Pith. sign in

REVIEW 4 major objections 5 minor 14 references

DAG-FM: A Foundation Model for Causal Discovery under Heterogeneous Causal Mechanisms

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a foundation model pre-trained on a heterogeneous mixture of LiNGAM, ANM, HNM, and PNL mechanisms converges to a unique, identifiable causal DAG from observational data, and demonstrates the claim with a leaf/parent p

desk verdict Solid empirical foundation-model work whose headline identifiability theorem doesn't actually cover the continuous prior it trains on; send it to referees, but expect a major theory fix. read the letter →

arxiv 2607.11510 v2 pith:FLVQH5D6 submitted 2026-07-13 cs.LG stat.ML

Yikang Chen , Zhengkang Guan , Haoyuan Qian , Xingxuan Zhang , Peng Cui , Yi Yang , Fei Wu , Kun Kuang This is my paper
classification cs.LGstat.ML
keywords causaldiscoveryfoundationmodelamortizedinferencefunctionalmodelsDAGidentifiabilityheterogeneousmechanismsmixtureofexpertsleafprediction
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

The paper tries to establish that amortized causal discovery can be both general and provably correct: instead of assuming one causal mechanism family, it designs a prior over four functional causal model families and proves that the observational distribution pins down the true DAG almost surely. If true, a foundation model trained on this prior will, as the sample size grows, map data to a unique graph rather than a Markov equivalence class. The paper's DAG-FM implements this by decomposing the DAG into leaf-node and parent-node prediction, with a mixture-of-leaf-experts router that identifies which mechanism family applies. On synthetic benchmarks mixing mechanisms and on two real-world datasets, the paper reports that DAG-FM outperforms existing classical and amortized causal discovery methods and scales to thousands of variables.

What carries the argument

The proof machinery is the Quantile Partial Effect (QPE) — the covariate-gradient of the conditional quantile function — which for the four families takes a finite linear-span form with known basis functions: (1) for LiNGAM, (1) for ANM, (1,y) for HNM, and h for PNL. Non-overlap of these spans lets the posterior identify the mechanism family; a Wronskian condition then singles out the unique effect variable, and recursion over leaves identifies the whole DAG. On the architecture side, a Tabular Interaction Block and two bilinear heads implement leaf and parent prediction, and a Mixture-of-Leaf-Experts router selects the family-specific expert.

What would settle it

For a PNL mechanism Y=h(g(X)+U) with h from the paper's continuous warping family (e.g., sinh-arcsinh), compute the Wronskian W_X(s_{X_j}, eta_X) defined in Assumption A.2(iv) at randomly sampled parameter values. If it is ever identically zero, the additional identifiability condition fails and Theorem 4.5 does not cover that part of the prior space; a null result across many samples would support the paper's finiteness-as-practice stance.

Watch

Extended reading notes

Core claim

The central claim is Theorem 4.5: under causal sufficiency, faithfulness, and the requirement that each local mechanism belongs to one of four identifiable families (LiNGAM, additive noise, heteroscedastic noise, post-nonlinear), the posterior probability of the true DAG is one, P(G|p(X))=1, so the DAG is identifiable almost surely from the observational distribution. Theorem 4.4 establishes that the posterior also identifies which mechanism family generated each effect. The authors argue that if the pre-training prior satisfies these conditions, then the amortized model Q_theta trained to maximize log-likelihood converges to the posterior, and hence to the unique graph. The method then cons

Load-bearing premise

The theorem's main identifiability condition assumes only finitely many post-nonlinear warping functions can occur, but the paper's actual prior samples warps continuously, so the guarantee rests on a finiteness approximation that the paper itself marks as 'highly likely' rather than proven.

Editorial extensions

If this is right

  • If Theorem 4.5 holds, the amortized model's MAP estimate converges to the true DAG as the sample size grows, so the method does not stop at a Markov equivalence class.
  • Heterogeneous mechanisms can be mixed inside one prior without breaking identifiability, so a single foundation model can serve datasets whose edges follow different FCM families.
  • The leaf/parent decomposition guarantees a valid DAG at every inference step and avoids quadratic adjacency prediction, enabling the reported scaling to 1,000 variables.
  • The routing probabilities from MoLE surface the inferred mechanism family, giving interpretability and a hook for human priors.
  • On the paper's benchmarks, DAG-FM reports higher precision/recall/F1 and lower structural Hamming distance than classical FCM algorithms and four concurrent amortized models, including on real-world data where several baselines run out of memory.

