acceptodds
Under review as a conference paper at ICLR 2027

CAUTION: Knowing When to Trust LLMs for Ensemble Causal Discovery

Abstract

Causal discovery from observational data remains challenging due to the fundamental limitations of purely statistical methods, such as statistical distinguishability within equivalence classes and sensitivity to finite sample sizes. While large language models (LLMs) offer a promising source of domain knowledge to complement statistical inference, existing LLM-augmented methods are vulnerable to LLM errors. A central challenge is therefore deciding when to trust LLM knowledge and when to trust statistical evidence. We propose CauTion, a trust-calibrated ensemble framework that addresses this challenge through annotation-free reliability estimation and selective arbitration. CauTion proceeds in three stages. First, an algorithm ensemble utilizes a consensus voting to resolve up to 96% of edges on which algorithms agree, achieving near-perfect accuracy on the filtered consensus edges. Second, a trust-calibrated arbitration mechanism estimates the relative reliability of the LLM and the algorithms via an annotation-free trust calibration procedure, which is then utilized to govern a trust-weighted voting process that restricts LLM arbitration exclusively to edges with unreliable algorithmic evidence. Third, a cycle repair step is applied to guarantee the final causal graph is validly acyclic. Experiments on six datasets demonstrate that CauTion consistently outperforms both data-centric and LLM-augmented baselines in SHD and F1, with strong robustness to LLM errors.

Primary area:
causal reasoning
Keywords:
Causal Discovery, Large language model

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_oVenOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CauTion, which combines statistical causal discovery algorithms with an LLM and estimates their reliability without human annotations. The method first accepts decisions on which PC, GES, and CaMML fully agree, leaving the remaining variable pairs for arbitration. Consensus pseudo-labels estimate LLM accuracy, while leave-one-out agreement estimates algorithm accuracy. Separate trust weights are computed for edge existence and direction. The LLM is queried only when its answer can change the weighted voting decision. A final repair procedure reverses or removes edges to eliminate cycles.

Experiments cover six bnlearn networks with five independently sampled datasets per network, comparing statistical and LLM-augmented methods. The paper also includes component ablations, comparisons across six LLM backends, and evaluation on a newly constructed Airport network. CauTion achieves the best or tied-best mean SHD and F1 in the main experiments, with particularly clear gains on larger networks.

Critical strengths

The method provides a concrete mechanism for limiting LLM errors. CauTion estimates reliability separately for existence and direction and queries the LLM only when its answer can affect the decision. The query threshold follows directly from the weighted voting rule, giving selective invocation a clear decision-based justification.

Critical weaknesses

The assumptions underlying the reliability estimates are not sufficiently validated. Section 3.3.2 assumes that LLM reliability transfers from consensus to disputed pairs and argues that sharing a domain and variable space supports this assumption. However, disputed pairs are selected through algorithm disagreement and may concentrate on harder, semantically ambiguous, or more complex relations. Reweighting existence classes adjusts class proportions but does not address within-class difficulty shifts. Likewise, leave-one-out evaluation measures agreement rather than correctness. With only three algorithms, shared biases can produce high agreement despite incorrect decisions. Taking the mean of the higher-performing half of these proxy estimates does not establish the ensemble’s actual accuracy. These quantities determine both trust weights and query selection, so this concern directly affects the central contribution.

Questions for authors

Using the known reference graphs for evaluation, could you report LLM existence and direction accuracy separately on consensus and disputed pairs? How large is the gap after class reweighting? How closely do leave-one-out estimates track individual-algorithm and ensemble accuracy on disputed pairs?

With algorithm outputs, LLM answers, and cycle repair held fixed, how do estimated trust weights compare with fixed weights? As a diagnostic upper bound, how do results change when reliability is computed using the reference graph?

Reviewer_SkptOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a framework that weights judgments from LLMs and statistical causal discovery methods on relationships between variable pairs to construct a more accurate causal graph. The framework uses consensus among statistical methods as a proxy for the correctness of LLM judgments. It then uses the accuracy estimated from consensus pairs to guide arbitration of disputed pairs.

