Pith. sign in

REVIEW 3 major objections 6 minor 51 references

A single pretrained model recovers causal graphs zero-shot across unknown mechanisms by internalizing statistical asymmetries from diverse synthetic structural causal models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A foundation model pretrained on diverse synthetic SCMs performs zero-shot causal graph recovery that beats classical and amortized baselines on broad synthetic suites and transfers modestly to real systems.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Solid amortized causal-discovery system with a clean variational design story; the almost-sure zero-shot guarantee is stronger than the real-data evidence supports. the 3 major comments →

arxiv 2607.11508 v1 pith:5V3LNTV5 submitted 2026-07-13 cs.LG cs.AIstat.ML

CDFM: Towards a General-Purpose Causal Discovery Foundation Model

Jie Qiao , Ruichu Cai , Zijian Li , Weilin Chen , Pengfei Hua , Boyan Xu , Zhengming Chen , Zhifeng Hao
show 1 more author
Peng Cui
This is my paper
classification cs.LG cs.AIstat.ML
keywords causal discoveryfoundation modelstructural causal modelszero-shot inferencevariational inferenceidentifiabilityamortized inferencesynthetic pretraining
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Causal discovery has long forced practitioners to pick a specialized algorithm for each dataset's assumed mechanism—linear non-Gaussian noise, additive noise, and the like—and to re-check those assumptions by hand. This paper argues that a foundation model pretrained on a vast space of synthetic structural causal models can instead perform zero-shot structural inference: given only observational data, it outputs a causal graph without assumption matching or dataset-specific training. The authors first show that causal direction is unidentifiable without mechanism constraints, then prove that exact marginal likelihood over a strictly identifiable mechanism space uniquely recovers the true graph as sample size grows. They turn that marginalization into a tractable variational objective whose three terms—mechanism inference, data reconstruction, and graph decoding—directly design the network. Pretraining on diverse graphs, continuous and discrete mechanisms, latent confounding, and missingness lets the model absorb the statistical footprints that distinguish cause from effect. On large synthetic suites, a physical light-tunnel system, and real bivariate cause-effect pairs, the frozen model outperforms classical and amortized baselines. If the claim holds, the field can move from fragmented, test-driven algorithms to a general-purpose causal discovery model.

Core claim

CDFM recovers causal structure from observational data across heterogeneous, initially unknown mechanisms in a zero-shot setting. Pretraining on a massive, diverse space of identifiable synthetic structural causal models lets the model internalize complex statistical asymmetries, so that maximizing the structural decoder under the variational posterior uniquely recovers the true graph almost surely as sample size grows.

What carries the argument

A variational evidence lower bound that treats unknown causal mechanisms as latent variables and decomposes the intractable marginal likelihood into three modules—mechanism inference Q(M|X), data reconstruction log P(X|G,M), and graph inference log P(G|M). Strict space identifiability of the mechanism catalog guarantees that the exact marginal likelihood isolates the true graph; with enough capacity the ELBO gap vanishes, so the architecture inherits that guarantee (Corollary 1).

Load-bearing premise

The synthetic catalog of causal mechanisms used for pretraining must both cover the processes that generate real data and be strictly identifiable, so no wrong graph can perfectly mimic the true observational distribution.