Reading between the lines

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

  • The theoretical guarantee is narrower than the empirical prior: the paper's continuous warping family violates its own finiteness assumption, so the strict identifiability result applies only to a finite subset of the prior; an unconditional proof would require either discretizing the warping family or extending the QPE framework to infinite families.
  • Because errors in leaf prediction cascade through the recursive pruning, the method's accuracy at high dimensions will depend on the reliability of early leaf choices; the reported scaling results could be stress-tested by measuring order divergence as a function of dimension.
  • The QPE basis-function lens suggests a direct test for adding new mechanism families: check whether their QPE spans lie outside the existing span; if they overlap, identifiability is lost.
  • The MoLE router's family predictions could be used as a deployment diagnostic: high-entropy routing on a new dataset would signal that the data-generating mechanism falls outside the four identifiable families, warning the user not to trust the recovered graph.
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 / 5 minor

Summary. The paper proposes DAG-FM, a two-stage amortized causal-discovery foundation model: a leaf-node predictor and a parent-node predictor, both implemented with a tabular interaction backbone and a Mixture-of-Leaf-Experts (MoLE) head. The training prior is designed to include LiNGAM, ANM, HNM, PNL, and a general mechanism family. The main theoretical claim is that, under Assumption 4.3 plus additional identifiability conditions, the posterior over DAGs is almost surely degenerate, so the amortized model asymptotically recovers the true DAG; the main empirical claim is state-of-the-art DAG recovery on synthetic heterogeneous benchmarks, homogeneous benchmarks, and two real-world datasets, with scaling to large n and d.

Significance. If the theoretical guarantee were valid, this would be an important step: it would give a principled design condition for amortized causal discovery under heterogeneous mechanisms and would place a foundation-model approach on solid identifiability ground. The empirical study is broad: it includes 24 baselines, multiple metrics with confidence intervals on synthetic data, OOD settings, scalability curves, and an ablation of the MoLE component. The architecture is well motivated by the order-then-prune decomposition, and the scalability to 24GB VRAM is practically relevant. However, the central theoretical result currently does not cover the prior space actually used, and some of the empirical superiority claims are not supported by the reported uncertainty. These issues are substantial but potentially addressable in revision.

