{"id":"d8071335-8400-489d-b2da-4c80851142ae","arxiv_id":"2607.11510","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DAG-FM recovers causal DAGs from tabular data by auto-regressive leaf-node and parent-node prediction with a Mixture-of-Leaf-Experts router over four mechanism families, reporting state-of-the-art benchmark results.","lead":"DAG-FM is a foundation model that reconstructs causal graphs from tables of observations by repeatedly predicting the next leaf variable and its parents with two transformer modules. It is designed to work when different variables obey different causal mechanisms, and it scales to datasets that most prior methods cannot fit in memory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.5's identifiability guarantee does not cover the actual prior: Assumption A.4 (finitely many PNL distortions) contradicts Section B.1's continuous warping families, so Corollary A.5's finite-basis construction collapses.","rationale":"The reader's weakest assumption correctly identifies the gap between Theorem 4.5's hypotheses and the prior space constructed in the paper. My reading of Appendix A confirms that A.4 is doing essential work: Corollary A.5 cannot construct the finite basis needed to invoke Lemma A.3 without it, and Section B.1's warping families are continuous, not finite. The paper's own text concedes that A.4 is only 'highly likely' to hold in training, but that is insufficient for a theorem stated about the prior space. The empirical work is substantial and the architecture is credible, so this is a reason to keep the verdict CONDITIONAL rather than reject the paper. The requested concrete test would settle whether the gap is merely a proof-technicality or a substantive failure of the theoretical guarantee; until then, the 'strictly identifiable DAGs' claim should be qualified to the finite-support setting.","tokens_in":36038,"tokens_out":5174,"duration_ms":53489,"concrete_test":"Analytical/numerical check of Corollary A.5 for a continuous PNL family: sample M=100 warps h_i from the Section B.1 sinh-arcsinh family, evaluate h_i'(h_i^{-1}(y)) on a fine grid of y, and compute the numerical rank of the M×grid matrix. If the rank grows with M (or exceeds k+2 for any proposed finite basis), the finite-basis construction b=(1,y,h_1,...,h_k) fails for the actual prior. Equivalently, re-derive Theorem 4.5 with h drawn from an uncountable family and identify the first step in Corollary A.5 that no longer holds; if no finite basis exists, amend the theorem to state support-finiteness explicitly and qualify the 'strictly identifiable' claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is Theorem 4.5: under Assumption 4.3 plus 'some additional identifiability conditions' — footnote 1 points to Assumption A.2(iv) and Assumption A.4 — the posterior satisfies P(G|p(X))=1. Corollary A.5, the step that makes PNL fit the QPE framework, constructs a finite basis b=(1,y,h_1,...,h_k) and requires A.4: only finitely many possible functions h=h'(h^{-1}) within the prior space. But Section B.1 defines PNL warps from continuous families (sinh-arcsinh with continuous epsilon/tau; residual flow with continuous slope/weights; asymmetric power with continuous mu and exponents). The support of the prior over h is uncountable, so A.4 is false for the prior space actually used. The authors' defense — 'training only proceeds for a finite number of steps, only a finite number of post-nonlinear functions h are sampled' — changes the object: finitely many draws from a continuous prior do not make the prior finite. If A.4 is read as applying only to the training sample, then Theorem 4.5 guarantees uniqueness only on that finite realization, not on unseen test SCMs, defeating the foundation-model claim. Assumption A.2(iv) is likewise asserted without verification; but A.4 alone breaks the stated proof. The paper's own caveats (General mechanism 'lacks a rigorous theoretical guarantee'; PNL-HNM identifiability 'remains open') reinforce that the headline 'strictly identifiable DAGs' overstates the theorem's coverage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":36317,"tokens_out":3718,"duration_ms":36646,"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":[{"comment":"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","section":"§4.1, Appendix A (Assumption A.4), Appendix B.1"},{"comment":"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.","section":"§4.1, Theorem 4.5 proof (Appendix A)"},{"comment":"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.","section":"§4.1, Table 4 and Theorem 4.4"},{"comment":"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.","section":"§5, Table 1 and Table 3"}],"minor_comments":[{"comment":"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.","section":"§4.4, Algorithm 1"},{"comment":"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.","section":"§5, Real-World Benchmarks"},{"comment":"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.","section":"Appendix A, after Eq. (7)"},{"comment":"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.","section":"§2, Related Work"},{"comment":"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.","section":"Appendix C, OOD"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's theoretical core relies heavily on a self-cited reference (Chen et al., 2026) for Lemma A.3 and the QPE decomposition. If that reference is not yet publicly available or not accepted, the proof is not self-contained. The