{"id":"c27eac82-cb1d-47e4-822f-548df19e676e","arxiv_id":"2504.14937","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A greedy algorithm, CaGreS, summarizes causal DAGs by contracting nodes while preserving conditional independences, with sound do-calculus and s-separation guarantees.","lead":"The authors propose a method to compress large causal DAGs into small summary graphs that preserve the conditional independence information needed for causal inference. The summary graphs can be used directly for effect estimation and may make complex causal models easier for humans to verify.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Robustness claim rests on unproven premise that summary-induced adjustment sets are harmless; over-adjustment via colliders or descendants of treatment could bias inference.","rationale":"I focused on Theorem 4.1, the s-separation algorithm, and the do-calculus soundness proofs, since those underpin the summary-DAG inference machinery. After tracing the definitions, the canonical DAG G_H is indeed a supergraph of every DAG compatible with H, so the intersection of CIs across compatible DAGs equals the CI set of G_H; Theorem 4.1 and Theorem 4.2 are internally consistent, and the do-calculus soundness argument via Corollary B.7.1 is valid. The genuinely load-bearing weakness is the robustness claim. Example 2 asserts that the larger adjustment sets produced by summarization are harmless, and Section 7 relies on this to argue that summaries mitigate DAG misspecification. No theorem or experiment establishes that backdoor adjustment on H or G_H never selects a descendant of treatment or a collider, both of which are standard sources of bias. This is a real gap because the paper's central contribution, as stated in the abstract and introduction, is that summaries are 'more robust to misspecification.' A conditional acceptance is appropriate: the core theory is sound, but the robustness claim is overreaching relative to the evidence. The reader's weakest assumption identifies exactly this unproven premise, and my independent read agrees.","tokens_in":35934,"tokens_out":30639,"duration_ms":281229,"concrete_test":"Enumerate all DAGs on n<=6 nodes. For each DAG G, each valid summary H obtained by contracting node pairs (with acyclicity maintained) and each ordered pair (T,Y) with a directed path T to Y, compute the set of backdoor-admissible adjustment sets in G and in the canonical DAG G_H (ordering T first within its cluster if clustered). Check whether any admissible set from G_H contains a node that is a descendant of T in G or a collider on a backdoor path in G, while G has an admissible set avoiding such nodes. If such a case exists, simulate data from G and compare ATE estimates using the G_H-derived adjustment set versus the G-derived one; a significant bias would falsify the over-adjustment premise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that summary DAGs are 'more robust to misspecification' depends on the assertion in Example 2 (Section 3.2) that the larger adjustment sets induced by summarization 'will only be ones that do not hurt the analysis.' This premise is not proven. The canonical DAG G_H (Definition 5) adds edges both within clusters (via a topological order) and between clusters (by mapping every H-edge to all cross-product edges). These added edges can create new backdoor paths between a treatment T and outcome Y that did not exist in the original DAG. A do-calculus or backdoor adjustment procedure run on H (or G_H) may therefore select an adjustment set containing a descendant of T or a collider on a proper backdoor path, which is known to induce bias in nonparametric causal inference. Section 7 provides only a qualitative case study with GPT-4-induced errors; it does not bound or characterize the adjustment sets that summarization can produce. In contrast, the do-calculus soundness results (Theorems 6.1-6.2) are carefully argued and appear correct; the soft spot is specifically the robustness/over-adjustment claim, which is load-bearing for the paper's motivation and for the recommendation to use summaries when the input DAG is unreliable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for summarizing causal DAGs by contracting nodes into 'summary causal DAGs,' with the goal of reducing graph size while preserving enough conditional-independence information for sound causal inference. The central theoretical contributions are: (i) a connection between node contraction and edge addition via a 'canonical causal DAG' G_H, leading to an equivalence between the recursive basis of a summary DAG H and that of G_H (Theorem 4.1); (ii) a notion of s-separation with a sound and complete algorithm based on d-separation in G_H (Theorem 4.2); (iii) soundness and completeness claims for do-calculus in summary DAGs (Theorems 6.1 and 6.2); and (iv) an NP-hardness result for the summarization problem (Theorem 3.2). The paper also introduces a greedy algorithm, CaGreS, and evaluates it on six datasets, including a GPT-4-based case study arguing that summaries are robust to misspecification of the input DAG. The