major comments (4)
  1. [§4.1, Appendix A (Assumption A.4), Appendix B.1] Assumption A.4 states that only finitely many post-nonlinear functions h are possible in the prior space, and Corollary A.5 builds the finite basis b = (1, y, h_1, ..., h_k) on this assumption. Section B.1, however, defines the PNL warping functions as continuous families: sinh-arcsinh with continuous (epsilon, tau), residual flow with continuous weights and base slope, and asymmetric power with continuous mu and exponents. The support of the prior over h is therefore uncountable, so A.4 is false for the prior used for training and testing. The paper's defense that only finitely many h are sampled during a finite training run changes the object of the theorem: if A.4 applies only to the finite training realization, Theorem 4.5 guarantees identifiability only on that realization, not on new test SCMs drawn from the continuous prior. This is a load-bearing gap for the headline claim that D
  2. [§4.1, Theorem 4.5 proof (Appendix A)] The proof of Theorem 4.5 is too sketchy for a central result. In the 'superfluous forward edge' and 'missing forward edge' cases, the argument relies on statements such as 'Markov blanket' and 'no reverse model can yield ... specific directional properties' without a formal derivation from causal sufficiency and faithfulness to the claimed contradictions. The 'reversed edge' case invokes Corollary A.5 but does not precisely show why the existence of two effect variables in a subset contradicts the corollary after conditioning on the rest of the graph. Since Theorem 4.5 is the main theoretical contribution, a rigorous, step-by-step proof is needed. If the proof is essentially that of Lemma A.3 applied recursively, the authors should state that reduction explicitly and verify all conditions after each leaf removal.
  3. [§4.1, Table 4 and Theorem 4.4] Theorem 4.4 claims that the QPE spaces of the four mechanism families are mutually non-overlapping and that therefore P(F | p(X,Y)) = 1. The proof rules out degenerate cases by saying, for example, that for ANM '∇g will never degenerate into a constant vector' because g is non-linear. But the prior in Section B.1 samples random functions from continuous distributions; there may be measure-zero but nonempty sets where the QPE of one family lies in the span of another family's basis (e.g., an HNM with h(x) constant on a region, or a PNL whose h(y) is locally affine). The proof should quantify these exclusions or show that they have prior probability zero so that the posterior identifiability claim holds almost surely. Without this, Theorem 4.4 is not fully established even for the finite-family case.
  4. [§5, Table 1 and Table 3] The abstract and Section 5 claim that DAG-FM 'significantly outperforms' other amortized methods. In Table 1, on the Hetero benchmark, the F1 confidence intervals for DAG-FM (0.67 ± 0.15) and FoundCause (0.57 ± 0.16) overlap substantially; nSHD also overlaps (0.17 ± 0.17 vs 0.23 ± 0.17). Only precision appears non-overlapping (0.71 ± 0.14 vs 0.57 ± 0.22), but that is a weak basis for a general 'significantly outperforms' claim. Table 3 reports real-world results without any error bars, making it impossible to assess the significance of the reported improvement on Sachs and Causal Chamber. The authors should add error bars to Table 3 (e.g., bootstrap over data subsets or multiple restarts) and apply a statistical test or at least report overlap-aware intervals for Table 1. If the claim is meant only as 'best point estimates', it should be phrased that way.
minor comments (5)
  1. [§4.4, Algorithm 1] The threshold alpha is fixed at 0.5 in the text but appears as a free parameter in Algorithm 1. If alpha is not tuned, state so explicitly; if it is, report sensitivity analysis.
  2. [§5, Real-World Benchmarks] The caption of Table 3 says 'compared with other amortized algorithms' but the table includes only four methods; the full comparison in Table 12 has many 'Out of Memory' entries. A footnote explaining the OOM conditions (e.g., GPU memory limit) would help.
  3. [Appendix A, after Eq. (7)] The QPE basis for PNL in Table 4 is written as h = h'(h^{-1}), but the notation is not defined at that point; define h, h', and h^{-1} explicitly.
  4. [§2, Related Work] The reference 'FoundCause (Blöbaum et al., 2026)' is cited without a publication venue. If it is a preprint, say so; if it is not available, the empirical comparison should be reproducible from a public checkpoint, which the paper does state.
  5. [Appendix C, OOD] In Figure 7, the 'Mechanism' OOD case uses PNL-HNM, but the theoretical identifiability of PNL-HNM is described as open. The text should clearly separate the empirical OOD result from the theoretical guarantee.

Circularity Check

2 steps flagged · score 4.0 of 10

Theorem 4.5's identifiability guarantee rests on a uniqueness lemma imported from the authors' own prior work and on Assumption A.4, which is made to hold by redefining the prior as the finite training realization; the empirical SOTA claims on external benchmarks remain independently grounded.

  1. uniqueness imported from authors [Appendix A, Lemma A.3 and Corollary A.5; footnote 1 and Theorem 4.5 proof]
    "Lemma A.3 (Chen et al. 2026, Corollary 3.8). If Assumption A.2 holds, then the variable Xi satisfying condition (iii) in Assumption A.2 is the unique effect variable. ... The proof is then completed by applying Lemma A.3."

    Theorem 4.5's conclusion P(G|p(X)) = 1 is obtained by Corollary A.5, whose proof says 'The proof is then completed by applying Lemma A.3.' Lemma A.3 is a uniqueness theorem cited to Chen et al. 2026, whose first author, Yikang Chen, is also the first author of this paper. The lemma is not proved here, and the 'additional identifiability conditions' in footnote 1 point exactly to the assumptions needed to invoke it. Thus the central theoretical guarantee is a self-citation chain: the authors' prior uniqueness theorem is used to declare the DAG forced, rather than being derived independently in this paper. The empirical benchmarks are external, so the circularity is confined to the theoretical layer.

  2. self definitional [Assumption A.4 versus Section B.1 (Post-processing / PNL warps)]
    "Assumption A.4. For the post-nonlinear function h in PNL, assume that there is only a finite number of possible functions h = h'(h^{-1}) within the prior space. ... Since the training only proceeds for a finite number of steps, only a finite number of post-nonlinear functions h are sampled from the PNL mechanisms. ... Sinh-Arcsinh: ... skewness parameter ε, sampled continuously from [−1.0,1.0] ... tail-weight parameter τ, sampled continuously from [0.5,2.0]."