Critical strengths
  1. The paper proposes a new calibration based method to weight the judgement of LLMs and statistical methods in causal graph discovery. By separating consensus variable pairs and disputed variable pairs, the paper provides an effective approach for combining the strength of statistical methods on numerical data and LLM judgement on uncertain variable pairs. This approach improves the performance of causal graph recovery and achieves high accuracy compared to the baselines.

  2. The experiments cover 6 datasets with 6 LLM backbones tested. The graph recovery performance measured by Structural Hamming Distance, F1 score and Structural Intervention Distance shows the proposed method has clear advantage compared to baselines.

Critical weaknesses
  1. The proposed method uses the variable pairs agreed by various statistical methods as pseudo-labels to infer the trust weights of LLMs on disputed pairs. However, the consensus and disputed sets may have different distributions and levels of difficulty. It is likely the disputed pairs may come from conflicting findings in domain research. LLM accuracy estimated from consensus pairs may not reliably reflect its accuracy on disputed pairs. The paper does not investigate the difference between the two sets of pairs.

  2. The LLM queries primarily rely on the models' internal knowledge, without explicitly retrieving supporting scientific literature or directly incorporating statistical evidence from the numerical data into individual LLM judgments. Although the proposed method combines LLM predictions with data-driven statistical outputs during arbitration, the lack of explicit evidence grounding may limit the reliability of LLM judgments in the LLM-based baselines.

Questions for authors

Related to weakness 1, the calibration method proposed in the paper can be seen as a memorisation test of LLMs on the target benchmark, i.e., if an LLM performs well on the consensus pairs of a benchmark, it is likely that the LLM has seen the benchmark and knows the causal relationships among its variables. The experimental results suggest that the LLMs are likely to have seen the Cancer dataset, but there is no clear pattern for the other datasets.

It is worth investigating whether familiarity with a benchmark contributes to the performance of the proposed method and the corresponding LLM-based baselines. It would be useful to prompt each LLM used in the experiments to reconstruct the causal graphs of the benchmarks to test whether it has memorised these datasets. This would help clarify the roles of LLMs and using the consensus pairs for trust calibration in the proposed method.

Reviewer_vpAiOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Goal:

  • use LLMs' domain knowledge within an ensemble of causal discovery methods while avoiding excessive LLM queries and overreliance on any individual information source or pre-determined weights that have to be guessed/assumed. Method:
  • Unanimous decisions from non-LLM algorithms are immediately accepted. The reliability of the ensemble and the LLM are estimated, and these are used to weight decisions on the remaining pairs to be decided, using LLM queries here. Finally, cycles are repaired using the LLM. Experiments:
  • Evaluation on six standard benchmarks, with ablations of the main components and robustness experiments across different base LLMs.
Critical strengths

Calibration scheme:

  • I found this to be clearly explained, and an effective way to use existing information (multiple different algorithms, many different edges) to compare algorithms without collecting more data, and without needing to picking trust weights in advance, i.e. 'out of a hat' without knowledge of peformance on this dataset.

Experiments:

  • The experiments section was clearly explained, and I appreciated the ablation and robustness sections to clearly separate components of the algorithm, which show that improvements hold across multiple base LLMs.
Critical weaknesses

Nothing major, see questions below.

Questions for authors

Comparison to other works:

  • my literature search identified both https://arxiv.org/pdf/2601.16715 and https://arxiv.org/abs/2405.01744 as works targeting the same problem with similar tools. I think some of the claims of the related works section should therefore be softened/updated with these in mind, and I would appreciate a discussion on the comparisons to these algorithms/why a comparison would not be reasonable.

Differences between consensus pairs and non-consensus pairs

  • Since LLM calibration uses consensus pairs (whereas algorithm calibration uses leave-one-out agreement across all pairs), I wonder how well the reliability estimates reflect actual performance on the disputed pairs it's used on. Could the authors compare the estimated reliabilities against ground-truth accuracies on those pairs? Would it be reasonable to test the LLM on all pairs and use the majority vote from the other algorithms as a pseudo-label, with some 'discount' for the fact that these pseudo-labels are not as reliable?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair,

We sincerely thank you and the reviewers for reading our manuscript so carefully and for offering constructive suggestions that helped us present our method more clearly, evaluate it more rigorously, and better define its scope.

