{"id":"01255712-e72e-4197-ae86-002a7afa4466","arxiv_id":"2507.00440","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper proposes a contrastive-learning-based causal graph regression framework that explicitly models the predictive power of confounding subgraphs and achieves state-of-the-art OOD generalization on graph regression benchmarks.","lead":"This paper builds a method for graph regression that separates a graph into causal and confounding parts and uses contrastive learning to make the causal part robust to distribution shift. It reports large accuracy gains on two out-of-distribution benchmarks (GOOD-ZINC, ReactionOOD), suggesting a general recipe for such tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The causal-intervention claim rests on Eq. (16)'s unvalidated embedding additivity: a GNN representation of C∪S is not generally Hc+Hs, and InfoNCE alignment with such sums does not by itself implement backdoor adjustment; the OOD gains may come from a generic regularizer.","rationale":"The reader's weakest assumption is exactly the one I find most load-bearing: the embedding-space counterfactual has no formal basis. I agree with CONDITIONAL because the empirical results are strong and the code repository is promised, but the causal interpretation is not established. The proposed check is an invariance test: if the learned causal representation is truly invariant to confounders, swapping S must not change predictions. If it does, the method's success cannot be attributed to backdoor adjustment. I do not see an internal logical contradiction that would require rejection; the concern is about missing support for a central mechanism, so the verdict should remain CONDITIONAL/UNCHANGED.","tokens_in":17587,"tokens_out":9093,"duration_ms":106495,"concrete_test":"Freeze a trained model on a synthetic graph regression dataset with known causal and confounding motifs. For each test graph i, keep its causal subgraph C_i fixed but replace its confounding subgraph S_i with S_j from another graph j (using hard masks derived from the trained attention), and measure the change in the causal readout prediction. If Eq. (17) has implemented backdoor adjustment, predictions should be invariant to the confounder replacement; if predictions shift by more than a small tolerance (e.g., the model's test RMSE), then LCI is not enforcing confounder invariance. Repeat on GOOD-ZINC by permuting Hs embeddings across examples in a minibatch at test time and comparing predicted y before/after permutation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed framework improves OOD generalization by (i) using confounders' predictive power and (ii) performing causal intervention via contrastive learning. The load-bearing step is Eq. (16): Hmix,ij = Hc,i + Hs,j is asserted to be a counterfactual representation of the graph obtained by replacing the confounder of graph i with that of graph j. For a GNN encoder, the embedding of the union of two subgraphs is f(A⊙M, X⊙M), not f(C)+f(S); representation addition is not justified. The paper itself only claims this 'can be understood as an implicit realization of backdoor adjustment' (Section 4.3), and no derivation or validation is provided. Backdoor adjustment requires stratifying or averaging over S, whereas Eq. (17) is an instance-discrimination objective that aligns Hg,i with one random Hc,i+Hs,j. It enforces a form of invariance, not the interventional distribution. Moreover, if S is predictive (the paper's premise), a representation that is invariant to S may discard predictive signal, which conflicts with the -βI(S;Y) term. If Eq. (16)/(17) is not actually implementing intervention, the stated causal mechanism is unsupported; the empirical gains could come from a generic contrastive regularizer and the 'causal graph regression' interpretation would not be established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies causal graph regression (CGR), i.e., graph-level regression under out-of-distribution shifts, and proposes a framework that (i) extends the graph information bottleneck objective by explicitly modeling the predictive power of confounding subgraphs through an additive term -βI(S;Y), and (ii) introduces a contrastive intervention loss that aligns an original graph representation with a representation formed by adding the causal embedding of one graph to the confounding embedding of another graph. The authors argue that this implements an implicit form of backdoor adjustment in representation space, and they report state-of-the-art OOD performance on the GOOD-ZINC benchmark and competitive results on ReactionOOD S-OOD datasets. The paper includes ablations and a parameter sensitivity analysis, and the code is publicly released.","tokens_in":17849,"tokens_out":8592,"duration_ms":91316,"significance":"If the central claims hold, the paper makes a useful contribution by adapting causal graph learning ideas to regression, a setting where label-based intervention techniques do not apply. The variational bounds for I(C;G) are standard, and the reduction of the mutual information terms to MSE under a constant-variance Gaussian assumption is internally consistent (apart from a presentation issue in Eqs. (13)-(14)). The empirical results are strong: the proposed method outperforms all baselines on all eight GOOD-ZINC settings and is best or second-best in most ReactionOOD settings. The