    Corollary A.5's finite basis b=(1,y,h1,...,hk) requires Assumption A.4, but the prior space used in Section B.1 samples PNL warps from continuous families, so its support is uncountable. The paper's justification of A.4 counts only the finite number of SCMs realized during training and 'consequently' treats the prior as finite. This changes the object: the theorem's guarantee applies to the finite set of already-sampled h, not to the continuous prior from which unseen test SCMs are drawn. The claim that the actual prior space guarantees identifiability is then equivalent to the training-realization assumption, i.e., the conclusion is made to fit the QPE finite-basis framework by defining the prior to be finite.

full rationale

Most of the empirical content is independent: the synthetic benchmarks are held-out SCMs sampled from the same prior (a standard in-distribution held-out evaluation), and Sachs/Causal Chamber are external real-world datasets; those results are not fitted parameters renamed as predictions. The circularity is confined to the theoretical guarantee. Theorem 4.5 depends on Corollary A.5, whose proof is completed by Lemma A.3, quoted as 'Chen et al. 2026, Corollary 3.8' — a uniqueness lemma from the same author group, invoked as an external mathematical fact. In parallel, Assumption A.4 is needed to construct the finite QPE basis, but the actual PNL prior is continuous; the paper's defense that only finitely many functions are realized during training redefines the quantified 'prior space' and therefore makes the identifiability guarantee apply to a finite training realization rather than the generative prior. These are load-bearing definitional/citation issues in theory, not in the externally benchmarked empirical claims. The paper also honestly flags that the 'General' mechanism lacks a rigorous guarantee, further narrowing the theoretical coverage. Overall: some self-citation and a definitional gap, but the central empirical result has independent content.

Assumptions & free parameters 6 free parameters · 6 assumptions · 3 invented entities

The empirical claim rests on held-out synthetic and real-world benchmarks; the theoretical claim rests on the QPE framework (self-cited), the standard FCM identifiability literature, and two paper-specific assumptions (A.2(iv), A.4) that are unverified. Most free parameters are prior-space and inference choices rather than fitted constants; the most consequential hand-set value is the parent threshold α=0.5, which directly controls graph sparsity at inference.

free parameters (6)
  • parent threshold α = 0.5
    Inference threshold in Algorithm 1 accepting an edge prediction; set by hand with no sensitivity analysis; directly controls precision/recall at inference.
  • MoLE top-k = k=1
    Router keeps one expert's prediction for the final leaf probability; the weighting scheme ('setting k=1') is not fully specified.
  • training table size n = 1024
    All pre-training tables have n=1024 rows; zero-shot claims across other sample sizes rest on generalization the paper does not analyze.
  • mechanism mixture probabilities = uniform over the six train families
    Equal-probability mechanism sampling is a design choice; Section D.2 shows performance degrades once the General mechanism exceeds 0.5 probability, so the uniform choice is load-bearing for reported averages.
  • SNR range for noise injection = [-5, 10] dB uniform
    Prior-space calibration choice balancing low- and high-noise regimes.
  • expected graph density and warp/discretization fractions = 0.25 (truncated exponential on [0,1])
    Hand-set prior hyperparameters biasing toward sparse graphs and moderate distortion; not justified by theory.
assumptions (6)
  • domain assumption Causal sufficiency and faithfulness (Assumption 4.3(i))
    Assumed for all training and test data; the future-work section concedes latent confounders are not handled, and real-world datasets such as Sachs are contested on exactly this ground.
  • domain assumption QPE identifiability machinery of Chen et al. (2026), imported wholesale (Table 4; Lemma A.3 = Chen et al. Cor. 3.8)
    The main theorems reduce to this self-cited reference with overlapping authorship; the derivations of the QPE-basis table are not reproduced.
  • ad hoc to paper Assumption A.2(iv): Wronskian determinants W(s,η) ≠ 0 for relevant variable pairs
    Genericity condition introduced for this analysis; never verified for the actual prior; Theorem 4.5's main-text statement hides it in footnote 1.
  • ad hoc to paper Assumption A.4: finitely many PNL functions h in the prior space
    Crafted so PNL fits the finite-basis QPE framework; the authors admit it only 'highly likely' holds because training samples finitely many SCMs; it is false for the continuous warping families in Section B.1.
  • domain assumption Mutual non-overlap of the four FCM families' QPE spaces (Theorem 4.4)
    Argued via 'will never' statements (e.g., non-linear g with non-constant gradient); the literal reading is false for piecewise-linear g; requires generic-position conditions from the FCM identifiability literature.
  • standard math Density regularity: densities exist, strictly positive, at least C^k
    Stated at the start of Appendix A before any analysis.
invented entities (3)
  • General causal mechanism f(X,U) with U as input dimension
    purpose: Broadens prior coverage beyond the four theory-backed families; used in training and the 'General' benchmark.
    A construction, not a postulated physical entity; the paper concedes it 'lacks a rigorous theoretical guarantee for strict DAG identifiability' (Section 4.2).
  • PNL-HNM mechanism Y=h(a(X)+b(X)U)
    purpose: Exclusively an OOD test mechanism (Section C.3).
    Invented test-only mechanism; the paper states its identifiability 'remains open', so OOD results on it have no theoretical handle.
  • Dummy expert in MoLE
    purpose: Absorbs source nodes and unseen cases in MoLE routing.
    Design component; it can mask routing failures, weakening the stated interpretability benefit.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DAG-FM: A Foundation Model for Causal Discovery under Heterogeneous Causal Mechanisms." pith.science (2026). https://pith.science/paper/FLVQH5D6

