acceptodds
Under review as a conference paper at ICLR 2027

CauRisk: When to Trust Causal Discovery Foundation Models under Distribution Shift

Abstract

Causal discovery foundation models can predict causal graphs for new datasets, but their confidence may not reflect graph accuracy when the underlying causal mechanisms change. We introduce CauRisk, a post-hoc framework that estimates graph-level prediction error under shifts in structural causal models (SCMs), without access to target ground-truth graphs at test time. Trained on synthetic source environments with known graphs, the risk estimator combines predictive uncertainty, stability across resampled datasets, and compatibility between the predicted graph and statistical patterns in the observed data. These diagnostics provide evidence about structural error without assuming that observational data can resolve non-identifiable causal directions. The discovery model remains frozen. CauRisk uses the estimated risk to decide which graph predictions to retain and which to defer. We specify an evaluation protocol that holds out entire causal mechanism and graph families, reserving target graphs exclusively for evaluation. The protocol compares CauRisk with uncertainty-only and stability-only estimators on failure detection, risk calibration, and risk–coverage trade-offs. It tests whether combining these diagnostics helps predict graph error beyond the source environments and supports decisions about when to rely on a predicted causal graph.

Primary area:
causal reasoning
Keywords:
causal discovery, foundation models, reliability estimation, distribution shift, uncertainty quantification, selective prediction

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_neHHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

In this paper, the authors introduced a post-hoc framework, namely CauRisk, that estimates graph-level prediction errors for causal discovery foundation models (Arrow, AVICI). Combining four diagnostics, uncertainty, stability, compatibility, and metadata, through mechanism-wise (worst-group) aggregation into a single risk score, CauRisk estimator calibrates and thresholds the result to decide which graph predictions to retain or defer. Three theoretical propositions are introduced to bound the results: selective ambiguity floor, a finite-grid validation bound, and a shift amplification after deferral. The framework is trained and evaluated on synthetic SCMs under three evaluation conditions: same-family, mechanism-shifts, and topology-shifts. The paper reports that across all three test conditions, CauRisk’s failure-detection ranking ability transfers well, unlike its numerical calibration and fixed acceptance threshold, especially under topology shifts, where the retained graphs have higher average error than the declared risk budget.

Critical strengths

1- The paper addresses an important problem: when to trust causal discovery foundation models’ predictions when deployed and evaluated across new datasets. 2- The authors supported their framework with theoretical analysis rather than relying solely on empirical results, and the framing effectively supported the motivation. 3- The experimental setup is designed to hold out mechanisms and graph-topology families from fitting and calibration, rather than reusing the same training setup, and uses these held-out families to assess transfer under distribution shift. 4- The paper’s main claim, that failure-detection ranking ability transfers under shifts while the numerical calibration and the fixed-threshold acceptance do not, is supported by the results and findings reported in the paper.

Critical weaknesses

1- The paper uses only synthetic data, and the mechanism and topology shifts evaluated are limited. All tasks use the same fixed parameters [d=10, n=500] and the same noise generation process, noise family, and scale for both training (discovery rows) and evaluation (audit rows), with only changes in the functional form or the edge-selection form. With this much similarity in data generation, the reported performance degradation is likely a lower bound on what a real deployment shift in datasets would look like. Also, generating data without latent confounding for a specific data type (continuous) is not always the case in real datasets, where data types can differ, and there will (in most cases) be latent confounding.

2- The paper includes several evaluation tables in the appendix and uses them in the analysis (Section 6: Results) without properly referencing them, such as in Q3 and Q4, leaving the reader to infer where each reported figure comes from. Also, it does not clarify how these tables connect to Table 1 in the main text. Each table presents a different evaluation process, such as ranking quality using AURC (Table 1), numerical calibration using MAE (Table 4), and the actual retain and defer decisions (Tables 8 and 13). It would help if the paper more clearly identified and distinguished these evaluation processes before presenting the numbers. They are conceptually distinct but presented consecutively, with no explicit distinction separating them. This made it harder for me, as a reader less familiar with this field, to follow what each table’s results were measuring. Additionally, the main algorithm of this method, which could provide a clearer description of the process, is placed in the last appendix of the paper (Appendix F) rather than in the main body.

