Pith. sign in

REVIEW 4 major objections 6 minor 23 references

CARL-GT: Evaluating Causal Reasoning Capabilities of Large Language Models

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Large language models perform at or near chance on causal reasoning when the input is a serialized data table, according to CARL-GT, a benchmark that tests causal graph reasoning, knowledge discovery, and decision-making.

desk verdict The CARL-GT benchmark is a sensible addition to causal-reasoning evaluation, but Table 2's impossible 0.00 accuracies make the main experimental evidence unreliable. read the letter →

arxiv 2412.17970 v1 pith:VK4PM4LB submitted 2024-12-23 cs.CL cs.LGstat.ME

Ruibo Tu , Hedvig Kjellström , Gustav Eje Henter , Cheng Zhang This is my paper · ORCID
classification cs.CLcs.LGstat.ME
keywords causalreasoningbenchmarklargelanguagemodelsknowledgediscoveryd-separationcounterfactualtabulardatadecisionmaking
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CARL-GT is a benchmark that tests large language models on three families of causal tasks: reading causal graphs, inferring causal structure from observational tables, and estimating intervention and counterfactual quantities. The paper's central finding is that today's open-source LLMs are still weak at these tasks, particularly when they must discover causes from tabular data: on d-separation judgments (whether two variables are independent once a third set is fixed) from tables, several models score near 0.5 on a 0-to-1 scale, the chance level, and on causal-direction questions three of five models return zero accuracy because they decline to answer. The benchmark generates its own graphs and tables with known cause-effect equations, so every question has a ground truth, and it supplies prompts that work without task-specific examples. The reason to care is that it offers a direct way to test whether LLM reasoning transfers from language and math to the structured-data settings real-world decision support actually uses.

What carries the argument

The central object is the CARL-GT benchmark: synthetic directed acyclic graphs $G=(V,E)$, tabular datasets generated from linear structural causal models defined by those graphs, and task families that ask for adjacency matrices, d-separation relationships, causal directions, intervention distributions, and counterfactual expectations. The design that carries the argument is the pairing of a causal graph with a serialized table for the same underlying data-generation process, so the same causal relationship can be probed both as a verbalized graph question and as a table-only discovery question; this makes the difficulty gap between the two settings interpretable. Ground-truth answers for intervention and counterfactual tasks are obtained by fitting structural causal models to larger samples, and the reported metrics are classification scores for d-separation and direction, F1 scores for graph recovery, and mean absolute error for inference. The prompt templates, with graph data in an expert-analyst text form and tables in Markdown, are part of the benchmark rather than an afterthought.

What would settle it

Run the same CARL-GT prompts with human-verified answer extraction or constrained decoding that forces the output into yes/no or a float, then compare classification scores and accuracy. If the models that currently score zero on table-based causal direction, such as Qwen2-7B, Gemma2-9B, and Mixtral-8x7B, jump above chance, the extraction step is responsible for the reported failure; if they remain at zero, the failure is in the models' causal reasoning itself.

Watch

Extended reading notes

Core claim

The paper claims that current large language models do not yet do reliable causal reasoning, and that the weakness is most visible when the evidence is a serialized table rather than a verbalized graph. On explicitly provided causal graphs, the tested models handle adjacency questions reasonably but struggle with d-separation and direction; on tables alone, d-separation classification hovers at chance and causal-direction questions produce zero accuracy for Qwen2-7B, Gemma2-9B, and Mixtral-8x7B because the models answer that the information is insufficient. The paper also claims that performance across task categories is not what one would expect from a single reasoning ability: graph-based adjacency prediction and counterfactual estimation correlate strongly, while intervention estimation correlates negatively with most other tasks. The contribution is the benchmark itself, offering generated causal graphs, matching tabular data, ground-truth answers, and reusable zero-shot prompts as a reproducible measuring stick for causal reasoning in LLMs.

Load-bearing premise

The load-bearing premise is that the pipeline that extracts final answers from each model's free-text response, an unspecified LLM asked to summarize the answer as yes, no, or a float, faithfully preserves the causal reasoning the evaluated model actually performed; if that extraction distorts, discards, or replaces answers, every reported score is suspect.