@misc{pith2026260711510,
  author       = {Pith},
  title        = {Pith review of: DAG-FM: A Foundation Model for Causal Discovery under Heterogeneous Causal Mechanisms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FLVQH5D6}},
  note         = {Machine review of arXiv:2607.11510}
}
read the original abstract

Causal discovery from observational tabular data remains fundamentally challenging, primarily due to the heterogeneity of underlying causal mechanisms and the high-dimensional combinatorial search space of Directed Acyclic Graphs (DAGs). In this paper, we propose \textbf{DAG-FM}, a novel foundation model architecture that amortizes causal discovery. Unlike direct matrix prediction, DAG-FM decomposes the causal discovery process into two auto-regressive stages using two specialized Transformer-based sub-modules: a leaf-node predictor and a parent-node predictor. To effectively model complex row-column interactions, we adopt a robust tabular interaction block to output feature-wise representations. Crucially, to handle diverse and unknown Functional Causal Model (FCM) assumptions in real-world scenarios, we introduce Mixture-of-Leaf-Experts (MoLE), allowing the model to dynamically route and adapt to identifiable mechanism families. Through an iterative inference algorithm, DAG-FM seamlessly extracts causal orderings and constructs valid DAGs. Extensive experiments demonstrate that DAG-FM achieves state-of-the-art performance on both synthetic benchmarks and complex real-world datasets, significantly outperforming traditional classical algorithms and recent foundation models in both accuracy and scalability.

Figures

Figures reproduced from arXiv: 2607.11510 by the authors.

Figure 1
Figure 1. Architecture of DAG-FM. (a) Tabular Interaction Block; (b) Leaf Prediction Head; (c) [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 1
Figure 1. Architecture and inference Procedure of DAG-FM. (a) Tabular Interaction Block; (b) Leaf [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figures from the paper (13 more)
Figure 2
Figure 2. Figure 2: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p021_2.png]
Figure 3
Figure 3. Figure 3: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p015_3.png]
Figure 3
Figure 3. Figure 3: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p022_3.png]
Figure 4
Figure 4. Figure 4: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 4
Figure 4. Figure 4: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]
Figure 5
Figure 5. Figure 5: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 5
Figure 5. Figure 5: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]
Figure 6
Figure 6. Figure 6: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 6
Figure 6. Figure 6: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 7
Figure 7. Figure 7: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]
Figure 8
Figure 8. Figure 8: Performance of DAG identification (Precision, F1, nSHD) and order identification (ODR) [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]
Figure 9
Figure 9. Figure 9: Performance of DAG identification (Precision, F1, nSHD) under non-uniform mechanism [PITH_FULL_IMAGE:figures/full_fig_p026_9.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 1 canonical work pages

  1. [1]

    Post-processing

    We then sample 1D random noise variables i.i.d. from these randomly instantiated distributions. Subsequently, these noise samples are normalized and participate in the forward inference of the random causal mechanisms with controlled signal-to-noise ratios (SNR). Random MechanismsThe building blocks of various causal mechanisms are randomly sampled functi...

    2025

  2. [7]

    with respect to the given basis functionsb, it can be shown thatW Xi (si,j,η i)≡0 for anyX j ∈X −i and anyx(Chen et al., 2026, Theorem 3.6 and Corollary 3.8). Assumption A.2 couples this finite linear span form with the cause-and-effect relationship, and rules out coincidences where this necessary condition also identically holds for other variablesX k ∈X...