To assist your assessment, we provide a brief summary of our work.

Paper overview

LLM domain knowledge can help resolve causal relationships that observational data leave ambiguous, but an incorrect LLM judgment can introduce a spurious edge or reverse its direction. A central challenge in LLM-augmented causal discovery is therefore deciding when to trust LLM knowledge and when to rely on statistical evidence without access to the true graph. We propose CauTion, an ensemble framework that uses agreement among statistical algorithms to estimate the relative reliability of the LLM and the algorithm ensemble separately for edge existence and direction. CauTion directly accepts decisions on which all the algorithms agree and queries the LLM on disputed pairs only when its judgments are estimated reliable. On each of six benchmarks, CauTion achieves the best or tied-best mean SHD and F1 among the methods compared.

Discussion with reviewers

  • Reviewer vpAi: The reviewer summarized that "I am satisfied with the discussion - my assessment of the paper remains positive, and I have no further questions."
  • Reviewer Skpt: The remained concerned is using LLM judgments without retrieving external scientific evidence, which we provide detailed analysis that external web search could expose benchmark ground-truth graphs and invalidate the evaluation. As our main focus is reliablity calibration, we leave external retrieval as a complementary extension for future implementation.
  • Reviewer oVen: At the time of writing, we have not yet received feedback on our rebuttal.

Recognized Strengths by Reviewers

  • Effective reliability calibration. Reviewer vpAi described the calibration scheme as “an effective way to use existing information” to compare algorithms without collecting more data. Reviewer Skpt similarly recognized “an effective approach” for combining statistical methods with LLM judgments on uncertain variable pairs.

  • Principled mitigation of LLM errors. Reviewer oVen recognized “a concrete mechanism for limiting LLM errors” and noted that the query threshold gives selective invocation “a clear decision-based justification”.

  • Strong empirical performance. Reviewer vpAi noted that “improvements hold across multiple base LLMs,” while Reviewer Skpt highlighted that the approach “achieves high accuracy compared to the baselines”.

  • Clear presentation of calibration and experiments. Reviewer vpAi described both the calibration scheme and the experiments as “clearly explained” and appreciated that the ablation and robustness sections “clearly separate components of the algorithm”.

Main Concerns and Questions Addressed

IssueReviewersResponse
Reliability transfer from consensus to disputed pairsvpAi, Skpt, oVenWe directly compared LLM accuracy on consensus and disputed pairs. LLM accuracies on consensus and disputed pairs were strongly correlated (Pearson’s r = 0.787/0.943 and Spearman’s ρ = 0.900/1.000 for existence/direction). Our LLM estimates achieved lower MAE than all tested baselines.
Validity of algorithm reliability estimatesvpAi, oVenWe evaluated our algorithm reliability estimates against actual accuracy on disputed pairs. The estimated and actual ensemble accuracies were strongly correlated (Pearson’s r = 0.761/0.922 and Spearman’s ρ = 0.700/0.900 for existence/direction). Our algorithm estimates achieved lower MAE than all tested baselines.
Compare estimated weights with fixed weightsoVenWe compared estimated weights with fixed weights as the reviewer required. Our estimated weights outperformed all fixed-weight settings.
Comparison with related methodsvpAiWe clarified the relationship to EMA and ALCM and added an EMA comparison with CauTion. CauTion achieved better performance on five benchmarks, with a tie on Cancer.
Potential benchmark memorizationSkptWe tested whether all six LLMs could reconstruct the benchmark graphs. 24 of 36 model–benchmark combinations obtained zero edge F1, demonstrating poor benchmark memorization across the tested LLMs.
Involving external scientific evidenceSkptWe provided a detailed analysis explaining that web search could expose benchmark ground-truth graphs and invalidate the evaluation. As our main focus is reliability calibration, we leave external retrieval as a complementary extension for future work.

We hope that our detailed responses have carefully addressed all concerns, and that these efforts will be taken into consideration in the final assessment. We are truly grateful for your time and expertise throughout this review process.

Best regards,
The Authors of Submission #61821

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.