Editorial extensions

If this is right

  • Evaluation of causal ability in LLMs should include table-only tasks, because verbalized-graph results alone overstate causal competence.
  • Open-source models today are not dependable for zero-shot knowledge discovery from observational data, so deployment in discovery or decision-support settings should not assume causal reliability.
  • Because question difficulty rises sharply from 10 to 51 nodes in the causal graphs, current models are unlikely to scale to realistic problem sizes without further work.
  • The synthetic generation gives ground truth for every question, so the benchmark can separate genuine causal computation from memorized facts about causality.
  • The observed negative correlation between intervention estimation and the other tasks suggests that causal reasoning is not a single competence; improvement work needs to target sub-skills separately.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper does not run: replace the LLM-based answer extraction with constrained decoding that forces a yes/no or float token, to test whether the near-chance scores are reasoning failures or measurement artifacts.
  • The correlation analysis rests on five models, so the reported task relationships are fragile; adding more models and per-instance scores could confirm or overturn them.
  • Because the data are generated by linear Gaussian mechanisms, the benchmark could be extended to nonlinear and discrete systems, and the ordering of task difficulty (adjacency versus d-separation versus direction) would probably change.
  • The near-universal refusals on table-based causal direction suggest that current pretraining does not instill the inductive bias needed to orient edges; a targeted training objective on synthetic causal tables is a concrete next step the benchmark could monitor.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes CARL-GT, a benchmark for evaluating causal reasoning in large language models using synthetic causal graphs and tabular data. The benchmark defines tasks in three categories: causal graph reasoning (adjacency, d-separation, causal direction), knowledge discovery from tables (d-separation, causal direction), and decision making (interventional and counterfactual inference). The authors evaluate five open-source LLMs in a zero-shot setting with serialized tables and graph descriptions, and report that models are particularly weak on table-based knowledge discovery, often performing at or below chance. They also analyze correlations between task performances and conclude that tasks in different categories are more correlated than tasks in the same category. The benchmark code is released.

Significance. If the measurement issues are resolved, CARL-GT is a useful and easily extensible synthetic benchmark for a relatively under-served capability area. The synthetic generation from structural causal models is principled, the release of code and prompts supports reproducibility, and the inclusion of serialized tabular data is practically relevant. However, the headline quantitative claim about table-based causal reasoning rests on Table 2, whose scoring appears internally inconsistent; until that is fixed, the magnitude of the reported weakness is unverified. The task-correlation analysis also needs stronger statistical support.

major comments (4)
  1. [Section 4.2, Table 2] The caption of Table 2 states that unknown LLM answers are replaced by random yes/no, but the table reports Acc=0.00 for Qwen2-7B, Gemma2-9B, and Mixtral-8x7B. For a binary question, uniform random replacement yields an expected accuracy of 0.5 regardless of label balance, so 0.00 is impossible under the stated procedure. This suggests that the replacement was either not applied, that replaced answers were always scored as incorrect, or that the entries are placeholders. Because this table is the primary evidence for the claim that LLMs are weak in table-based causal reasoning, the scoring procedure must be corrected and the table regenerated or clearly annotated.
  2. [Table 2] The Mixtral row appears in Table 2 with Acc=0.00 even though the caption says Mixtral results are omitted due to GPU memory limits. This is a direct contradiction. Either the row should be removed or the caption amended; the text in Section 4.2 also states that Mixtral could not answer the causal-direction questions, so the table should distinguish between "not evaluated" and "evaluated but refused."
  3. [Section 4.1, Templates 4.2 and 4.3] The two-step answer extraction uses an unspecified LLM-based extractor. If the extractor is the same model being evaluated, a model that fails to comply with the output format or refuses to answer could have its verbatim response misparsed, systematically biasing the reported metrics. The paper should specify the extraction model (or state that human annotation was used) and provide evidence that extraction errors do not drive the results, particularly for the 0.00 accuracy values.
  4. [Section 4.5, Table 6] The correlation analysis uses ranking vectors from only five LLMs, yielding only 5 data points for each coefficient. With such a small sample, the reported correlations (e.g., -0.9, 0.9) are highly unstable and not statistically meaningful; the paper even concedes that more LLMs would be needed. Since the abstract presents the task-relationship finding as a contribution, this analysis should be either expanded to more models or explicitly framed as an illustrative observation without the quantitative correlation claim.
minor comments (6)
  1. [Abstract] The phrase "casual reasoning" appears in the abstract and should be corrected to "causal reasoning."
  2. [Section 1] The sentence "Therefore, we introducing a benchmark" should be "Therefore, we introduce a benchmark."
  3. [Section 3.2] The word "summerized" should be "summarized."
  4. [Section 4.1] The spelling of "MarkDown" and "Markdown" is inconsistent; please use one form throughout.
  5. [Template 4.1] The edge notation in the template (e.g., "V 1− > V2") has inconsistent spacing and should be formatted uniformly, e.g., "V1->V2."
  6. [References] The MathVista reference appears twice (Lu et al., 2023 and Lu et al., 2024) with the same title and author list; this duplication should be removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CARL-GT is an empirical benchmark; its conclusions are measured LLM outputs against independently generated SCM ground truth, not a derivation that reduces to its inputs.