3- As per my understanding, the compatibility scope is limited: it only tests conditional independencies implied by the predicted graph’s local Markov property, which does not reflect full faithfulness. Because of this, it cannot test whether a graph’s existing edge between adjacent variables should exist. It is excluded from the tested independence queries by the Markov property itself. So an extra added edge in this case cannot be directly identified by this diagnostic. I would appreciate further clarification from the authors on this point.

4- The relationship between the paper’s proposition theory and the deployed decision rule remains unclear to me. Propositions 1 and 2 seem to support earlier steps in the pipeline, while the final defer/retain decision is chosen by an empirical method that the paper itself explains is different from risk-control methods (Section 3.3). It is not clear to me from the text whether the propositions provide only motivation or an actual guarantee for the deployed decision-rule mechanism. I would appreciate further clarification on this point.

5- Minor Comments: Section 6: Results reports results for Q1- Q4 with minimal explanation/analysis, and the results are spread out across 14 tables that are not connected or properly summarised. Figure 2 in the paper has a minor misplacement of the y-axis title.

Questions for authors

1- Could the authors validate their CauRisk framework against a real-world dataset, even on a small scale, to check whether the learned diagnostic can track graph errors on a real dataset, given that all current evaluation is on synthetic SCMs?

2- Since only a small set of variables is used across all evaluations (d=10), could the authors extend their analysis to a larger scale to check whether their diagnostic patterns still hold? Or is the choice of (d=10) a backbone constraint rather than an evaluation choice?

Reviewer_3BoEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of evaluating whether a graph inferred by an amortized causal discovery method is trustworthy. For this, the authors train a risk estimator based on uncertainty, resampling stability, data-graph compatibility and other metadata to provide a score estimating the graph's structural error. The method is evaluated on artificial datasets.

Critical strengths

The paper addresses an important problem for trusting the output of causal discovery algorithms. Although the authors focus on amortized (foundation) models as the basis for discovering causal graphs, the approach does not seem inherently restricted to them, which makes it potentially useful more broadly. Graph falsification is a hard yet very important topic for trusting inferred graphs. As far as I know, this is the first work addressing graph-level reliability specifically for causal discovery foundation models in this way.

Critical weaknesses

The baseline comparison could include more methods focusing on (unsupervised) graph falsification. Even simple baselines, such as the number or fraction of Markov-condition violations or a likelihood-based score, would be useful, alongside more direct methods for causal graph falsification. In that sense, it is unclear how much value the approach truly adds over such "simpler" proxies. Besides, it lacks a theoretical guarantee that the accepted graphs satisfy the desired error budget under a distribution shift.

Questions for authors

The direction of this work seems overall novel to me. A few questions:

  • Why do you focus on having an amortized model produce the graph? Couldn’t you simply treat the causal discovery algorithm as a black box here (e.g., with the PC algorithm producing the graph)?
  • While the experiments are fair, the comparisons could include some directly related methods or at least some "typical" proxies, such as the number or fraction of Markov-condition violations or a likelihood-based score. A comparison with work directly focused on causal graph falsification, such as "Toward Falsifying Causal Graphs Using a Permutation-Based Test" by Eulig et al., would also be insightful. Adding these address my biggest concern regarding about how much better your assessment is compared with these "simpler" approaches. I acknowledge that some related methods focus on falsification rather than structural error. Nevertheless, could their outputs be compared on graph-error ranking or failure detection, while keeping the interpretation of falsification distinct from your retain/defer decision?
  • While there are some good theoretical insights, the implemented acceptance rule provides an empirical criterion rather than a guarantee that accepted graphs meet the specified error budget under a distribution shift. Could you clarify how the theoretical bounds relate to this rule and what additional assumptions would be needed for such a guarantee?

If these points can be addressed, I am happy to increase my score!

Reviewer_SDxsOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper provides an algorithm to, once structural causality model produces an DAG, estimate/evaluate reliability of that DAG matching underlying graph that produced the data.

Critical strengths

Approach seems sound. It is important, when deploying method, to evaluate if it actually worked correctly before using the result, so the question has merit.