What would settle it

Freeze the pretrained model and evaluate zero-shot on held-out SCM families whose functional form and noise were deliberately excluded from the pretraining generator; if AUROC collapses toward chance or classical methods that match the true assumption while CDFM fails, the covering and transfer claim is refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners can replace multi-algorithm assumption-checking workflows with one pretrained model plus light preprocessing.
  • Zero-shot recovery improves with sample size and degrades more slowly than baselines as the number of variables grows to 100.
  • The same frozen model transfers to physical systems and heterogeneous real bivariate cause-effect pairs without retraining.
  • Controlled diagnostics show the model respects classical boundaries: near-chance under linear Gaussian SCMs, high accuracy once non-Gaussianity, nonlinearity, or V-structures appear.
  • Structure-aware quantile reconstruction yields competitive missing-value imputation as a free by-product.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Expanding the synthetic catalog with newly discovered identifiable families could grow zero-shot coverage without redesigning the architecture.
  • The paper's closed loop—real-world failures feeding richer generators—suggests causal discovery may evolve more like large-scale pretraining than like hand-crafted algorithm design.
  • The same mechanism-as-latent variational template may transfer to related tasks such as causal effect estimation under unknown mechanisms.
  • Systematic failure on a real domain outside the pretraining taxonomy would yield a concrete map of missing mechanism classes rather than a generic assumption-violation message.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes CDFM, a pretrained foundation model for zero-shot causal structure recovery from observational data. It argues that classical causal discovery is fragmented by mechanism-specific assumptions, and that a single model can instead amortize inference over a broad space of unknown mechanisms M. The theoretical core (Theorems 1–3, Corollary 1) restates that direction is unidentifiable without mechanism constraints, that known identifiable mechanisms yield unique graphs via likelihood, and that exact marginal likelihood over a strictly identifiable mechanism space recovers G* almost surely; a variational ELBO then decomposes learning into mechanism encoding Q(M|X), reconstruction log P(X|G,M), and graph decoding log P(G|M). The architecture (column Set-Transformer embeddings, alternating row/column attention with soft causal masks, quantile imputation, MoE bilinear graph head) is designed to mirror that decomposition and is pretrained on a large synthetic generator spanning multiple graph models and mechanism families (linear/hetero, RFF, PNL, CAM, CPT, discrete ANM, ordinal, physical, time-lag, measurement error, latent confounding, MCAR). Zero-shot synthetic benchmarks (15 families, D up to 100, N up to 4000) show higher AUROC/F1 than AVICI, TabCausal, Arrow, PC, GES, DirectLiNGAM, and NOTEARS; Causal Chamber and Tübingen results are competitive; controlled diagnostics check consistency with classical identifiability (after fine-tuning).

Significance. If the central empirical claim holds—that one frozen model can recover structure across heterogeneous unknown mechanisms better than specialized algorithms without per-dataset assumption matching—this would be a genuine paradigm shift for applied causal discovery and a useful bridge between amortized inference and foundation-model practice. Strengths include an explicit variational design principle linking architecture to the ELBO, a broad multi-family pretraining generator, systematic synthetic scaling tables (sample size, dimension, 15 mechanisms), competitive transfer to Causal Chamber and Tübingen, and controlled checks that the model respects non-identifiable regimes (linear Gaussian) and V-structure/confounding asymmetries. The theory itself is standard asymptotic Bayesian/identifiability reasoning rather than a new theorem, but it usefully organizes the design. Code is promised, which would aid reproducibility.

major comments (3)
  1. [Corollary 1 / §4–5.4 / §6.4] The strongest claim (abstract, Corollary 1, §4–5) is almost-sure recovery of G* for unknown mechanisms via pretraining over a strictly identifiable space M and vanishing variational gap. Section 5.4 only enumerates a finite taxonomy (linear/hetero/RFF/PNL/CAM/CPT/discrete ANM/ordinal/physical/time-lag + measurement error + 15% latent + MCAR). The paper never verifies that this space is free of observationally equivalent (G′,M_G′) mimics, nor that real processes lie inside it. Real-world results (Causal Chamber A1/A2; Tübingen accuracy 0.671 in Table 6) are competitive but small-scale and do not establish membership of the true mechanisms in M. The §6.4 diagnostics explicitly fine-tune rather than evaluate pure zero-shot. The synthetic wins and ELBO architecture therefore do not yet license the almost-sure guarantee for arbitrary real data; the claim should be restated as empirical genera
  2. [§6.2 Tables 1–3 / Figure 6] Primary evaluation (Tables 1–3, Figures 4–6) trains and tests inside the same synthetic generator families. That is useful for controlled scaling, but it is mild circularity of evaluation: strong AUROC/F1 partly reflects prior coverage rather than pure out-of-distribution discovery. The paper should either (i) hold out entire mechanism families or graph models from pretraining and report zero-shot degradation, or (ii) clearly separate “in-family amortized recovery” from the foundation-model claim of general-purpose discovery. Without that, the gap to classical methods that are not trained on the same generator is hard to interpret as pure scientific progress.
  3. [Corollary 1 / Eq. (6) / §5.5] Corollary 1 assumes sufficient capacity so the variational gap vanishes and Q*(M|X) approaches the true posterior, after which the structural decoder recovers G*. The manuscript provides no evidence that the trained model is near this optimum (ELBO gap estimates, posterior calibration, or reconstruction–structure consistency diagnostics on held-out SCMs). Soft causal masks, loss weights (α_margin, γ_F1, β_acyc, α_impute), and the fixed graph-adaptive threshold calibrator are free parameters that can dominate finite-sample behavior. Without such diagnostics, the theory remains an aspirational justification rather than a binding explanation of the reported experiments.
minor comments (6)
  1. [Figure 1 / Introduction] Figure 1 and the paradigm-shift narrative overstate the contrast with classical workflows; many practitioners already use ensembles or assumption checks. Soften the rhetoric so the contribution is the amortized multi-mechanism model, not the elimination of all assumptions.
  2. [§5.2 Eq. (12) / §5.3] Notation for the soft mask S_ij in Eq. (12) uses W_ji and temperature T; later the bilinear head also uses a temperature τ. Clarify that these are distinct parameters and report sensitivity.
  3. [Table 3 / Figure 5] Table 3 and Figure 5 report “best baseline” per mechanism; please also report which baseline wins each cell in the main text or appendix for transparency (the heatmap in Figure 6 helps but is dense).
  4. [§6.3 Table 6] Tübingen evaluation uses 95 pairs and reports accuracy 0.671; state whether the official weighted accuracy protocol and pair weights were used exactly, and whether any pairs were excluded for missingness or non-numeric types.
  5. [Appendix A] Appendix proofs of Theorems 1–3 are standard; a short pointer to Peters et al. / Pearl for the impossibility result and to asymptotic Bayesian model selection would reduce length without loss.
  6. [§2 Related Work] Several arXiv citations (TabCausal, Arrow, Causcale, etc.) are concurrent or very recent; ensure final versions and dates are updated, and that novelty relative to AVICI / CSIvA / CausalPFN is stated more precisely in Related Work.