overarching claim is that summary causal DAGs can be used directly for causal inference and are more robust to errors in the original DAG.","tokens_in":36194,"tokens_out":16789,"duration_ms":158772,"significance":"If the theoretical results hold, the paper addresses a genuinely useful problem: causal DAGs in high-dimensional settings are hard to verify, and a principled summarization framework could help analysts inspect and reason about them. The core idea of encoding a summary's CI information through a canonical supergraph is natural and the proofs of Theorem 4.1 and the s-separation results appear plausible and are presented in detail. The experiments cover six datasets and include a comparison against several baselines; the authors also ship code and data, which supports reproducibility. However, the advertised robustness property ('more robust to misspecification') rests on unproved over-adjustment claims, and the formal problem definition and NP-hardness proof contain inconsistencies. These issues are central to the paper's framing and need repair. The contribution is potentially useful for the data-management and causal-inference communities, but the current version does not fully support its strongest claims.","major_comments":[{"comment":"The formal definition of the superiority order is reversed relative to the informal explanation and the example. The text before Definition 3 says H1 is superior to H2 if the RB of H2 is implied by the RB of H1, and Example 8 uses H1 ≻ H2 because Σ_RB(H1) implies Σ_RB(H2). However, Definition 3 states that G2 ≻ G1 if for every σ ∈ Σ_RB(G1), Σ_RB(G2) implies σ. Under this formal definition, a DAG with a stronger RB (i.e., fewer CIs, more edges) would be called superior, which contradicts the paper's own goal of preserving CIs and contradicts Example 8. Because the definition of 'maximal' in Problem 1 relies on this order, the formal problem statement is incoherent as written. Please correct the direction of the implication and ensure that Example 8 and the NP-hardness discussion are consistent with the corrected definition.","section":"Section 3.2, Definition 3 and Example 8"},{"comment":"The NP-hardness proof does not establish Theorem 3.2 as stated. The appendix proves (Lemma B.2 and Theorem B.3) that it is NP-hard to decide whether the number of within-cluster non-edges (or, equivalently, a related within-cluster quantity) is below a threshold τ. The proof then argues that since the within-cluster count lower-bounds the total number of added edges |E(G_H)| − |E(G)|, NP-hardness of the total follows. This implication is logically invalid: NP-hardness of a necessary condition for a property does not imply NP-hardness of the property itself. Moreover, the reduction targets the edge-count objective, not the RB-implication maximality objective of Problem 1, so even a correct edge-count hardness result would not directly prove Problem 1 NP-hard. A new proof, or a revised theorem statement that explicitly targets the edge-count formulation as a separate result, is needed.","section":"Section 3.2 and Appendix B.1, Theorem 3.2"},{"comment":"The robustness claim that summarization-induced over-adjustment 'will only be ones that do not hurt the analysis' is asserted without proof and is not implied by the paper's theorems. In the canonical DAG G_H, added edges can create new backdoor paths or cause variables in a treatment's cluster to become descendants of the treatment; standard causal inference shows that adjusting for descendants of treatment or for colliders can induce bias. The do-calculus soundness results (Theorems 6.1 and 6.2) guarantee that if a do-calculus derivation succeeds on H, the corresponding interventional equality holds in every compatible DAG, but they do not constrain which adjustment set a user or automated procedure will select on H. Section 7 provides only a qualitative GPT-4 case study and does not characterize the set of adjustment sets that summaries can produce. Please either provide a formal characterization of when over-adjustment is harmless, or revise the robustness claim to a provable statement (e.g., identifiability on H implies identifiability on every compatible DAG, which is a conservative robustness property rather than a claim that larger adjustment sets are always safe).","section":"Section 3.2 (Example 2) and Section 7"},{"comment":"The suggestion to order the treatment U before all other nodes in its cluster when computing ATE over G_H conflicts with Definition 5, which requires the ordering ⟨X1,...,Xn⟩ to be a topological order of the original DAG G. If U has ancestors (e.g., confounders) within its cluster, a topological order of G must place those ancestors before U; reordering U first can produce a canonical DAG that is not a supergraph of G, in which case E(G) ⊆ E(G_H) fails and Theorem 4.1, Theorem 6.1, and Lemma B.6 no longer apply. This makes the ATE computation procedure unsupported. Please specify how ATE is computed when treatment or outcome is part of a cluster while preserving the supergraph property used in the soundness proofs.","section":"Section 6, 'ATE Computation over