full rationale

CARL-GT contains no derivation chain in which a claimed prediction is equivalent by construction to an input or fitted parameter. The benchmark data are synthetic tables generated from structural causal models with random parameters (Section 4.1), and the evaluation metrics compare LLM responses to ground-truth answers computed independently, e.g., d-separation from graph algorithms and intervention/counterfactual expectations from learned SCMs on 15,000 samples. The conclusion that LLMs are weak at tabular causal discovery is not forced by construction: the models could in principle answer correctly, and several models score above chance on d-separation AUC. The self-citations to Tu et al. (2024) for data generation, question selection, and metrics are construction references to a reusable, stated-assumption synthetic-data framework (linear uniform SCMs); they are not invoked as a uniqueness theorem or as the sole justification for the central claim, so they are not load-bearing and do not raise the circularity score under the independent-support rule. The Table 2 anomaly (0.00 accuracy for Qwen2-7B, Gemma2-9B, and Mixtral under a stated uniform random replacement for unknown answers) is a scoring-validity/correctness issue, not circularity, because it does not make the reported weakness true by definition; it affects whether the numbers are trustworthy, not whether they reduce to their inputs. No equation is defined in terms of the quantity it claims to explain, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The benchmark does not introduce new theoretical entities, particles, or forces. It creates a new evaluation dataset and task suite, which are not counted as invented entities in the sense of causal or physical postulates. The only free variables are the random parameters in the synthetic SCMs, which are not fitted to data and are not free parameters in the derivation sense.

assumptions (3)
  • domain assumption Synthetic data generated from linear uniform structural causal models are representative of real-world tabular causal reasoning scenarios.
    The benchmark's validity relies on the premise that synthetic tabular data and graph descriptions capture the causal reasoning skills needed in real-world applications like healthcare. Stated in Section 3.1 and experimental setup (Section 4.1).
  • domain assumption LLM-based answer extraction produces faithful structured answers.
    Template 4.3 assumes that asking a language model to extract neighbors, yes/no answers, or float numbers from a response does not introduce systematic bias. This is load-bearing for all metrics.
  • domain assumption Ground-truth intervention and counterfactual quantities computed from estimated structural causal models on 15,000 samples are accurate.
    In Section 4.2 (Decision-making), the authors state the ground truth is computed by learning SCMs from data rather than from the true generative model, which assumes the estimation error is negligible.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CARL-GT: Evaluating Causal Reasoning Capabilities of Large Language Models." pith.science (2026). https://pith.science/paper/VK4PM4LB

@misc{pith2026241217970,
  author       = {Pith},
  title        = {Pith review of: CARL-GT: Evaluating Causal Reasoning Capabilities of Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VK4PM4LB}},
  note         = {Machine review of arXiv:2412.17970}
}
read the original abstract

Causal reasoning capabilities are essential for large language models (LLMs) in a wide range of applications, such as education and healthcare. But there is still a lack of benchmarks for a better understanding of such capabilities. Current LLM benchmarks are mainly based on conversational tasks, academic math tests, and coding tests. Such benchmarks evaluate LLMs in well-regularized settings, but they are limited in assessing the skills and abilities to solve real-world problems. In this work, we provide a benchmark, named by CARL-GT, which evaluates CAusal Reasoning capabilities of large Language models using Graphs and Tabular data. The benchmark has a diverse range of tasks for evaluating LLMs from causal graph reasoning, knowledge discovery, and decision-making aspects. In addition, effective zero-shot learning prompts are developed for the tasks. In our experiments, we leverage the benchmark for evaluating open-source LLMs and provide a detailed comparison of LLMs for causal reasoning abilities. We found that LLMs are still weak in casual reasoning, especially with tabular data to discover new insights. Furthermore, we investigate and discuss the relationships of different benchmark tasks by analyzing the performance of LLMs. The experimental results show that LLMs have different strength over different tasks and that their performance on tasks in different categories, i.e., causal graph reasoning, knowledge discovery, and decision-making, shows stronger correlation than tasks in the same category.