Circularity Check

2 steps flagged

Mild evaluation circularity: synthetic benchmarks reuse the same 15 mechanism families as the pretraining generator, so strong zero-shot AUROC partly reflects in-family coverage rather than pure out-of-M discovery; ELBO/theory and real-world tests are independent.

specific steps
  1. fitted input called prediction [§5.4 Causal Prior Construction and Pretraining + §6.2 Synthetic Benchmark / Table 3]
    "The core of our data generator is a broad taxonomy of structural causal mechanisms... Continuous additive mechanisms include linear... RFF... post-nonlinear... CAM... Discrete and mixed-type... We construct a comprehensive synthetic evaluation suite spanning 15 distinct structural causal mechanism families... CDFM achieves competitive or superior performance across most nonlinear, discrete, and corrupted observation settings."

    The 15 test families (CAM, CPT, Discrete ANM, Linear, Linear-Hetero, ..., Time-Lag) are identical to the generator taxonomy used for supervised pretraining of (X,G) pairs. The model is therefore fitted on samples from M and then scored on held-out samples from the same M; the high AUROC/F1 is the expected supervised generalization within the training distribution, not an independent prediction of graphs generated by mechanisms outside that space. This is the classic ‘fit then predict closely related quantity’ pattern, albeit at the level of the entire amortized map rather than a single scalar.

  2. fitted input called prediction [§6.4 Empirical Validation of Identifiability]
    "Moreover, to ensure CDFM has been exposed to the full range of diagnostic tasks, we will fine-tune CDFM on these specific tasks. ... To ensure the theoretical boundary is practically detectable, CDFM is fine-tuned on synthetic datasets generated from the same linear Gaussian model family."

    The controlled checks that ‘CDFM behaves consistently with theoretical identifiability boundaries’ are performed after fine-tuning on the exact low-dimensional families being diagnosed (ANM with (b,q), linear-Gaussian V-structures, latent-confounded pairs). The resulting accuracy curves therefore reflect a model that has been re-fitted to the diagnostic distribution, not the frozen zero-shot foundation model whose almost-sure recovery is claimed by Corollary 1. The ‘validation’ is forced by the fine-tuning step.

