acceptodds
Under review as a conference paper at ICLR 2027

Consistency evaluation of benchmarks used for causal discovery

Abstract

In graphical causal model, causal discovery aims to construct a causal graph based on numerical data and domain knowledge in plain text. However, the evaluation of causal discovery methods remains a challenge in the area as the progress of domain researches often makes benchmark causal graphs contain misaligned knowledge. This problem especially affects the evaluation of large language model (LLM) based causal discovery methods as they are sensitive to the new discoveries in the literature. This work is the first to systematically study the quality of benchmark causal graphs. Specifically, we design a pipeline that automatically retrieves relevant research papers from scientific databases, and prompts LLMs to check the consistency between the benchmark causal graphs and domain research papers. We evaluate 10 popular real-world benchmarks, for which our pipeline in total proceeds 35,340 domain papers. Our results show that popular benchmarks vary significantly in their consistency with domain research, with clear implications for causal discovery research.

Primary area:
causal reasoning
Keywords:
Causal discovery, benchmark evaluation

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

This paper examines whether commonly used causal discovery benchmark graphs agree with domain literature. The authors first survey benchmark usage in recent causal discovery papers, identifying 69 empirical benchmarks, and then select ten frequently used benchmarks for further analysis.

The proposed pipeline retrieves papers for variable pairs, extracts conditional independence relations from benchmark graphs, and uses an LLM to assess agreement between these relations and reported statistical findings. Approximately 35,000 papers are processed. The fraction of inconsistent results among answerable cases is used to measure disagreement with domain knowledge. A manual check of 100 randomly sampled papers reports 90% verifier accuracy. The main finding is that inconsistency rates vary substantially across benchmarks, with lower rates for Sachs and higher rates for Asia and Child.

Critical strengths

The paper addresses an important and often overlooked evaluation problem. Lower structural error against an empirical benchmark does not automatically imply better agreement with the real world. This distinction is particularly relevant for LLM methods that use variable semantics and may disagree with a fixed reference graph. The authors attempt to build a traceable literature-auditing workflow. The work combines a benchmark-usage survey, large-scale retrieval, conditional-relation checking, and manual validation. Requiring supporting passages from the source papers also facilitates later inspection. With a sounder core methodology, this resource could help benchmark maintenance.

Critical weaknesses

The graph-based extraction does not correctly and fully implement the required conditional independence reasoning. Section 3.1 interprets d-connection as correlation, but the Markov property only guarantees that d-separation implies conditional independence. The converse requires additional assumptions such as faithfulness, and statistical dependence need not imply correlation in its usual sense. More concretely, the appendix algorithm considers only collider-free paths and finds minimal sets blocking those paths. It does not check paths opened by conditioning on colliders or their descendants. It also immediately returns independence when no collider-free path exists. For example, in X → C ← Y, marginal independence does not imply independence conditional on C. This algorithm is insufficient for the arbitrary conditioning-set comparisons described in the main text. Checking only minimal separating sets also does not establish the relation for the conditioning set actually used in a paper. Since the inconsistency rates depend on these judgments, this issue could change the main findings.

Questions for authors

Is the appendix algorithm the implementation used for the reported results? If so, how are paths opened by conditioning on colliders or their descendants handled? Could all actual queries be checked using a complete d-separation procedure, with updated inconsistency rates and rankings?

How does the pipeline handle a paper whose conditioning set is not a minimal separating set? Is faithfulness assumed when interpreting d-connection? How does the verifier distinguish statistical dependence, correlation, predictive ability, and statistical nonsignificance?

Reviewer_e4mXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper aims to re-evaluate existing causal discovery benchmarks in the context of LLMs by examining how social and contextual knowledge has changed over time. The authors investigate discrepancies between the assumptions underlying these benchmarks and current domain knowledge available to LLMs.

Critical strengths

The motivation for re-examining existing causal discovery benchmarks is valuable. The public release of the code is also a positive aspect of the work.

Critical weaknesses

1. Overall presentation and completeness

The paper has several issues related to its organization, presentation, and completeness. For example, Appendix A is empty, and the AI usage statement is incomplete. There are also inconsistencies between the tables and the accompanying descriptions. For instance, the paper reports reviewing 86 papers in one section and 96 in another. Some figures are also difficult to interpret.

2. Overgeneralization of the main claims

The authors argue that LLMs may struggle with existing causal discovery benchmarks because their knowledge reflects current social and contextual understanding, which may differ from the assumptions considered valid when the benchmarks were originally developed. However, the title and several parts of the manuscript appear to suggest that existing causal discovery benchmarks are fundamentally flawed. This claim seems too broad given the evidence presented. The authors should narrow their claims to the use of these benchmarks for LLM-based causal discovery or the evaluation of LLMs.

3. Inconsistency between datasets and ground-truth graphs

The paper's recommendation to update or replace misaligned benchmarks does not sufficiently distinguish between inconsistencies with current domain knowledge and inconsistencies between the data and their ground-truth graphs. For semi-synthetic benchmarks such as Asia, data are generated from the specified causal graph, making it the correct ground truth for the data-generating process by construction. Even if the graph reflects historical assumptions that differ from current knowledge, this does not make it incorrect for the generated data. The reported inconsistency therefore concerns the relationship between the graph and current domain knowledge, rather than a mismatch between the graph and the data. Similarly, the Sachs dataset is based on experimentally studied protein-signaling mechanisms that are less susceptible to the social and historical changes considered in the paper, which may partly explain its low inconsistency rate. The concern is therefore more directly relevant to LLM-based methods that rely on external knowledge than to conventional data-driven causal discovery. If the authors propose updating these benchmarks, the ground-truth graph and the corresponding data-generating process must be updated together to preserve a valid evaluation setting. For real-world benchmarks, the data and ground-truth graph should likewise reflect the same temporal and contextual setting. Otherwise, modifying the graph alone could introduce inconsistencies and unfairly penalize data-driven methods. Furthermore, as mentioned earlier, the title and main text treat conventional causal discovery benchmarks as fundamentally problematic. This extends the point that this critique must be restricted solely to the evaluation of LLM-driven causal discovery, while also highlighting the inherent limitation of non-data-driven, LLM-based causal discovery