Figures

Figures reproduced from arXiv: 2412.17970 by the authors.

Figure 1
Figure 1. Our benchmark, CARL-GT, for evaluating causal reasoning capabilities of LLMs from three aspects, i.e., causal graph reasoning, knowledge discovery, and deci￾sion making. Fig. 1a shows that a task consists of input data, such as causal graphs and tables, questions about causal statements, ground-truth answers, and parsers for evalu￾ation. Furthermore, Fig. 1b shows that each aspect has multiple tasks, e.g., the causa… view at source ↗
Figure 2
Figure 2. Relationships between tasks in the same categories. [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Relationships between causal inference-based and other tasks. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Correlation between causal graph reasoning and causal discovery-based [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 1 canonical work pages

  1. [4]

    Which modality should I use - text, motif, or image? : Understanding graphs with large language models

    Debarati Das, Ishaan Gupta, Jaideep Srivastava, and Dongyeop Kang. Which modality should I use - text, motif, or image? : Understanding graphs with large language models. In Findings of the Association for Computational Linguistics: NAACL 2024,

    work page 2024

  2. [5]

    Gpt4graph: Can large language models understand graph structured data ? an empirical evalua- tion and benchmarking

    16 Jiayan Guo, Lun Du, Hengyu Liu, Mengyu Zhou, Xinyi He, and Shi Han. Gpt4graph: Can large language models understand graph structured data ? an empirical evalua- tion and benchmarking. arXiv:2305.15066,

    arXiv

  3. [7]

    Measuring mathematical problem solving with the math dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874,

    arXiv

  4. [9]

    Jinhyuk Lee, Anthony Chen, Zhuyun Dai, Dheeru Dua, Devendra Singh Sachan, Michael Boratko, Yi Luan, S´ebastien MR Arnold, Vincent Perot, Siddharth Dalmia, et al

    URL https://openreview.net/forum?id=e2wtjx0Yqu. Jinhyuk Lee, Anthony Chen, Zhuyun Dai, Dheeru Dua, Devendra Singh Sachan, Michael Boratko, Yi Luan, S´ebastien MR Arnold, Vincent Perot, Siddharth Dalmia, et al. Can long-context language models subsume retrieval, rag, sql, and more? arXiv preprint arXiv:2406.13121,

    arXiv

  5. [10]

    Clevr-math: A dataset for compositional language, visual and mathematical reasoning

    Adam Dahlgren Lindstr ¨om and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. arXiv preprint arXiv:2208.05358,

    arXiv

  6. [12]

    Inter-gps: Interpretable geometry problem solving with formal language and symbolic reasoning

    Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang, Xiaodan Liang, and Song- Chun Zhu. Inter-gps: Interpretable geometry problem solving with formal language and symbolic reasoning. arXiv preprint arXiv:2105.04165,

    arXiv

  7. [13]

    Mathvista: Evaluat- ing mathematical reasoning of foundation models in visual contexts

    17 Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluat- ing mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255,

    arXiv

  8. [14]

    Chen Meiqi, Peng Bo, Zhang Yan, and Lu Chaochao

    URL https:// openreview.net/forum?id=KUNzEQMWU7. Chen Meiqi, Peng Bo, Zhang Yan, and Lu Chaochao. Cello: Causal evaluation of large vision-language models. arXiv preprint arXiv:2406.19131,

    arXiv

Show all 23 references
  1. [15]

    doi: 10.1162/tacl a 00446

    ISSN 2307-387X. doi: 10.1162/tacl a 00446. URL https://doi.org/10.1162/tacl_a_00446. Ankur Parikh, Xuezhi Wang, Sebastian Gehrmann, Manaal Faruqui, Bhuwan Dhingra, Diyi Yang, and Dipanjan Das. ToTTo: A controlled table-to-text generation dataset. In Bonnie Webber, Trevor Cohn,...

    doi

  2. [18]

    Gpqa: A graduate- level google-proof q&a benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate- level google-proof q&a benchmark. arXiv preprint arXiv:2311.12022,

    arXiv

  3. [19]