full rationale

The core derivation (Theorems 1–3 + Corollary 1 + ELBO decomposition into Q(M|X), reconstruction, and graph decoder) is self-contained with explicit proofs in the appendix; it does not redefine the target graph in terms of itself, nor does it import a uniqueness theorem via self-citation. Architecture and losses are a standard variational + supervised amortization of that bound. The only circularity is evaluative and mild: the primary synthetic suite (Tables 1–3, Figs. 4–6) is drawn from exactly the same taxonomy enumerated for pretraining (§5.4), so reported gains measure how well the network has internalized the generator’s asymmetries rather than recovery of mechanisms outside M. Controlled identifiability diagnostics (§6.4) further fine-tune rather than evaluate pure zero-shot. Real-world results (Causal Chamber, Tübingen) and the mathematical statements themselves remain non-circular. Score 3 reflects this limited, non-load-bearing evaluation loop without elevating it to algebraic tautology.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

The central claim rests on classical causal assumptions plus a strong ‘strict mechanism-space identifiability’ premise, uniform priors, architectural inductive biases, and many training/calibration knobs. Invented pieces are the CDFM system and its latent mechanism encoder, not new physical entities. Free parameters are loss weights, soft-mask temperatures, and a synthetic-fitted graph threshold model.

free parameters (5)
  • Loss weights α_margin, γ_F1, β_acyc, α_impute
    Balance BCE, margin, soft-F1, acyclicity, and quantile imputation in L_total (§5.5); chosen for training stability, not derived.
  • Soft causal mask parameters β, τ, T
    Control attention penalties from continuous edge weights (§5.2); hand/architecture choices that shape structure-aware reconstruction.
  • Graph-adaptive threshold calibrator coefficients
    Logit-linear model fitted once on held-out synthetic oracle thresholds using ten graph-level features (§5.3); decision boundary is data-fitted, not theory-fixed.
  • Architecture sizes (E, L, inducing points, MoE experts, quantiles Q)
    Embedding dim, attention depth, inducing counts, expert count, and quantile levels are design hyperparameters affecting capacity and the ELBO approximation.
  • Pretraining mixture rates (e.g., 15% latent confounding, 50% MCAR masking, edge density {1,2,3})
    Generator schedule choices that define the empirical prior over M and G (§5.4); directly shape what ‘generalization’ means.
axioms (6)
  • standard math Without restrictions on structural equations, causal direction is unidentifiable from any joint (Theorem 1 / Functional Representation Lemma).
    Standard causal inference fact used to motivate priors on mechanisms (§4).
  • domain assumption The pretraining mechanism space M is strictly identifiable: no alternative G',M_G' can match P(X|M*,G*) for G'≠G*.
    Load-bearing for Theorems 3 and Corollary 1; not verified for real data, only assumed for the synthetic taxonomy.
  • domain assumption Uniform uninformative priors P(G) and P(M) so maximizing joint/ELBO tracks marginal model selection.
    Stated in §4 ELBO discussion; simplifies theory but is not true of real scientific priors.
  • ad hoc to paper Target graphs are bow-free acyclic directed mixed graphs after marginalizing latents; directed edge wins over bidirected when both present.
    Encoding convention in §3 that restricts the hypothesis class the model is scored against.
  • ad hoc to paper Sufficient model capacity makes the variational gap vanish at the global optimum, so ELBO optimization recovers exact marginal likelihood maximizers.
    Corollary 1 premise; asymptotic idealization not established for the finite network.
  • domain assumption Causal Markov and faithfulness (for V-structure diagnostics) and classical FCM asymmetries (ANM, non-Gaussianity, etc.) as sources of signal in pretraining.
    Implicit throughout mechanism families and §6.4 identifiability checks.
invented entities (3)
  • CDFM architecture (column Set-Transformer embeddings + alternating row/col attention + soft causal mask imputation + MoE bilinear graph head) no independent evidence
    purpose: Instantiate Q(M|X), reconstruction, and P(G|M) for zero-shot structure inference.
    New system composition; pieces borrow from set transformers and amortized discovery but the full stack is paper-specific.
  • Latent neural mechanism variable M encoded by Q(M|X) no independent evidence
    purpose: Treat unknown causal mechanisms as latents so marginalization becomes an ELBO with tractable modules.
    Conceptual/ variational construct; no external measurement of M outside the network’s internal state.
  • Large-scale multi-family SCM pretraining generator as ‘causal knowledge’ prior no independent evidence
    purpose: Cover identifiable asymmetries so the model internalizes statistical footprints of many mechanisms.
    Engineering prior, not a new physical law; quality judged only by downstream benchmarks.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CDFM: Towards a General-Purpose Causal Discovery Foundation Model." pith.science (2026). https://pith.science/paper/5V3LNTV5