ablations (Figure 4 and Appendix A.5) show that both the enhanced GIB loss and the contrastive intervention loss contribute to the gains. The code release is a definite strength. The main weakness is that the causal intervention mechanism, which is central to the paper's framing, rests on an unvalidated assumption that embedding addition produces a valid counterfactual representation; without a derivation or direct validation, the OOD improvement could be attributed to a generic contrastive regularizer.","major_comments":[{"comment":"The counterfactual representation Hmix,ij = Hc,i + Hs,j is asserted without justification. For a GNN encoder, the embedding of a graph with subgraphs C and S is f(A⊙M, X⊙M), which is not generally equal to f(C) + f(S). The paper states that this addition 'can be understood as an implicit realization of backdoor adjustment', but no derivation or empirical validation is provided. Since the causal intervention claim is load-bearing for the paper's title and contributions, the authors should either (a) provide a theoretical condition under which GNN embeddings are additive across subgraphs, or (b) run an ablation that isolates the additive counterfactual assumption, e.g., replacing Hmix with a non-additive mixing operation (such as concatenation or a learned combination) or with Hc,i plus random noise. Without such evidence, the OOD gains can be explained by a generic contrastive regularizer that encourages invariance to confounders, and the paper's causal graph regression interpretation is not established.","section":"Section 4.3, Eq. (16)"},{"comment":"The paper claims to be the first to explicitly consider the predictive role of confounding features in graph regression, but the experiments do not include any baseline that adapts existing causal graph learning methods (e.g., CAL or DisC) to regression. The introduction states that vanilla adaptation of CGL to regression is dwarfed by ERM, yet no such adaptation appears in the comparison tables. Without this baseline, it is unclear whether the observed improvements come from the proposed enhanced GIB objective, from the contrastive intervention, or simply from the architectural choices shared with the baselines. Adding an adapted CAL-style baseline, or at least an ablation that removes only the -βI(S;Y) term while retaining the rest of the framework, would strengthen the claim that modeling confounder predictiveness is the key novelty.","section":"Section 5.2, Tables 1-3"}],"minor_comments":[{"comment":"The derivation of Eq. (14) is presented in a way that confuses mutual information with conditional entropy. Eq. (13) approximates H(Y|C), not I(C;Y). Since I(C;Y) = H(Y) - H(Y|C) and H(Y) is constant, maximizing I(C;Y) is equivalent to minimizing H(Y|C), which is approximately the squared error. The text should say 'the objective for maximizing I(C;Y) reduces to minimizing the least-squares loss' rather than 'I(C;Y) reduces to the least-squares loss' and should not write LCP as a negative MSE without clarifying the target is to minimize -LCP.","section":"Section 4.2, Eq. (14)"},{"comment":"The InfoNCE formula in Eq. (17) is missing a closing parenthesis in the denominator; the intended expression is -1/B Σ_i log [ exp(sim(Hg,i, Hmix,ij)) / Σ_{k≠i} exp(sim(Hg,i, Hg,k)) ]. The notation should be fixed for clarity.","section":"Section 4.3, Eq. (17)"},{"comment":"The subsection title contains a typo: 'Datsets' should be 'Datasets'.","section":"Section 5.1, title"},{"comment":"When deriving the upper bound for I(C;G), the notation µϕ(G) is used both for the mean of the conditional distribution p(C|G) and, in Eq. (8), as the quantity whose squared norm is averaged. Since the covariance is assumed to be the identity, this is acceptable, but the text could clarify that µϕ(G) is the mean vector of the embedding distribution.","section":"Section 4.2, Eq. (5)-(8)"}],"recommendation":"major_revision","confidential_remarks":"The empirical results are impressive and the code release is a plus. The main risk to the paper's contribution is the unvalidated additive counterfactual assumption in Eq. (16); if the authors can provide a theoretical justification or a convincing ablation that isolates this assumption, the paper would be a strong candidate for acceptance. The missing regression-adapted CGL baseline also weakens the stated novelty, but that is more easily addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful thing in this paper is the recipe, not the causal story. The authors combine an enhanced GIB loss that lets confounders contribute to prediction with an InfoNCE-style alignment loss, and on GOOD-ZINC and ReactionOOD it beats the classification-based baselines, often by a wide margin. That is a real, practically useful addition to the graph OOD toolbox, and the paper is the first to explicitly tackle regression in this setting. Credit where due: the ablation suggests both loss terms pull their weight, the sensitivity analysis is honest about hyperparameter dependence, and the code is promised.