{"id":"79c7b13d-c6c3-45c3-8ce0-d87c562574f9","arxiv_id":"2607.22934","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"ABCDEFG provably recovers interventional equivalence classes of causal graphs from large-scale, unknown-target perturbation data by mapping the problem to extended factor graphs.","lead":"Scientists built a Bayesian method that learns cause-effect networks from cell-perturbation data, even when the exact genes targeted by a drug are unknown. It scales to thousands of genes and comes with a proof that the most likely graph it finds matches the true network up to an equivalence class.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix Def B.12 defines the parametric family F_I* without intervention-dependent densities, so Assumption 1 of Theorem 2.5 is unsatisfiable for any nontrivial intervention set; the identifiability theorem is vacuous as stated.","rationale":"The reader identified 'sufficient capacity' as the weakest assumption, and I agree that it is load-bearing. However, the more precise problem is internal: the formal definition of the parametric family in Def B.12 does not actually depend on the intervention index k. The conditionals are written as p(f_enc(U_j ⊙ X)) and p(f_enc(V_i ⊙ Z)) with no intervention input, yet the definition requires these conditionals to differ across interventions for targeted nodes. This makes Assumption 1 contradictory as stated, rather than merely hard to verify for finite architectures. The main text's architecture does include an intervention mask, but the gap between that informal description and the formal definition is exactly where the central identifiability result needs support. If Def B.12 is corrected to include intervention-specific inputs, the theorem may still hold under a suitable capacity assumption, but the paper must also show that the corrected family contains the relevant interventional distributions. This reinforces the reader's CONDITIONAL verdict: the theoretical contribution needs a major revision, but the experimental framework and the acyclicity-by-construction mechanism are not invalidated by this issue. I therefore keep the reader's verdict unchanged rather than moving to reject.","tokens_in":37319,"tokens_out":17854,"duration_ms":164967,"concrete_test":"Independently re-derive Def B.12 with the intervention index k carried through: substitute the paper's mask operation [U_j ⊙ X ; W^j ⊙ I] into f(k)(Z_j|X_pi) and f(k)(X_i|Z_pi), and check whether f(k) ≠ f(1) can hold for i ∈ I_k with a single shared encoder/decoder. If the defining equations are invariant to k, Assumption 1 of Theorem 2.5 is unsatisfiable for any nontrivial intervention set. If the mask is added, demonstrate that the resulting family can represent the interventional conditionals of a simple two-node linear-Gaussian SEM with a soft mean shift on the target node; otherwise the parameterization is too restrictive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix Def B.12 defines the parametric family F_I*(D) as {{f(k)(X,Z): k in [n_I*]} : f(k)(X,Z) = prod_i f(k)(X_i|Z_pi_i^D) prod_j f(k)(Z_j|X_pi_j^D)}, with f(k)(Z_j|X_pi) = p(f_enc(U_j ⊙ X)) and f(k)(X_i|Z_pi) = p(f_enc(V_i ⊙ Z)). The right-hand sides are identical for every k; there is no intervention-specific parameter or input. The definition immediately imposes f(k)(X_i|Z_pi) ≠ f(1)(X_i|Z_pi) iff i ∈ I_k, which is contradictory unless U, V, or the network parameters vary with k. Thus F_I*(G*) contains only the observational conditional family, so Assumption 1 of Theorem 2.5 — {p(k)(X)} ∈ F_I*(G*) — cannot hold for any I* with a nonempty intervention. The main theoretical guarantee therefore does not apply to the formal model as defined. The main text mentions an intervention mask [U_j ⊙ X ; W^j ⊙ I], but this input is absent from Def B.12 and no argument shows that adding a one-hot intervention vector to a shared encoder/decoder yields a family containing arbitrary soft/hard interventional conditionals. This is not merely an unverifiable capacity assumption; as written, the assumption is internally inconsistent and needs correction before Theorem 2.5 supports the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ABCDEFG, a variational Bayesian method for causal discovery from interventional data. It represents causal structures as extended factor DAGs with intervention nodes, uses Gumbel-softmax or sum-product networks to sample edge matrices that are acyclic by construction, and fits a VAE with causal masks for distributional estimation. The authors prove (Theorem 2.5 / B.15) that the argmax of the ELBO-optimal posterior graph distribution is I*-Markov equivalent to the true graph under assumptions of sufficient capacity, I-faithfulness, positivity, and finite entropy, and (Theorem 2.13 / B.22) that the underlying f-DAG is identifiable up to factor permutation under additional structural conditions. Experiments compare ABCDEFG with score-based and Bayesian baselines on simulated data and on a single-cell growth-factor perturbation screen.","tokens_in":37678,"tokens_out":10341,"duration_ms":103696,"significance":"If