paper also cites several 2026 works (e.g., FoundCause, TabCausal, CauScale) that may not be publicly accessible; for a journal submission, the authors should ensure that all compared methods and their checkpoints are identifiable. The A.4 assumption issue is not merely cosmetic: it directly impacts the validity of the main theorem for the actual prior, and the authors should be asked to either revise the prior or the theorem. On balance, the empirical contribution is substantial and the architectural ideas are worth publishing after these theoretical overclaims are resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things up front. First, the empirical core is genuinely useful: DAG-FM is a well-engineered candidate for a single amortized model across FCM families, and the benchmark coverage is unusually thorough. Second, the theoretical guarantee as stated does not hold for the prior the paper actually uses, and this is not a cosmetic issue.\n\nCredit where it's due. The leaf-then-parent decomposition is sensibly borrowed from Dhir et al. and Thompson et al. and correctly credited. The Mixture-of-Leaf-Experts router is the real new piece, and the ablation shows it helps on heterogeneous benchmarks. The prior-space construction is careful — SNR control, post-hoc standardization to avoid gaming — and the scale of experiments (24 baselines, OOD tests, ablations) is impressive. The empirical story mostly holds, though \"significantly outperforms\" outruns the confidence intervals in a few spots (Table 1 F1: 0.67±0.15 vs 0.57±0.16; Table 6: TabCausal beats DAG-FM on LiNGAM F1), and the real-world tables carry no error bars. I also don't see code or checkpoints mentioned, which matters for reproducibility.\n\nThe load-bearing problem is Theorem 4.5. Its hypotheses are hidden in a footnote: Assumption A.4 says only finitely many PNL post-nonlinear functions h are possible in the prior. But Section B.1 samples PNL warps from continuous families — sinh-arcsinh with continuous skew/tail parameters, residual flows with continuous weights, asymmetric power with continuous exponents. The support of h is uncountable, so A.4 is false for the prior as defined. The authors' defense — that training draws finitely many SCMs, so only finitely many h appear — changes the object. A finite realization from a continuous prior is not a finite prior. If A.4 is read as applying only to the training sample, the theorem guarantees uniqueness on those sampled SCMs, not on unseen test SCMs, which is exactly what a foundation model claim needs. The paper's own caveats about the General mechanism and PNL-HNM identifiability being unproven honestly undercut the abstract's \"strictly identifiable DAGs\" framing.\n\nThere are smaller gaps too. The proof of Theorem 4.4 asserts PNL and HNM QPE spaces are non-overlapping based on dimension intuition; a one-dimensional span can intersect a two-dimensional span. Assumption A.2(iv) is stated without verification. All of this is addressable: restate Theorem 4.5 with full hypotheses, either restrict the prior to a genuinely finite family or qualify the convergence claim, and ship code.\n\nWho is this for? People working on amortized causal discovery and tabular foundation models; they'll find the architecture and evaluation useful, but should not take the identifiability guarantee at face value. Does it deserve a serious referee? Yes. It's a substantial paper with real empirical work and honest caveats; it needs major theory revision, not a desk reject.","headline":"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.","tokens_in":37020,"tokens_out":3663,"would_cite":true,"duration_ms":38285,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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","keywords":["causal discovery","foundation model","amortized inference","functional causal models","DAG identifiability","heterogeneous mechanisms","mixture of experts","leaf prediction"],"falsifier":"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.","tokens_in":35707,"feed_emoji":"🕸️","tokens_out":7429,"duration_ms":71452,"temperature":0.7,"pith_summary":"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.","feed_headline":"A foundation model recovers the unique DAG from heterogeneous causal data","feed_subtitle":"Leaf-by-leaf prediction and expert routing identify the exact causal graph, scaling to 1,000 variables.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Foundation model recovers exact DAG from heterogeneous causes","AI model finds unique causal graph under mixed mechanisms","Leaf-by-leaf prediction nails causal DAG at scale","From mixed data to unique DAG: a foundation model","Causal discovery foundation model scales to 1,000 variables"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Foundation model recovers exact DAG from heterogeneous causes","AI model finds unique causal graph under mixed mechanisms","Leaf-by-leaf prediction nails causal DAG at scale","From mixed data to unique DAG: a foundation model","Causal discovery foundation model scales to 1,000 variables"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000561,"raw_usage":{"total_tokens":2500,"prompt_tokens":741,"completion_tokens":1759,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":1681}},"tokens_in":485,"tokens_out":1759,"duration_ms":11284,"temperature":1.0,"reasoning_tokens":1681,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T04:22:22.680673+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":2}