Critical weaknesses

I didn't see strong practical, or theoretical applications -- while paper seems sound, I did not feel that I learned anything. The underlying algorithm that produced DAG might have an issue with applicability -- maybe it worked on synthetic data but didn't work in practice. On one hand something like this could be needed to evaluate correctness. But on the other hand, given only synthetic experiments, is it likely to be more correct than the approach that produced the graph in the first place?

Questions for authors

Do you think this will be applicable to a practical problem? Or alternatively, is there a significant theoretical result?

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

This paper studies graph-level reliability estimation for causal discovery foundation models under structural causal model (SCM) shifts and proposes CauRisk. The framework uses synthetic source environments with known graphs to train risk estimators around frozen discovery backbones. It combines predictive uncertainty, bootstrap stability, and data–graph compatibility through view-specific experts and mechanism-wise aggregation. Source-only isotonic calibration is then used to estimate graph error and decide which predictions to retain or defer. Experiments on Arrow and AVICI evaluate failure detection, numerical calibration, and risk–coverage trade-offs under held-out mechanism and topology shifts.

Critical strengths
  1. The paper clearly distinguishes graph-error ranking, numerical calibration, and selective acceptance. This is a meaningful and practically relevant problem, since confidence scores alone do not establish the reliability of a predicted causal graph.

  2. The source-only training and calibration protocol keeps target graphs out of risk estimation while leaving the discovery backbone frozen. This helps isolate reliability estimation from changes to the underlying graph predictor.

  3. The paper evaluates two causal discovery backbones across mechanism and topology shifts, with comparisons against diagnostic-specific and joint-feature baselines.

Critical weaknesses
  1. CauRisk does not consistently outperform simpler joint-feature baselines. Under Arrow topology shift, its AURC is 0.8444 versus 0.7901 for all-feature trees. Could the authors clarify the benefit of the proposed aggregation?

  2. The source-selected acceptance rule exceeds its risk budget under topology shift and defers all graphs at the stricter budget. Could the authors investigate alternative calibration or threshold-selection strategies?

  3. All experiments use synthetic datasets with only 10 variables and 500 observations. Could the authors include larger graphs, varying sample sizes, and datasets with externally validated reference graphs?

Questions for authors

See Weakness

Comments on the paper

1 comment
AuthorsOpenReview ↗

CauRisk Discussion Summary

Dear Area Chair and Reviewers,

Thank you for your reviews and constructive feedback. CauRisk addresses a practical question: after a discovery model returns a graph, how can we assess its structural error when the true graph is unknown? We learn how observable diagnostics relate to graph error on source tasks with known graphs, then use those diagnostics to score predictions from frozen discovery models on new datasets.

A risk score can help us decide which graphs to inspect first, even when a threshold chosen on source data misses its risk budget under shift. CauRisk achieves lower AURC than permutation-based graph falsification in all six Arrow and AVICI settings. The paired 95% bootstrap intervals favor CauRisk in each setting, including after averaging over exact-score ties (Table 4 and Appendix B.9).

In our offline evaluation, flagging about one in five predictions by CauRisk score identifies 23.5%–35.2% of high-error graphs, compared with 20.0%–20.4% under random review (Appendix I). The all-feature tree baseline finds the same share of high-error graphs and achieves lower AURC under Arrow topology shift (Table 1). The review experiment thus gives a concrete use for both learned scores.

The distinction matters for acceptance decisions. Under topology shift, the source rule set at a budget of 0.5 retains graphs with mean losses of 0.6111 for Arrow and 0.5500 for AVICI (Table 2). Appendix B.8 compares four source-only rules, and Proposition 3 shows why target coverage matters: the same amount of joint distribution shift contributes a larger term to the retained-risk bound when fewer target predictions are retained.

The extended tests cover graphs with up to 50 variables, different sample sizes and noise conditions, and latent-root perturbations. The causalAssembly experiments use industrial station graphs with generated observations (Appendix H.2). The Sachs case uses real cell measurements from one biological system and compares the predicted graphs with a literature reference (Appendix H.3).

Together, these results demonstrate the value of learning graph-level structural-error scores while distinguishing useful error ranking from the stronger requirements of reliable numerical acceptance under distribution shift.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.