the theoretical claims were established, the paper would make a substantial contribution: it combines scalable Bayesian uncertainty quantification with acyclicity-by-construction and unknown intervention targets, and its f-DAG identifiability result would extend the factor-graph line of work. The paper also provides broad empirical evaluation and a detailed appendix. However, the main identifiability theorem currently rests on an internally inconsistent definition of the parametric family, and the proof of the f-DAG identifiability lemma is not at the required level of rigor. The central theoretical guarantee therefore needs substantive correction before the paper can be accepted.","major_comments":[{"comment":"Assumption 1 of Theorem 2.5 is unsatisfiable as stated. In Def. B.12 the functions are defined as f^(k)(Z_j|X_pi)=p(f_enc(U_j ⊙ X)) and f^(k)(X_i|Z_pi)=p(f_enc(V_i ⊙ Z)); the right-hand sides do not depend on k. The definition then requires f^(k)(X_i|Z_pi)≠f^(1)(X_i|Z_pi) iff i∈I_k, which is contradictory for any nonempty intervention. Moreover, F_{I*}(D) is a set of joint densities over (X,Z), while the assumption says it contains the marginal interventional distributions p^(k)(X). The intervention mask [U_j⊙X; W^j⊙I] mentioned in §2.5 is absent from Def. B.12, and no argument shows that such a shared encoder yields arbitrary soft/hard interventional conditionals. Consequently Theorem 2.5/B.15 is vacuous as written; the definition of F_{I*} must be corrected and the capacity assumption made precise before the identifiability claim is meaningful.","section":"Appendix B, Def. B.12 / Eq. (14); Thm 2.5, Assumption 1"},{"comment":"The proof constructs a PMF q' by moving ε mass from \\hat G to G*, then argues L(q')>L(q*) contradicts optimality of q*. But the optimization is over q(G;Λ), a restricted parametric family (Gumbel softmax or SPN). The constructed q' is an arbitrary perturbation of q* and is not shown to be representable in that family. Without a representability argument (or a theorem stated for the unconstrained variational family), the contradiction only holds for global optimization over all PMFs, not for the actual ABCDEFG objective. This is a load-bearing gap in the main identifiability proof.","section":"Appendix B.2, proof of Theorem B.15"},{"comment":"The f-DAG identifiability result depends on Lemma B.18, which is stated as a sketch in the main text and whose appendix proof is dense and in places not fully justified. For example, the proof of Prop. B.21 repeatedly asserts the existence of factors/nodes with certain properties and concludes a 'v-structure mismatch' without fully verifying all cases; the final step for type-F1 factors says flipping parent/child 'does not introduce v-structure' without considering interactions among multiple F1 factors. Since Theorem 2.13 is a central claimed contribution, the lemma needs a complete, self-contained proof or a precise statement of which parts remain conjectural. The strong conditions in Def. 2.7/D_m should also be clearly flagged as unverifiable structural assumptions, not a general identifiability guarantee.","section":"Theorem 2.13 / Lemma B.18 and Appendix B.3"}],"minor_comments":[{"comment":"Please align the theoretical definition with the implemented architecture: the one-hot intervention input [W^j⊙I] appears only in the method description, not in the parametric family used in the theorems. Also clarify whether f^(k)(X|G;Φ) denotes the X-marginal of the joint density or the joint density itself; the current notation switches between the two.","section":"Section 2.5 / Def. B.12"},{"comment":"The abstract claims scaling to 'thousands of nodes,' but the largest simulation has 500 nodes and the real-data experiment uses 1,000 highly variable genes. Please either add a thousand-node experiment or soften the scaling claim.","section":"Abstract / Section 3.2"},{"comment":"Several table entries appear garbled (e.g., Table 5 rows for ABCDEFG: '0 0 0120 0 212' and '010 12521 1 26 28 26 22 17 25'). These need to be cleaned so the reported results are readable.","section":"Appendix C, Tables 5-6"},{"comment":"The claim that the unknown-target setting follows by 'replacing the causal DAG with an I-DAG' is not backed by a formal theorem or proof. Please provide a statement or explicitly label this as a heuristic extension.","section":"Section 2.5, unknown-target paragraph"}],"recommendation":"major_revision","confidential_remarks":"The main contribution appears to be the f-DAG identifiability lemma; the DAG-level ELBO identifiability is largely a variational rephrasing of Brouillard et al. The internal inconsistency of Def. B.12 is the most serious issue. If the authors can correct that definition, prove the feasibility of the perturbation argument, and supply a complete proof of Lemma B.18, the paper could become a solid contribution. I would not accept the current version because the central theorems are not supported as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper deserves a look, but the headline theory is broken as