Summary DAGs'"},{"comment":"The algorithm description says that a topological order for the nodes of H is established and that the order of nodes within a cluster is 'arbitrary,' but the equivalence and soundness results of Theorem 4.1 and Theorem 4.2 require the within-cluster order to be a topological order of the original DAG G. If the within-cluster order is arbitrary, the canonical DAG G_H may not be a supergraph of every DAG compatible with H, and d-separation over G_H can return CIs that are not valid in H. Concretely, for a cluster {A,B} with an edge to C, ordering A before B yields (A ⊥ C | B) in G_H, but this CI does not hold in a compatible DAG with edges B→A→C. The paper should state that the s-separation algorithm assumes the within-cluster order is inherited from the original DAG's topological order, or should prove that the output is independent of that order under the given definition of compatibility.","section":"Section 4.2.1, s-separation algorithm"}],"minor_comments":[{"comment":"The primary quality metric, 'number of additional edges in the canonical causal DAG,' is exactly the cost function that CaGreS minimizes. Comparisons against baselines on this metric are therefore partly self-referential, even though the C1 overlap experiments and the percentage-of-implied-CIs metric provide independent evidence. Please present the independent metrics as the headline results, or add a statement acknowledging the circularity and justifying why edge count is still a meaningful comparison.","section":"Section 8.3 and 8.4"},{"comment":"The notation G(P) is used in the definition of maximality ('G′ ∈ G(P)') but is never defined. Please define the set of I-Maps or the set of candidate DAGs explicitly.","section":"Section 3.2, Definition 3"},{"comment":"Definition 7 defines s-separation for subsets X,Y,Z ⊆ V(H) but the written condition 'f^{-1}(X) and f^{-1}(Y) are d-separated by f^{-1}(Z)' uses f on cluster nodes; this is fine once f is understood to map cluster nodes to their constituent original nodes, but the notation should be made explicit for readers, especially because f is overloaded in the paper.","section":"Section 4.2, Definition 7"},{"comment":"Theorem 6.2 establishes only the existence of a compatible DAG G′ in which the relevant d-connection holds; it does not explicitly construct a distribution or interventional model in which the do-calculus equality fails. If 'completeness' is intended in the standard sense of do-calculus completeness, the proof should be extended by invoking the completeness of d-separation for DAGs to produce a distribution that is Markov for G′ (and hence for H).","section":"Section 6, Theorem 6.2"},{"comment":"The statement 'resulting in 55 detected edges' followed by the counts 21 correct, 1 inverted, 1 missed, and 33 additional is arithmetically consistent (21+1+33=55), but the sentence structure could be clearer; it should explicitly note that the 55 detected edges include the 33 extra edges and exclude the missed one.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"The paper's central theoretical machinery (canonical DAG, s-separation, do-calculus soundness) is plausible and worth preserving, but the formal problem definition contains a reversed implication, the NP-hardness proof has a logical gap, and the advertised robustness property is not proven. These are substantial but fixable issues. In addition, the relationship to prior work on cluster DAGs (reference [9], Anand et al., AAAI 2023) should be clarified: the paper claims a more streamlined proof of do-calculus soundness for cluster DAGs, but it should state explicitly which results are new beyond that prior work, since the current wording makes the novelty hard to assess."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a real contribution. It gives you a principled problem formulation for summarizing causal DAGs and a solid theoretical core for doing inference on the summaries. It is not, however, a demonstrated robustness tool, and the paper's headline robustness claim is the weakest part.\n\nWhat is actually new: Theorem 4.1 — a summary DAG's recursive basis is equivalent to that of its canonical expanded DAG — is clean, and the proof in the appendix is detailed and goes through. s-separation is a genuine extension of d-separation, sound and complete via the canonical DAG. The do-calculus soundness and completeness results are carefully argued and, as far as I can tell, correct; the authors credit the earlier cluster-DAG work they build on. CaGreS is a sensible greedy heuristic with a well-motivated cost function, code and data are released, and on the independent usability benchmark — causal-effect intervals computed with DoWhy on the summaries, compared against the original DAGs — CaGreS beats the adapted baselines. That result is not self-referential, and it is the paper's best evidence.