4. Reliance on LLMs to assess ground-truth validity

The proposed evaluation framework appears to rely on LLMs to determine whether existing ground-truth relationships remain valid, without sufficient involvement from domain experts. Given that the central question concerns changes in domain knowledge over time, assessments by multiple domain experts would provide a more reliable basis for evaluating the proposed revisions. A more convincing approach would involve several experts from the relevant domains, with LLMs serving as supplementary tools to support or cross-check their assessments.

5. Insufficiently defined criteria for inconsistency

The criteria used to determine whether a causal relationship is inconsistent with current knowledge are not sufficiently explained. This concern is closely related to the reliance on LLM-based judgments. Clear and explicit criteria are needed to make these assessments interpretable and reproducible.

Questions for authors

Relevance of the datasets listed in the appendix

Figure 10 in Appendix C includes datasets such as MATH, ImageNet, and GSM8K, which, to my knowledge, are not standard causal discovery benchmarks. Could the authors clarify why these datasets are included and verify whether they were correctly identified and extracted as causal discovery benchmarks?

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

This paper studies whether commonly used causal discovery benchmarks still agree with current domain research. The authors first survey popular benchmarks, then build an LLM-based pipeline that retrieves research papers, extracts conditional association results, and compares them with the relations implied by d-separation in each benchmark graph. The pipeline processes 35,340 papers and reports an inconsistency rate for each benchmark. The paper finds large differences across benchmarks and argues that older benchmarks tend to be less aligned with recent research. The authors assess the LLM verifier by manually checking 100 cases and report an overall accuracy of 90%.

Critical strengths
  1. The paper studies an important and often overlooked problem. Causal discovery methods are usually evaluated against benchmark graphs that are treated as ground truth. Checking whether these graphs still agree with current scientific knowledge is valuable, especially because many widely used benchmarks are old and continue to be reused. If done reliably, this type of audit could improve how the community selects and maintains benchmarks.
  2. The paper proposes a scalable way to connect benchmark graphs with a large body of domain research. The pipeline combines paper retrieval, d-separation queries, and paper analysis across several widely used benchmarks and thousands of papers. The authors also release the prompt. This level of scale and transparency is useful and could provide a starting point for a stronger benchmark-auditing system.
Critical weaknesses
  1. The evaluation does not clearly separate disagreement with the benchmark graph from differences in scientific setting. Section 3.2 (page 6) says the verifier checks whether a paper studies the two variables and controls for a covariate set Z, but the paper does not specify criteria for comparability of population, variable definition, measurement, time period, region, study design, or experimental condition. It is also unclear whether the paper's covariate set must exactly match Z, or whether related variables and additional covariates are accepted. This matters because conditional associations can differ across settings without directly contradicting the benchmark. Section 4.2 (page 9) acknowledges that studies with additional latent confounders "may not invalidate the previous benchmarks," yet the paper does not explain whether such cases are identified as Unknown, excluded, or reported separately. The inconsistency rate may therefore mix direct disagreement with the benchmark graph and differences in scientific setting, which limits how the main conclusions can be interpreted.
  2. The reported validation is insufficient to establish the reliability of the LLM verifier under OCR conversion and possible truncation. Page 7 states that PDFs are converted with Mistral OCR and analyzed by Qwen3-30B-A3B-Thinking-2507, and Appendix D.4 (page 32) notes that the input is "possibly truncated." The paper does not report how often truncation occurs, how retained text is selected, or whether the relevant variable definitions, covariates, and results remain in the input. It also does not separately evaluate OCR quality, evidence extraction, variable/covariate matching, and final classification. The 100-paper manual check reports only an overall accuracy of 0.9, without breakdown by answer type, benchmark, OCR quality, truncation status, or matching difficulty. Since Unknown judgments are excluded from the inconsistency rate and the benchmark rankings depend on automated Consistent/Inconsistent decisions, it remains unclear whether the main findings are stable under verifier errors.
Questions for authors
  1. In Section 3.2, page 6, what exact criteria are used to match a variable in a paper to a benchmark node? Must the definition and measurement method match, or can related variables and proxies be accepted?
  2. Does the pipeline check whether the paper and benchmark use comparable populations, regions, time periods, study designs, and experimental conditions? If these factors are not checked, how does the analysis distinguish disagreement with the graph from a difference in scientific setting?
  3. Section 4.2, page 9, notes that studies with additional latent confounders may not invalidate earlier benchmarks. Are such cases separated from direct graph disagreements in the reported results? If not, how do the inconsistency rates and benchmark rankings change after excluding clear setting or covariate mismatches?
  4. Appendix D.4, page 32, states that the paper content may be truncated. What fraction of papers are truncated, how is the retained content selected, and what happens when the relevant methods, covariates, tables, or results are not included?
  5. For the 100 manually checked papers, what was the sampling procedure? Can the authors report accuracy separately for the extracted relation labels (Unknown, Correlated, and Independent) and for the final consistency labels (Unknown, Consistent, and Inconsistent)? Were multiple annotators used, and if so, what was their agreement?
  6. How sensitive are the reported inconsistency rates and benchmark rankings to OCR errors, truncation, variable-matching errors, and classification errors? Do the main findings remain similar in an error-sensitivity analysis based on the manually estimated error rates?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.