@misc{pith2026260711508,
  author       = {Pith},
  title        = {Pith review of: CDFM: Towards a General-Purpose Causal Discovery Foundation Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5V3LNTV5}},
  note         = {Machine review of arXiv:2607.11508}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Causal discovery, the process of recovering underlying causal structures from observational data, is a fundamental pursuit across scientific disciplines. Over the past decades, numerous algorithms have been developed to tackle this challenge through workflows tailored to the specific causal mechanisms underlying each type of dataset, demonstrating effectiveness across a wide range of applications. However, as the volume and heterogeneity of real-world data continue to grow, this dataset-specific approach inevitably leads to a fragmented, test-driven paradigm that struggles to scale to the demands of modern scientific discovery. To address this, we formulate the Causal Discovery Foundation Model (CDFM) as a unified, general-purpose framework for zero-shot structural inference. To ensure reliable generalization across unknown domains, we first investigate the theoretical boundaries of causal identifiability, revealing the indispensable role of causal prior mechanisms in this process. Building on these insights, we formulate a principled variational framework that treats unknown causal mechanisms as latent variables and mathematically decomposes the intractable marginal likelihood into distinct, tractable learning modules. The variational decomposition provides a conceptual design principle for the architecture design of CDFM, while comprehensive causal knowledge guides the large-scale synthesis of our pretraining data. By pretraining on a massive, highly diverse space of synthetic structural causal models, CDFM successfully internalizes complex statistical asymmetries. Extensive experiments demonstrate that CDFM consistently outperforms traditional algorithms, driving a paradigm shift toward a general-purpose causal discovery foundation model.

Figures

Figures reproduced from arXiv: 2607.11508 by Boyan Xu, Jie Qiao, Peng Cui, Pengfei Hua, Ruichu Cai, Weilin Chen, Zhengming Chen, Zhifeng Hao, Zijian Li.

Figure 1
Figure 1. Figure 1: Paradigm shift from traditional causal discovery to CDFM. Instead of selecting algorithms and verifying assumptions for each dataset, CDFM performs lightweight preprocessing and directly infers causal structures using a pretrained foundation model. The model is pretrained on diverse synthetic causal mechanisms and structures, enabling generalized causal reasoning across heterogeneous data distributions. De… view at source ↗
Figure 2
Figure 2. Figure 2: The foundation model paradigm for causal discovery. Causal knowledge guides the generation of diverse synthetic data. Pre-training on this data allows the foundation model to internalize causal asymmetries. Zero-shot inference on real-world data enhances our understanding of unknown mechanisms, which in turn helps us refine and expand the causal knowledge, ultimately driving towards a universally generaliz… view at source ↗
Figure 3
Figure 3. Figure 3: The overall architecture of CDFM. The overview of the architecture of CDFM, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance over sample size and the number of variables. (a) AUROC as a function of sample size N; (b) AUROC as a function of the number of variables D. Results are averaged over all 15 mechanism families. steeper performance degradation due to the combinatorial explosion of conditional independence tests and score evaluations in high dimensions. Performance across causal mechanisms. To further examine wh… view at source ↗
Figure 5
Figure 5. Figure 5: Mechanism robustness. AUROC comparison between CDFM and the best-performing baseline for each of the 15 mechanism families. CDFM wins on all 15 mechanisms in AUROC. Causal Chamber. Causal Chamber contains experimentally validated causal discovery tasks generated from automatically controlled physical systems. We evaluate CDFM on the light-tunnel data using Tasks A1 and A2. For Task A1, we use the observati… view at source ↗
Figure 6
Figure 6. Figure 6: Method × mechanism AUROC heatmap. Rows correspond to methods, columns to mechanism families. CDFM (top row) achieves the highest AUROC across all 15 mechanisms [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Empirical validation of identifiability. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 3 linked inside Pith

  1. [1]

    Annadani, S

    Y. Annadani, S. Bauer, A. Foster, and P. Tigas. Amortized active causal induction with deep reinforcement learning. In Advancesin Neural Information Processing Systems (NeurIPS 2024), 2024

    2024

  2. [2]

    Balazadeh Meresht, H

    V. Balazadeh Meresht, H. Kamkari, V. Thomas, J. Ma, B. Li, J. Cresswell, and R. Krishnan. Causalpfn: Amortized causal effect estimation via in-context learning. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, editors,Advances in Neural Information Processing Systems, volume 38, pages 154945–154984. Curran Associates, Inc., 2025