\n\nThe soft spots, in proportion: the robustness claim in the abstract and Section 7 rests on the assertion in Example 2 that the extra adjustments made by the summary 'will only be ones that do not hurt the analysis.' That sentence is doing a lot of work and is not proven. The canonical DAG adds cross-product edges that can open new backdoor paths, and nothing rules out summary-induced adjustment sets with colliders or descendants of treatment, which bias nonparametric estimates. The stress-test note is somewhat too severe in one respect: the soundness results do show that do-calculus on a summary is conservative for any compatible DAG, so the danger is not that the summary fabricates identifications, but that the robustness story is asserted rather than derived. Still, the gap is real, and Section 7's single GPT-4 case study is qualitative evidence, not a characterization. The other fair concern is circularity: the primary quality metric in Section 8.3 is the number of added edges, which is exactly CaGreS's cost. The independent overlap results mitigate this, but the headline comparisons lean on the circular metric. Minor: no error bars anywhere, and the NP-hardness reduction technically targets minimum added edges rather than the formal maximal-RB objective.\n\nBottom line: this paper deserves a serious referee. The theory and algorithm are worth taking seriously; the robustness claim needs either a real proof or a substantially scaled-down statement. Send it out, and push hard on Section 7 and Example 2. The audience is causal inference, data management, and graph summarization researchers — anyone who wants to work with noisy high-dimensional DAGs.","headline":"Good theory, honest experiments in places, but the headline robustness claim rests on an unproven over-adjustment premise—send to review with a hard push on Section 7.","tokens_in":36710,"tokens_out":11618,"would_cite":true,"duration_ms":100043,"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":"This paper claims that a causal DAG compressed by node contraction can be used directly for causal inference, because its recursive basis of conditional independencies matches that of the expanded canonical DAG.","keywords":["causal DAG summarization","conditional independence","recursive basis","do-calculus","d-separation","s-separation","node contraction","greedy graph summarization"],"falsifier":"Simulate data from a DAG in which merging nodes puts a mediator or collider inside the treatment's cluster, then compute the average treatment effect using the adjustment set that the summary DAG's do-calculus procedure returns; if that estimate differs systematically from the true effect in a way that adjusting for the original DAG would not, the robustness premise fails.","tokens_in":35749,"feed_emoji":"🧩","tokens_out":6371,"duration_ms":59663,"temperature":0.7,"pith_summary":"This paper tries to make high-dimensional causal DAGs usable by compressing them into smaller summary DAGs that still support reliable causal inference. Its central move is to view node contraction as edge addition: expanding a summary's cluster nodes into a canonical DAG adds exactly the edges that the contraction erased. The paper proves that a summary DAG and its canonical expansion have equivalent recursive bases, so they encode the same conditional-independence statements; on this basis it gives a sound and complete s-separation algorithm and sound and complete do-calculus rules for summary DAGs. A greedy algorithm, CaGreS, constructs such summaries by merging the node pair that adds the fewest edges, and experiments on six datasets show the summaries support causal-effect estimates close to those of the full DAG.","feed_headline":"Summarized causal DAGs preserve the information inference needs","feed_subtitle":"Node contraction acts like edge addition, so a small summary still encodes the same conditional independencies as the original DAG.","key_machinery":"The load-bearing object is the canonical causal DAG $G_H$: expand every summary node (cluster) into its constituent original nodes, order them topologically, add edges from earlier to later nodes inside each cluster, and lift cluster edges to all ordered pairs between clusters. The recursive basis $\\Sigma_{RB}$ is the set of at most $n$ conditional independencies stating that each node is independent of its non-descendants given its parents; Theorem 4.1 shows $\\Sigma_{RB}(H) \\equiv \\Sigma_{RB}(G_H)$, which turns node contraction into a precise statement about added edges. This equivalence powers the s-separation algorithm (run d-separation on $G_H$) and the soundness and completeness of do-calculus on summaries, and it motivates CaGreS's merge cost: the number of edges the contraction would add to the canonical DAG.","core_discovery":"The central claim is that a summary causal DAG $H$, built by contracting groups of nodes of a causal DAG $G$, carries enough causal information to be used directly for inference. Formally, if $G_H$ is the canonical causal DAG obtained by expanding each cluster of $H$ into a complete directed acyclic subgraph and inheriting cluster-level edges, then $\\Sigma_{RB}(H) \\equiv \\Sigma_{RB}(G_H)$: the recursive bases are equivalent, so every conditional independence that can be read from one can be read from the other. Because the canonical DAG is a supergraph of the original DAG, any do-calculus