\n\nThe soft spot is the load-bearing causal claim. Equation (16) states that a mixed representation Hmix = Hc + Hs is a counterfactual graph representation, and that aligning Hg with Hmix via InfoNCE implements backdoor adjustment. Neither step is derived. For a GNN, the embedding of the union of two subgraphs is not generally the sum of their embeddings, and instance discrimination does not estimate P(Y | do(C)). The paper's own wording, 'can be understood as an implicit realization,' concedes the point. The empirical gains probably come from a generic contrastive regularizer that encourages invariance to S, which is fine as a trick but not a causal discovery method. There is also a conceptual tension: if S is genuinely predictive (the paper's premise), then invariance to S is not what you want in all shifts, and the interaction between -beta I(S;Y) and the invariance term is never analyzed.\n\nMinor issues: GroupDRO is listed as a baseline but doesn't appear in the tables; the appendix figures lack error bars; and there is no regression-specific OOD baseline, so the comparison is mostly against classification methods adapted to regression. None of these are fatal, but they add noise.\n\nWho should read this: anyone working on OOD generalization for graph regression, especially molecular property prediction. The method is simple and likely to transfer. The causal vocabulary should be taken as motivation, not as a guarantee.\n\nI'd send it to peer review. The recipe is worth publishing even if the causal story is softened. Ask the authors to either justify or drop the backdoor-adjustment claim, and to add a proper regression baseline. If they do, it's a solid paper.","headline":"Useful empirical recipe for graph regression OOD, but the causal-intervention claim is an analogy, not a derivation.","tokens_in":18427,"tokens_out":2946,"would_cite":true,"duration_ms":33192,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","62D20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Graph regression beats OOD baselines by letting confounders keep their predictive power.","keywords":["causal graph regression","out-of-distribution generalization","graph information bottleneck","confounding subgraph","contrastive learning","backdoor adjustment","graph neural networks"],"falsifier":"Run the method on a synthetic graph regression dataset with a known generative model where the confounder $S$ has no causal effect on $Y$ but is strongly correlated in training; if the contrastive intervention loss still improves out-of-distribution performance compared to ERM, then the improvement cannot be attributed to backdoor adjustment, or the mixed representation assumption must be violated in a way that happens to help.","tokens_in":17304,"feed_emoji":"🧪","tokens_out":2690,"duration_ms":29630,"temperature":0.7,"pith_summary":"This paper argues that existing causal graph learning methods fail on regression tasks because they assume confounding subgraphs contain no predictive information, an assumption that rarely holds in practice. It proposes a framework that lets the confounder contribute to prediction through an enhanced graph information bottleneck objective, and replaces label-dependent intervention with a contrastive alignment loss that does not need discrete classes. If the framework works as claimed, it gives graph regression models a practical route to out-of-distribution generalization without discarding useful signal. The authors report that their method achieves the best out-of-distribution error on all eight GOOD-ZINC settings and on six of ten ReactionOOD settings.","feed_headline":"Confounders keep predicting: new recipe for graph regression OOD","feed_subtitle":"A contrastive causal loss lets spurious features carry signal, beating prior OOD methods on eight GOOD-ZINC settings.","key_machinery":"The load-bearing objects are the enhanced graph information bottleneck objective (Eq. 4), which adds a confounder-predictiveness term $-\\beta I(S;Y)$ to the classical GIB loss, and the contrastive intervention loss (Eq. 17), which replaces label-conditioned counterfactual mixing with unsupervised InfoNCE alignment between $H_{g,i}$ and $H_{c,i}+H_{s,j}$. These are computed with variational Gaussian bounds, reducing the mutual-information terms to least-squares goals and an $\\ell^2$ penalty on the causal embedding mean.","core_discovery":"The paper claims to be the first to explicitly treat confounding features as predictive in graph regression, and to show that doing so improves generalization. Its central proposal is a two-part loss: an enhanced GIB objective $-I(C;Y)+\\alpha I(C;G)-\\beta I(S;Y)$ that keeps the confounder $S$ informative for the target $Y$ while still compressing the causal subgraph $C$, and a contrastive intervention loss $\\mathcal{L}_{\\mathrm{CI}}$ that aligns the original graph representation with randomly mixed causal-and-confounding representations. The mixed representation $H_{\\mathrm{mix},ij}=H_{c,i}+H_{s,j}$ is treated as a counterfactual graph, and the InfoNCE-style alignment is presented as an implicit backdoor adjustment in representation space. The paper reports that this recipe yields state-of-the-art out-of-distribution performance on graph regression benchmarks, improving over prior causal graph learning methods that discard confounding information.","pith_inferences":["The random addition $H_{c,i}+H_{s,j}$ is a representation-space counterfactual; whether it faithfully reflects the underlying graph structure is not proven, so the method may be most reliable when causal and confounding substructures are separable in embedding space.","The contrastive alignment could be reinterpreted as a regularizer that encourages the causal representation to be invariant to