    2025

  3. [3]

    Bühlmann, J

    P. Bühlmann, J. Peters, and J. Ernest. CAM: Causal additive models, high-dimensional order search and penalized regression. The Annals of Statistics, 42(6):2526–2556, 2014

    2014

  4. [4]

    R. Cai, J. Qiao, K. Zhang, Z. Zhang, and Z. Hao. Causal discovery from discrete data using hidden compact representation. InAdvancesin Neural Information Processing Systems, volume 31, 2018

    2018

  5. [5]

    R. Cai, Z. Zhang, Z. Hao, and M. Winslett. Understanding social causalities behind human action sequences. IEEE Transactions on Neural Networks and Learning Systems, 28(8):1801–1813, 2017

    2017

  6. [6]

    D. M. Chickering. Optimal structure identification with greedy search.Journal of Machine Learning Research, 3:507–554, 2002

    2002

  7. [7]

    A. E. Gamal and Y.-H. Kim.Network Information Theory. Cambridge University Press, Cambridge, U.K., 2011

    2011

  8. [8]

    J. L. Gamella, J. Peters, and P. Bühlmann. Causal chambers as a real-world physical testbed for ai methodology. Nature Machine Intelligence, 7(1):107–118, 2025

    2025

  9. [9]

    Ghysels, J

    E. Ghysels, J. B. Hill, and K. Motegi. Testing for granger causality with mixed frequency data.Journal of Econometrics, 192(1):207–230, 2016

    2016

  10. [10]

    Hollmann, S

    N. Hollmann, S. Müller, L. Purucker, A. Krishnakumar, M. Körfer, S. B. Hoo, R. T. Schirrmeister, and F. Hutter. Accurate predictions on small data with a tabular foundation model. 637(8045):319–326. 19

  11. [11]

    P. O. Hoyer, D. Janzing, J. M. Mooij, J. Peters, and B. Schölkopf. Nonlinear causal discovery with additive noise models. In Advancesin Neural Information Processing Systems, volume 21, pages 689–696, 2009

    2009

  12. [12]

    Immer, C

    A. Immer, C. Schultheiss, J. E. Vogt, B. Schölkopf, P. Bühlmann, and A. Marx. On the identifiability and estimation of causal location-scale noise models. InInternational Conference on Machine Learning, pages 14316– 14332. PMLR, 2023

    2023

  13. [13]

    N. R. Ke, S. Chiappa, J. Wang, A. Goyal, J. Bornschein, M. Rey, T. Weber, M. Botvinick, M. Mozer, and D. J. Rezende. Learning to induce causal structure. InInternational Conference on Learning Representations, 2023

    2023

  14. [14]

    J. Lee, Y. Lee, J. Kim, A. R. Kosiorek, S. Choi, and Y. W. Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. InProceedings of the 36th InternationalConference on MachineLearning, volume 97 ofProceedings of Machine Learning Research, pages 3744–3753. PMLR, 2019

    2019

  15. [15]

    Leiner, B

    J. Leiner, B. Manzo, A. Ramdas, and W. Tansey. Scalable causal structure learning via amortized conditional independence testing. InProceedings of the FourthConference on Causal Learning and Reasoning, volume 275 of Proceedings of Machine Learning Research, pages 174–200. PMLR, 07–09 May 2025

    2025

  16. [16]

    Li, S.-Y

    Z.-R. Li, S.-Y. Liu, T.-Z. Wang, and H.-J. Ye. Tabcausal: Pretraining across causal environments for tabular causal discovery.arXiv preprint arXiv:2605.31156, 2026

    Pith/arXiv arXiv 2026

  17. [17]

    Lorch, S

    L. Lorch, S. Sussex, J. Rothfuss, A. Krause, and B. Schölkopf. Amortized inference for causal structure learning. In Advancesin Neural Information Processing Systems, volume 35, 2022

    2022

  18. [18]

    S. Löwe, D. Madras, R. Zemel, and M. Welling. Amortized causal discovery: Learning to infer causal graphs from time-series data. In B. Schölkopf, C. Uhler, and K. Zhang, editors,Proceedings of the First Conference on Causal Learning and Reasoning, volume 177 ofProceedings of Machine Learning Research, pages 509–525. PMLR, 11–13 Apr 2022