written. The stress-test note is correct: Appendix Def B.12 defines the parametric family F_I*(D) using conditional densities f(k)(X_i|Z_pi) = p(fenc(V_i ⊙ Z)) and f(k)(Z_j|X_pi) = p(fenc(U_j ⊙ X)), with no k-dependent parameter or input. The right-hand side is identical for every k, yet the definition asserts f(k) != f(1) iff i ∈ I_k. That is a contradiction. Consequently, Assumption 1 of Theorem 2.5 — that the parametric family contains the ground truth interventional distributions — cannot hold for any nonempty intervention, and the theorem is vacuous as stated. The main text describes a mask [U_j⊙X; W^j⊙I] that would make the family intervention-dependent, but that mask is absent from the formal definition. This is fixable, but it is load-bearing: the central claim is that the MAP estimate identifies the true graph up to I-Markov equivalence, and the current appendix does not support that claim.\n\nWhat is genuinely new and good: the extended factor graph construction with intervention nodes, the acyclicity-by-construction generative process over Y and B, and the SPN parameterization of edge correlations. The ELBO derivation is clear and explicitly reduces to DCDI's score when β→0 and q is a delta, which is a useful framing. The experiments are extensive — 24 simulated settings, up to 1000 nodes, and a real growth-factor screen — and ABCDEFG beats several strong baselines, especially when edges are jointly distributed via SPNs. The appendix is thorough and the limitations section is honest about cycles and low-rank approximations.\n\nSofter spots beyond the Def B.12 issue: Lemma B.18 / Theorem 2.13 rely on the strongly restricted class D_m and a proof that is sketched in the main text and lengthy in the appendix; I did not verify every case, but the reader's skepticism seems fair. The real-data evaluation is recall-only against a random-graph baseline, so the practical claims are weaker than the simulations. No code or data are provided, which makes reproducibility uncertain.\n\nWho this is for: researchers in scalable Bayesian causal discovery and single-cell perturbation analysis. The method is promising and the experimental work is honest, but the theory needs repair. I would send it to peer review rather than desk-reject; a serious referee should demand a corrected parametric family definition and a re-proof (or clear qualification) of the identifiability results under the actual model architecture.","headline":"Promising method and extensive experiments, but the main identifiability theorem is formally vacuous as stated because the parametric family in Def B.12 is not intervention-dependent; fix that before trusting the guarantees.","tokens_in":38199,"tokens_out":3164,"would_cite":false,"duration_ms":34513,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A variational Bayesian model over extended factor graphs can scale causal discovery to thousands of nodes, handle unknown interventions, and recover the true causal graph up to an intervention-specific equivalence class.","keywords":["causal discovery","factor graphs","variational inference","interventional data","unknown intervention targets","acyclicity","identifiability","single-cell perturbation"],"falsifier":"Run ABCDEFG on a ground-truth graph that violates one of the three identifiable-f-DAG conditions in Definition 2.7—for instance, a factor with no unique parent or child—with ample data and a large encoder, and test whether the MAP f-DAG is I-Markov equivalent to the truth up to permutation; failure would refute Theorem 2.13. For Theorem 2.5, construct a small known DAG whose interventional conditionals cannot be represented by the feed-forward encoder, then check whether the MAP graph is still I-Markov equivalent; if not, Assumption 1 is the load-bearing failure.","tokens_in":37119,"feed_emoji":"🧬","tokens_out":5839,"duration_ms":58145,"temperature":0.7,"pith_summary":"ABCDEFG attacks the combination of scalability, acyclicity, unknown intervention targets, and identifiability that no single existing causal-discovery method handles. The proposal is to represent causal structure as an extended factor graph, sample graphs from a generative process that is acyclic by construction, and fit the posterior via a VAE-style ELBO. The paper's central theoretical claim is that the MAP estimate of the optimized posterior is I-Markov equivalent to the true DAG under four standard assumptions, with a further theorem identifying the underlying factor graph up to permutation for a restricted class of f-DAGs. Empirical results show the method outperforming score-based and Bayesian baselines on simulated data and recovering known and novel growth-factor targets in a single-cell screen. A sympathetic reader would care because this is a rare demonstration of a Bayesian causal-discovery method operating at thousand-node scale with interventional data.","feed_headline":"One model finds causal graphs and intervention targets at scale","feed_subtitle":"ABCDEFG guarantees acyclic samples and recovers the true DAG up to an equivalence class—tested on single-cell