derivation that succeeds on $H$ also succeeds on $G$, and the paper shows the three do-calculus rules remain sound and complete over summary DAGs. The paper further shows that contracting nodes is safest when it adds few edges, and packages this into CaGreS, which greedily merges the pair with the smallest canonical-edge cost. If these claims hold, a user can verify a small summary instead of a large DAG and still perform causal estimation.","pith_inferences":["If Theorem 4.1 is right, the same equivalence should let summaries be composed: summarizing a summary along a coarser partition should yield a DAG whose canonical expansion is a supergraph of the previous canonical expansion, so inference remains sound under repeated compression.","The robustness claim suggests a concrete trade-off: larger clusters shrink the graph but enlarge adjustment sets and may hide which compatible DAG generated the data; comparing CaGreS summaries against an oracle that picks the best compatible DAG would quantify that cost.","The contraction-as-edge-addition view points to a natural weighted generalization: assign costs to lost bidirected or directed edges and use the same greedy merge for mixed graphs, a direction the paper's appendix begins."],"forward_implications":["Summary DAGs can be handed directly to do-calculus and backdoor-adjustment routines, so users can reason about interventions without expanding clusters.","Verification cost drops: checking a $k$-node summary replaces inspecting all $n(n-1)/2$ possible edges of the original DAG.","A summary is compatible with many original DAGs, so causal statements made from it hold in every compatible DAG; this is the sense in which summaries are robust to missed or spurious edges in the input.","CaGreS gives a practical route to summaries with $k$ nodes in $O((n-k)n^3)$ time, with caching and low-cost-merge optimizations that preserve quality."],"supporting_citations":[{"why":"Supplies the causal-DAG model, d-separation, and do-calculus whose soundness and completeness the summary framework must preserve.","marker":"[74]"},{"why":"Defines the recursive basis and proves the semi-graphoid and d-separation completeness used in Theorem 4.1.","marker":"[33]"},{"why":"Establishes the logic of causal models underpinning equivalence of conditional-independence sets.","marker":"[32]"},{"why":"Shows the recursive basis and d-separation are sound and complete for inferring conditional independencies, used to transfer RB equivalence to s-separation.","marker":"[107]"},{"why":"Extends do-calculus to clustered causal graphs; the paper streamlines that result via canonical DAGs.","marker":"[9]"},{"why":"Supplies the bottom-up greedy contraction strategy that CaGreS adapts to causal information preservation.","marker":"[34]"},{"why":"Baseline k-Snap general summarization algorithm by node grouping that CaGreS is compared against and improves upon.","marker":"[103]"}],"fun_headline_variants":["Causal DAG summaries preserve inference power","Compress causal DAGs without losing inference","Smaller causal graphs, same causal conclusions","Summarize causal DAGs yet keep do-calculus intact","Causal DAG summarization that stays inferable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that summaries are robust to misspecification assumes that the extra adjustment sets created by merging nodes are always harmless, never a collider or a descendant of treatment that would bias the estimate, and the paper does not prove that do-calculus on a summary DAG cannot select such a harmful set.","fun_headline_variants_meta":{"raw":{"variants":["Causal DAG summaries preserve inference power","Compress causal DAGs without losing inference","Smaller causal graphs, same causal conclusions","Summarize causal DAGs yet keep do-calculus intact","Causal DAG summarization that stays inferable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1300,"prompt_tokens":957,"completion_tokens":343,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":269}},"tokens_in":573,"tokens_out":343,"duration_ms":3132,"temperature":1.0,"reasoning_tokens":269,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:38:13.672089+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data from a DAG in which merging nodes puts a mediator or collider inside the treatment's cluster, then compute the average treatment effect using the adjustment set that the summary DAG's do-calculus procedure returns; if that estimate differs systematically from the true effect in a way that adjusting for the original DAG would not, the robustness premise fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the causal-DAG model, d-separation, and do-calculus whose soundness and completeness the summary framework must preserve."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows the recursive basis and d-separation are sound and complete for inferring conditional independencies, used to transfer RB equivalence to s-separation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline k-Snap general summarization algorithm by node grouping that CaGreS is compared against and improves upon."}],"review_version":1}