    2022

  19. [19]

    Y. Ma, D. Frauen, E. Javurek, and S. Feuerriegel. Foundation models for causal inference via prior-data fitted networks. InThe FourteenthInternational Conference on Learning Representations, 2026

    2026

  20. [20]

    J. M. Mooij, J. Peters, D. Janzing, J. Zscheischler, and B. Schölkopf. Distinguishing cause from effect using observational data: methods and benchmarks.Journal of Machine Learning Research, 17(32):1–102, 2016

    2016

  21. [21]

    Ni and B

    Y. Ni and B. Mallick. Ordinal causal discovery. InProceedings of the Thirty-EighthConference on Uncertaintyin Artificial Intelligence, volume 180 ofProceedings of Machine Learning Research, pages 1530–1540. PMLR, 2022

    2022

  22. [22]

    J. Pearl. An introduction to causal inference.The international journal of biostatistics, 6(2):7, 2010

    2010

  23. [23]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al. Scikit-learn: Machine learning in Python.Journal of MachineLearning Research, 12:2825–2830, 2011

    2011

  24. [24]

    B. Peng, S. Chen, J. Tian, Y. Qiao, and C. Lu. Causcale: Neural causal discovery at scale.arXiv preprint arXiv:2602.08629, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    Peters, D

    J. Peters, D. Janzing, and B. Schölkopf. Identifying cause and effect on discrete data using additive noise models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, volume 9 of Proceedings of Machine Learning Research, pages 597–604. PMLR, 2010

    2010

  26. [26]

    Peters, D

    J. Peters, D. Janzing, and B. Scholkopf.Elements of causal inference: foundations and learning algorithms. MIT press, 2017

    2017

  27. [27]

    J. Qiao, R. Cai, S. Wu, Y. Xiang, K. Zhang, and Z. Hao. Structural hawkes processes for learning causal structure from discrete-time event sequences. InProceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, 2023

    2023

  28. [28]

    J. Qiao, Y. Xiang, Z. Chen, R. Cai, and Z. Hao. Causal discovery from poisson branching structural causal model using high-order cumulant with path analysis. InProceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI 2024), 2024

    2024

  29. [29]

    J. Qiao, Z. Zeng, R. Cai, Z. Chen, and Z. Hao. On the Identifiability of Poisson Branching Structural Causal Model Under Latent Confounding. InProceedings of the Forty-Third International Conferenceon Machine Learning International Conference on Machine Learning (ICML-2026), 2026. 20

    2026

  30. [30]

    J. Qu, D. Holzmüller, G. Varoquaux, and M. Le Morvan. TabICL: A tabular foundation model for in-context learning on large data. InInternational Conference on Machine Learning, pages 50817–50847. PMLR

  31. [31]

    Robertson, A

    J. Robertson, A. Reuter, S. Guo, N. Hollmann, F. Hutter, and B. Schölkopf. Do-pfn: In-context learning for causal effect estimation. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, editors, Advancesin Neural Information Processing Systems, volume 38, pages 174811–174848. Curran Associates, Inc., 2025

    2025

  32. [32]

    Rubinsteyn and S

    A. Rubinsteyn and S. Feldman. fancyimpute: A variety of matrix completion and imputation algorithms implemented in python. Version 0.7.0

  33. [33]

    Runge, X.-A

    J. Runge, X.-A. Tibau, M. Bruhns, J. Muñoz-Marí, and G. Camps-Valls. The causality for climate competition. In NeurIPS 2019 Competition and Demonstration Track, pages 110–120. Pmlr, 2020

    2019

  34. [34]

    Sachs, O

    K. Sachs, O. Perez, D. Pe’er, D. A. Lauffenburger, and G. P. Nolan. Causal protein-signaling networks derived from multiparameter single-cell data.Science, 308(5721):523–529, 2005

    2005

  35. [35]

    Satopaa, J

    V. Satopaa, J. Albrecht, D. Irwin, and B. Raghavan. Finding a Kneedle in a haystack: Detecting knee points in system behavior. In2011 31st International Conference on Distributed Computing Systems Workshops, pages 166–171. IEEE, 2011

    2011

  36. [36]