screens.","key_machinery":"The carrying object is the extended factor graph: a DAG whose vertices are n feature nodes, n_I intervention nodes, and m latent factor nodes, with edges only from nodes to factors or factors to nodes. The causal graph is the half-square node graph, i.e. a Boolean product of node-to-factor and factor-to-node matrices; intervention edges enter through an intervention-to-factor matrix. Acyclicity is guaranteed by ordering factors and assigning each node to a partition between consecutive factors. The proof machinery is the Bayesian score function L(q) = E_q[S_I*(G)] - beta*KL(q||p(G)), which connects the ELBO to the score S_I*(G) used in score-based causal discovery.","core_discovery":"On its own terms, the paper's central discovery is that the ELBO-optimal variational posterior over extended factor graphs has a MAP estimate that is I*-Markov equivalent to the true causal DAG (Theorem 2.5), and that within the identifiable f-DAG class G_m the underlying factor graph is recoverable up to factor permutation (Theorem 2.13). This transfers score-based identifiability into a Bayesian setting and makes acyclicity a construction property rather than a constraint.","pith_inferences":["If the sufficient-capacity assumption is not met by the finite encoder/decoder, the practical guarantee becomes heuristic; a natural test is to measure calibration and graph-recovery error as encoder width grows on a known ground-truth graph.","The identifiable-f-DAG conditions imply the method will be most reliable when biological modules have unique marker genes and clear separations; pathways with overlapping or shared factors would be expected to merge in the recovered factor graph.","Because the framework separates factors as latent modules, the learned factors could serve as hypotheses about pathway structure; one could test this by checking whether factor membership predicts independent perturbation-response measurements not used in training.","The authors note real gene-regulatory networks contain cycles; extending the generative process to cyclic graphs would require relaxing the construction, and the identifiability theorem would need a new equivalence notion."],"forward_implications":["If Theorem 2.5 holds, optimizing the ELBO yields a posterior whose MAP graph is I-Markov equivalent to the true DAG, giving a Bayesian counterpart to score-based identifiability results.","Because graph samples are acyclic by construction, downstream users skip the pruning step that score-based methods need after thresholding.","Jointly modeling unknown intervention targets means the method can infer both the regulatory graph and which genes a drug or growth factor acts through, in one pass.","The SPN variant models correlations among edge probabilities, which the paper shows is necessary for accuracy on SPN-simulated nonlinear data.","On the 31k-cell growth-factor screen, the method recovers known Gene Ontology targets at recall 0.325-0.376 versus 0.196 for random factor graphs, and reconstructs held-out interventions with lower MSE than baselines."],"fun_headline_variants":["Causal discovery scales to thousands of nodes via amortized Bayes","Causal discovery despite unknown intervention targets","Bayesian causal discovery with identifiability guarantees","ABCDEFG: Amortized Bayes for large-scale causal discovery","One Bayesian model: DAG, interventions, and uncertainty"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The chain of guarantees rests on the assumption that the neural encoder/decoder can represent the true interventional distributions exactly, and that the true graph lies in a restricted class of low-rank factorizations whose conditions are not checkable from data.","fun_headline_variants_meta":{"raw":{"variants":["Causal discovery scales to thousands of nodes via amortized Bayes","Causal discovery despite unknown intervention targets","Bayesian causal discovery with identifiability guarantees","ABCDEFG: Amortized Bayes for large-scale causal discovery","One Bayesian model: DAG, interventions, and uncertainty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00087,"raw_usage":{"total_tokens":3576,"prompt_tokens":688,"completion_tokens":2888,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":2820}},"tokens_in":432,"tokens_out":2888,"duration_ms":20660,"temperature":1.0,"reasoning_tokens":2820,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:05:01.659775+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ABCDEFG on a ground-truth graph that violates one of the three identifiable-f-DAG conditions in Definition 2.7—for instance, a factor with no unique parent or child—with ample data and a large encoder, and test whether the MAP f-DAG is I-Markov equivalent to the truth up to permutation; failure would refute Theorem 2.13. For Theorem 2.5, construct a small known DAG whose interventional conditionals cannot be represented by the feed-forward encoder, then check whether the MAP graph is still I-Markov equivalent; if not, Assumption 1 is the load-bearing failure.","supporting_citations":[],"review_version":1}