    2026

  3. [20]

    Except for the specified mechanisms, all other configurations are identical to the prior space construction during training. When evaluating specific properties, we vary only the target hyperparameter during sampling (e.g., varying only sample sizenwhen evaluating sensitivity to sample size), while keeping all other settings identical toHetero. Real-World...

    2005

  4. [1992]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pp. 4171–4186,

    2019

  5. [2002]

    0.23 0.45 0.31 0.18 0.76 0.500.620.55 0.38 0.19 ICA-LiNGAM (Shimizu et al., 2006)0.20 0.45 0.28 0.16 0.85 0.05 0.05 0.05 0.03 0.41Direct-LiNGAM (Shimizu et al., 2011)0.19 0.35 0.25 0.14 0.76 0.05 0.05 0.05 0.03 0.40 CAM (B¨uhlmann et al.,

    2006

  6. [2005]

    Shohei Shimizu, Patrik O

    doi: 10.1126/science.1105809. Shohei Shimizu, Patrik O. Hoyer, Aapo Hyv ¨arinen, and Antti Kerminen. A linear non-gaussian acyclic model for causal discovery.Journal of Machine Learning Research, 7(72):2003–2030,

    doi:10.1126/science.1105809 2003

  7. [2013]

    doi: https://doi.org/10.1016/j.cell.2013.03.030

    ISSN 0092-8674. doi: https://doi.org/10.1016/j.cell.2013.03.030. Jiaru Zhang, Rui Ding, Qiang Fu, Huang Bojun, zizhen Deng, Yang Hua, Haibing Guan, Shi Han, and Dongmei Zhang. Learning identifiable structures avoids bias in DNN-based supervised causal learning. InThe 28th International Conference on Artificial Intelligence and Statistics, 2025a. Kun Zhang...

    doi:10.1016/j.cell.2013.03.030 2013

  8. [2016]

    doi: 10.1093/ije/dyv341

    ISSN 0300-5771. doi: 10.1093/ije/dyv341. 12 DAG-FM: A Foundation Model for Causal Discovery under Heterogeneous Causal Mechanisms Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwan...

    doi:10.1093/ije/dyv341

Show all 14 references
  1. [2018]

    13 DAG-FM: A Foundation Model for Causal Discovery under Heterogeneous Causal Mechanisms A Proofs We provide the proofs for the theories in the main text under the identifiability framework based on Quantile Partial Effects (QPE) (Chen et al., 2026). This framework decouples t...

    2026

  2. [2019]

    Out of Memory Out of MemoryDAGMA-Linear (Bello et al., 2022)0.20 0.10 0.13 0.07 0.47 0.04 0.03 0.03 0.02 0.32DAGMA-Nonlinear (Bello et al., 2022)0.25 0.20 0.22 0.12 0.51 Out of Memory A VICI (Lorch et al.,

    2022

  3. [2020]

    0.10 0.10 0.10 0.05 0.65 0.07 0.05 0.06 0.03 0.34GraNDAG (Lachapelle et al., 2020)0.00 0.00 0.00 0.00 0.40 1.000.03 0.05 0.03 0.20DAG-GNN (Yu et al.,

    2020

  4. [2022]

    doi: 10.3390/jmmp6010010

    ISSN 2504-4494. doi: 10.3390/jmmp6010010. Johnny Xi, Hugh Dance, Peter Orbanz, and Benjamin Bloem-Reddy. Distinguishing cause from ef- fect with causal velocity models. InForty-second International Conference on Machine Learning,

    doi

  5. [2025]

    0.23 0.50 0.31 0.19 0.80 0.07 0.13 0.09 0.05 0.53 NOTEARS-Linear (Zheng et al., 2018)0.33 0.10 0.15 0.08 0.40 0.15 0.05 0.08 0.04 0.25NOTEARS-Nonlinear (Zheng et al., 2018)0.07 0.05 0.06 0.03 0.58 0.18 0.10 0.13 0.07 0.28GOLEM (Ng et al.,

    2018

  6. [2026]

    Mooij, Dominik Janzing, and Bernhard Sch ¨olkopf

    Jonas Peters, Joris M. Mooij, Dominik Janzing, and Bernhard Sch ¨olkopf. Causal discovery with continuous additive noise models.Journal of Machine Learning Research, 15(58):2009–2053,

    2009

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.