    Shimizu, P

    S. Shimizu, P. O. Hoyer, A. Hyvärinen, and A. Kerminen. A linear non-gaussian acyclic model for causal discovery. Journal of Machine Learning Research, 7:2003–2030, 2006

    2003

  37. [37]

    Shimizu, T

    S. Shimizu, T. Inazumi, Y. Sogawa, A. Hyvärinen, Y. Kawahara, T. Washio, P. O. Hoyer, and K. Bollen. DirectLiNGAM: A direct method for learning a linear non-gaussian structural equation model.Journal of Machine Learning Research, 12:1225–1248, 2011

    2011

  38. [38]

    Spirtes, C

    P. Spirtes, C. Glymour, and R. Scheines.Causation, Prediction, and Search. MIT Press, Cambridge, MA, 2 edition, 2000

    2000

  39. [39]

    D. J. Stekhoven and P. Bühlmann. Missforest—non-parametric missing value imputation for mixed-type data. Bioinformatics, 28(1):112–118, 2012

    2012

  40. [40]

    Swelam, L

    O. Swelam, L. Purucker, J. Robertson, H. Raum, J. Boedecker, and F. Hutter. Does tabPFN understand causal structures? In EurIPS 2025 Workshop: AI for Tabular Data, 2025

    2025

  41. [41]

    Thompson, H

    R. Thompson, H. Zhao, D. M. Steinberg, and E. V. Bonilla. Arrow: A foundation model for causal discovery. arXiv preprint arXiv:2605.07204, 2026

    Pith/arXiv arXiv 2026

  42. [42]

    Troyanskaya, M

    O. Troyanskaya, M. Cantor, G. Sherlock, P. Brown, T. Hastie, R. Tibshirani, D. Botstein, and R. B. Altman. Missing value estimation methods for DNA microarrays.Bioinformatics, 17(6):520–525, 2001

    2001

  43. [43]

    van Buuren and K

    S. van Buuren and K. Groothuis-Oudshoorn. mice: Multivariate imputation by chained equations in R.Journal of Statistical Software, 45(3):1–67, 2011

    2011

  44. [44]

    M. Wu, B. Qiang, W. Shi, Y. Song, et al. Sample, estimate, aggregate: A recipe for causal discovery foundation models. Transactions on Machine Learning Research, 2025. Published March 2025

    2025

  45. [45]

    Xiang, J

    Y. Xiang, J. Qiao, Z. Liang, Z. Zeng, R. Cai, and Z. Hao. On the identifiability of poisson branching structural causal model using probability generating function. In Advances in Neural Information Processing Systems, volume 37, pages 11664–11699, 2024

    2024

  46. [46]

    R. Yao, T. Verdonck, and J. Raymaekers. Causal discovery in mixed additive noise models. InProceedings of The 28th International Conference on Artificial Intelligence and Statistics, volume 258 ofProceedings of Machine Learning Research, pages 3088–3096. PMLR, 2025

    2025

  47. [47]

    N. Yin, T. Gao, and Y. Yu. Learning causal graphs at scale: A foundation model approach. 2025

    2025

  48. [48]

    N. Yin, T. Gao, Y. Yu, and Q. Ji. Effective causal discovery under identifiable heteroscedastic noise model. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16486–16494, 2024

    2024

  49. [49]

    Zhang and A

    K. Zhang and A. Hyvärinen. On the identifiability of the post-nonlinear causal model. InProceedings of the Twenty-FifthConference on Uncertainty in Artificial Intelligence, pages 647–655. AUAI Press, 2009. 21

    2009

  50. [50]

    Zhang, G

    X. Zhang, G. Ren, H. Yu, H. Yuan, H. Wang, J. Li, J. Wu, L. Mo, L. Mao, M. Hao, et al. Limix: Unleashing structured-data modeling capability for generalist intelligence.arXiv preprint arXiv:2509.03505, 2025

    arXiv 2025

  51. [51]

    Zheng, B

    X. Zheng, B. Aragam, P. K. Ravikumar, and E. P. Xing. DAGs with NO TEARS: Continuous optimization for structure learning. InAdvancesin Neural Information Processing Systems, volume 31, 2018. 22 A Proofs and Derivation A.1 Derivation of the Variational ELBO To approximate the intractable marginal likelihood where the mechanismM is integrated out, we introd...

    2018

This paper was first reviewed by grok-4.5 on July 14, 2026.