score-based methods are missing from the comparison, and the evaluation of related approaches is limited. Relevant work includes, and probably a lot more and more recent work:

[1] Jalaldoust, Amirkasra, Kateřina Hlaváčková-Schindler, and Claudia Plant. “Causal discovery in Hawkes processes by minimum description length.” In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 6, pp. 6978–6987. 2022.

[2] Cueppers, Joscha, Sascha Xu, Ahmed Musa, and Jilles Vreeken. “Causal Discovery from Event Sequences by Local Cause-Effect Attribution.” In The Thirty-eighth Annual Conference on Neural Information Processing Systems.

[3] Noorbakhsh, Kimia, and Manuel Rodriguez. “Counterfactual temporal point processes.” Advances in Neural Information Processing Systems 35 (2022): 24810–24823.

Questions for authors
  • In Section 4.1, you state that individual incidents occur relatively independently. Could you clarify what this means and how it justifies splitting the data at the incident level? I am not criticizing this decision; I simply do not follow the reasoning.

  • What exactly does splitting at the incident level mean? Are you splitting the set of incidents or the data within each incident?

  • If incidents occur independently, how can labels from past incidents be helpful?

  • Why does the random baseline outperform, or perform nearly as well as, methods such as PC and NOTEARS?

  • Has the notation PaGPa_G been introduced, if yes where?

  • How do you derive a topological order?

Reviewer_FuGhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an RCA method named EvoCause. It uses LLMs to synthesize the constraints specified by domain knowledge to learn a causal graph. The paper releases a new RCA benchmark named TeleRCA based on a production telecommunications network.

Critical strengths

Benchmark dataset. The paper presents an RCA benchmark dataset TeleRCA, based on a production telecommunication network that may be useful for the community.

Critical weaknesses
  • Lacks of technical contributions. It argues that the existing causality-based RCA methods rely on a fixed learned graph (L15), but it goes on saying the proposed method, namely EvoCause, incorporates side information from the expert to retain \textit{the best graph on a labeled alignment set} (L19). Hence, the proposed method also gives a fixed learned graph. Many constraint-based discovery methods can take side information into account in the learning process. Fundamentally, the only difference is the use of LLMs. It is not surprising that the PC algorithm with background knowledge outperforms the ordinary PC. Incorporating background knowledge before failure time for root cause analysis can be done with any constraint-based observational causal discovery, whether it is used directly by a user or an LLM, the idea is not new given [1].

    • Reference - [1] Ikram, Azam, et al. "Root cause analysis of failures from partial causal structures." The 41st Conference on Uncertainty in Artificial Intelligence. 2025.
  • Lacks of clarity and correctness. First, the paper sets up the tone that causal discovery is able to recover Markov equivalence class in general and that additional knowledge from experts can help refine the graph, but in the actual learning stage of section 3.4, the first step of EvoCause is to have a causal discovery algorithm that recovers a DAG. It also incorrectly cites PC algorithm for this purpose as PC algorithm doesn’t return a DAG by design. The design of incorporating the expert label is also very ad-hoc from the perspective in causal discovery. The interpretations of the causal graphs output by directly providing background knowledge to a causal discovery algorithm verse using optimization procedures as a post-hoc step to search for a graph that best align with the additional constraints can be different. The paper does not provide a discussion on this.

  • Issues with citations. Line 122: the paper incorrectly cites a relevant work ‘CIRCA (Ikram et al., 2022) treats failures as interventions’ The algorithm from Ikram et al., 2022 is called RCD, not CIRCA. Within ML conferences, below are some relevant peer-reviewed causality-based RCA methods that should be cited:

    • [1] Ikram, Azam, et al. "Root cause analysis of failures from partial causal structures." The 41st Conference on Uncertainty in Artificial Intelligence. 2025.
    • [2] Lee, Kenneth, Zihan Zhou, and Murat Kocaoglu. "Root Cause Analysis of Failures in Microservices via Bayesian Root Cause Discovery." Forty-third International Conference on Machine Learning. 2026.
    • [3] Nagalapatti, Lokesh, et al. "Robust root cause diagnosis using in-distribution interventions." International Conference on Learning Representations. Vol. 2025. 2025.
    • [4] Han, Xiao, et al. "Root cause analysis of anomalies in multivariate time series through granger causal discovery." International Conference on Learning Representations. Vol. 2025. 2025.
Questions for authors
  1. Why should every zero-in-degree node be reviewed as a root cause?
  2. The paper says EvoCause supports different causal discovery algorithms without changing their learning procedures and preserves graph validity. But, in causal discovery, fully incorporating side information as additional constraints is not same as simply modifying the graphs according to the side information after getting the output of some causal discovery algorithms. The graphs coming from these two different approaches may not be the same. The paper seems to opt in for the latter. Could the authors please clarify what it means to preserve graph validity?
  3. Why does the characterization of the root-cause-consistent graph set matter? Isn’t it just the space of graphs that share the same set of zero-in-degree nodes consistent with the background knowledge provided by experts?
  4. Why should EvoCause use causal discovery algorithms and find the graph from the output that best align with the expert labels instead of directly incorporating the expert labels as additional constraints into the causal discovery method?
  5. Do the baselines in the experiment also have access to the expert labels?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.