    Llm processes: Numerical predictive distributions conditioned on natural language

    18 James Requeima, John Bronskill, Dami Choi, Richard E Turner, and David Duvenaud. Llm processes: Numerical predictive distributions conditioned on natural language. arXiv preprint arXiv:2405.12856,

    arXiv

  4. [20]

    Causal evaluation of language models

    Chen Sirui, Peng Bo, Chen Meiqi, Wang Ruiqi, Xu Mengying, Zeng Xingyu, Zhao Rui, Zhao Shengjie, Qiao Yu, and Lu Chaochao. Causal evaluation of language models. arXiv preprint arXiv:2405.00622, 2024a. Chen Sirui, Xu Mengying, Wang Kun, Zeng Xingyu, Zhao Rui, Zhao Shengjie, and ...

    arXiv

  5. [21]

    URL https: //www.kaggle.com/m/3301

    doi: 10.34740/KAGGLE/M/3301. URL https: //www.kaggle.com/m/3301. Avijit Thawani, Jay Pujara, Filip Ilievski, and Pedro Szekely. Representing numbers in nlp: a survey and a vision. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Comput...

    2021 doi

  6. [22]

    Causality for tabular data synthesis: A high-order structure causal bench- mark framework

    Ruibo Tu, Zineb Senane, Lele Cao, Cheng Zhang, Hedvig Kjellstr ¨om, and Gustav Eje Henter. Causality for tabular data synthesis: A high-order structure causal bench- mark framework. arXiv preprint arXiv:2406.08311,

    arXiv

  7. [23]

    Rendering graphs for graph reasoning in multimodal large language models

    Yanbin Wei, Shuai Fu, Weisen Jiang, James T Kwok, and Yu Zhang. Rendering graphs for graph reasoning in multimodal large language models. arXiv preprint arXiv:2402.02130, 2024a. Yanbin Wei, Shuai Fu, Weisen Jiang, Zejian Zhang, Zhixiong Zeng, Qi Wu, James T. Kwok, and Yu Zhang...

    arXiv

  8. [2015]

    doi: 10.3115/v1/P15-1142

    Association for Computational Linguistics. doi: 10.3115/v1/P15-1142. URL https://aclanthology.org/ P15-1142. Judea Pearl, Madelyn Glymour, and Nicholas P Jewell. Causal inference in statistics: A primer. John Wiley & Sons,

    doi

  9. [2017]

    doi: 10.18653/v1/P17-1167

    Association for Computational Linguistics. doi: 10.18653/v1/P17-1167. URL https://aclanthology.org/P17-1167. Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Deven- dra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guil- laume Lample...

    arXiv

  10. [2018]

    Training verifiers to solve math word problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168,

    arXiv

  11. [2020]

    doi: 10.18653/v1/2020.emnlp-main.89

    Association for Computational Linguistics. doi: 10.18653/v1/2020.emnlp-main.89. URL https://aclanthology.org/ 2020.emnlp-main.89. Panupong Pasupat and Percy Liang. Compositional semantic parsing on semi- structured tables. In Chengqing Zong and Michael Strube (eds.), Proceedin...

    2020 doi

  12. [2021]

    neurips.cc/paper_files/paper/2021/file/ 68d30a9594728bc39aa24be94b319d21-Paper-round1.pdf

    URL https://datasets-benchmarks-proceedings. neurips.cc/paper_files/paper/2021/file/ 68d30a9594728bc39aa24be94b319d21-Paper-round1.pdf. Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Jun...

    2021 arXiv

  13. [2022]

    Evaluating large language models on graphs: Performance insights and comparative analysis

    Chang Liu and Bo Wu. Evaluating large language models on graphs: Performance insights and comparative analysis. arXiv preprint: arXiv 2308.11224,

    arXiv

  14. [2023]

    The es- sential role of causality in foundation world models for ‘ied ai

    Tarun Gupta, Wenbo Gong, Chao Ma, Nick Pawlowski, Agrin Hilmkil, Meyer Scetbon, Ade Famoti, Ashley Juan Llorens, Jianfeng Gao, Stefan Bauer, et al. The es- sential role of causality in foundation world models for ‘ied ai. arXiv preprint arXiv:2402.06665,

    arXiv

  15. [2024]

    Think you have solved question answering? try arc, the ai2 reasoning challenge

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457,

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.