confounder changes, which suggests a direct connection to invariant risk minimization that the paper does not explore.","A testable extension is to apply the same two-part loss to node-level or link-level regression, where the definition of a causal subgraph may need a different mask structure.","One could probe the claim about confounder predictiveness by measuring the correlation between the learned confounder representation and the target on a held-out set where the spurious correlation is intentionally broken."],"forward_implications":["If the framework generalizes as reported, graph regression under distribution shift no longer needs to throw away confounders; modeling their predictive role should become a standard component of causal graph learning.","The contrastive intervention loss provides a label-free way to implement backdoor adjustment in representation space, which could make causal intervention applicable to continuous-label tasks beyond graphs.","The enhanced GIB objective suggests that information-theoretic disentanglement in regression should explicitly budget for confounder information rather than assuming it is noise.","The reported benchmark results, if reproduced, would give practitioners a default causal method for graph regression rather than forcing them to adapt classification-specific tools."],"supporting_citations":[{"why":"Supplies the random-addition intervention and the causal-attention framework that this paper adapts to regression.","marker":"Sui et al. (2022)"},{"why":"Provides the structural causal model and the backdoor-adjustment formulation that the contrastive intervention is said to realize implicitly.","marker":"Sui et al. (2024)"},{"why":"Defines the graph information bottleneck objective that the paper extends with the confounder term.","marker":"Wu et al. (2020)"},{"why":"Supplies the InfoNCE loss that the causal intervention loss is built on.","marker":"Oord et al. (2018)"},{"why":"Provides the Gaussian mutual-information approximation used to make the GIB objective computable.","marker":"Miao et al. (2022)"},{"why":"Supplies the GOOD-ZINC benchmark and the covariate/concept shift splits used for evaluation.","marker":"Gui et al. (2022)"},{"why":"Supplies the ReactionOOD benchmark with its structural out-of-distribution regression tasks.","marker":"Wang et al. (2023)"}],"fun_headline_variants":["Confounders as predictors: causal graph regression reworked","Keep confounders, boost OOD graph regression","Causal graph regression: confounders aid generalization","Spurious features join causal graph regression","Contrastive intervention improves graph OOD regression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method treats the vector sum of a causal representation and a confounding representation as a valid counterfactual graph, and assumes that aligning with such sums implements backdoor adjustment; if that mixed representation is not a faithful stand-in for intervening on the confounder, the claimed causal invariance does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Confounders as predictors: causal graph regression reworked","Keep confounders, boost OOD graph regression","Causal graph regression: confounders aid generalization","Spurious features join causal graph regression","Contrastive intervention improves graph OOD regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000771,"raw_usage":{"total_tokens":3387,"prompt_tokens":889,"completion_tokens":2498,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":2425}},"tokens_in":505,"tokens_out":2498,"duration_ms":17435,"temperature":1.0,"reasoning_tokens":2425,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:16:24.167241+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a synthetic graph regression dataset with a known generative model where the confounder $S$ has no causal effect on $Y$ but is strongly correlated in training; if the contrastive intervention loss still improves out-of-distribution performance compared to ERM, then the improvement cannot be attributed to backdoor adjustment, or the mixed representation assumption must be violated in a way that happens to help.","supporting_citations":[{"cited_title":"Causal attention for interpretable and generalizable graph classification","cited_arxiv_id":null,"evidence_quote":"Supplies the random-addition intervention and the causal-attention framework that this paper adapts to regression."},{"cited_title":"Enhancing out-of-distribution generalization on graphs via causal attention learning","cited_arxiv_id":null,"evidence_quote":"Provides the structural causal model and the backdoor-adjustment formulation that the contrastive intervention is said to realize implicitly."},{"cited_title":"Interpretable and generalizable graph learning via stochastic attention mechanism","cited_arxiv_id":null,"evidence_quote":"Provides the Gaussian mutual-information approximation used to make the GIB objective computable."},{"cited_title":"Good: A graph out-of-distribution benchmark","cited_arxiv_id":null,"evidence_quote":"Supplies the GOOD-ZINC benchmark and the covariate/concept shift splits used for evaluation."},{"cited_title":"Towards out-of-distribution generalizable predictions of chemical kinetics properties","cited_arxiv_id":"2310.03152","evidence_quote":"Supplies the ReactionOOD benchmark with its structural out-of-distribution regression tasks."}],"review_version":1}