{"id":"af9f1b23-0e70-409a-9e85-379806f3a5de","arxiv_id":"2411.19245","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A contrastive representation learning approach that aims to separate causal from non-causal latent factors in high-dimensional treatments for unbiased CATE estimation.","lead":"This paper proposes a contrastive learning method to learn representations of high-dimensional treatments, such as text, molecules, and products, that are meant to isolate the causal factors driving an outcome and discard spurious ones. The authors claim this yields unbiased causal effect estimates, and they test the approach on synthetic and real-world datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own SCM makes do(T) have zero effect on Y, so the CATE estimand in Eq. (2) is degenerate; Theorem 1's counterexample violates the stated structural equation, and the claimed bias motivating the contrastive method is unsupported.","rationale":"Reading the paper in good faith, the intended contribution is clear: learn ψ(T) that discards non-causal latents so that downstream CATE estimation is unbiased, with Theorem 3 transferring block-identifiability from von Kügelgen et al. However, the theoretical edifice rests on a causal model in which the observed treatment T is a child of the latents TC and TnC and is not a parent of Y. Under the paper's own Definition 1, do(T=t) is a well-defined intervention with zero effect on Y. The paper never defines the effect it claims to estimate except through τ(T, T′, x), which is identically zero in this SCM. If the authors intended the effect of the latent causes TC, then the contrastive representation is a proxy for an unobserved cause, not a treatment representation, and the language of Theorems 1–2 is misleading. The proof of Theorem 1 confirms the problem: the constructed alternative model changes the structural equation for Y. In the actual model, E[Y | T=(TC, TnC), X] = ρTC + δX, independent of TnC, so backdoor adjustment is not biased. This is not merely a proof gap; it removes the stated motivation for the method. I therefore agree with the reader that the paper should be rejected, and I would not change the verdict. The empirical results do not rescue the claim because the PEHE metric tests invariance to non-causal treatment components, not the causal effect of do(T), and no machine-checked proof or reproducible implementation is provided to offset the theoretical issue.","tokens_in":10630,"tokens_out":8246,"duration_ms":75802,"concrete_test":"Implement the Theorem 1 data-generating process exactly as written, e.g. X ~ N(0,1), TC = αX + εTC, TnC = βX + εTnC, Y = ρTC + δX + εY, T = (TC, TnC), and fit the true regression E[Y | T, X] by OLS. Evaluate E[Y | T=(tc, tnc), X=x] − E[Y | T=(tc, tnc′), X=x] for fixed tc and tnc ≠ tnc′ across the support of X: if this difference is zero, as it is analytically because Y does not depend on TnC, then the Theorem 1 claim of bias is refuted. Separately, symbolically compute do(T=t) in the Figure 1 SCM to verify that E[Y | do(T=t), X=x] is independent of t, which would confirm that the paper's CATE estimand is degenerate for the treatment T.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Under the model in Section 3 (T = m(TC, TnC), Y = f(TC, X, εY), Figure 1), T is not a cause of Y. An intervention do(T=t) replaces only the structural equation of T and leaves TC, X, and εY unchanged; hence E[Y | do(T=t), X=x] = E[Y | X=x] for every t, and τ(T, T′, x) ≡ 0. The paper's unbiasedness claims (Theorems 1–2) therefore either concern a vacuous zero effect or silently substitute the latent intervention do(TC), which is a different estimand from the one defined in Section 2. The proof of Theorem 1 illustrates the same slippage: its 'alternative model' Y = αTC + σ²βTnC + σ²X + εY adds a direct dependence of Y on TnC, violating the structural equation Y = f(TC, X, εY) stated at the start of Section 3. In the paper's own linear example, T = (TC, TnC) and Y = ρTC + δX + εY, so E[Y | T=(tc, tnc), X=x] = ρtc + δx, which does not depend on tnc; direct backdoor adjustment is unbiased and the claimed counterexample does not exist. Because the whole motivation for learning a contrastive treatment representation is the asserted bias of direct adjustment, and because the intended causal quantity is never coherently defined for the observed T, the central claim is unsupported unless the model and estimand are reformulated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies causal effect estimation when treatments are high-dimensional structured objects such as text, molecules, or product reviews. It models a treatment T as a function of causal latent factors TC and non-causal latent factors TnC, with outcome Y depending only on TC and covariates X. The paper claims that backdoor adjustment using T directly is biased when TnC is correlated with X, and it proposes a contrastive learning method to learn a representation ψ(T) that discards TnC. The theoretical contributions are stated as three theorems: Theorem 1 asserts bias of direct adjustment; Theorem 2 asserts that unbiased estimation is possible if and only if the representation contains no TnC information; and Theorem 3 asserts that the contrastive approach block-identifies TC. Experiments on synthetic, molecule, and recommender datasets compare the contrastive method to baselines using the PEHE metric.","tokens_in":10961,"tokens_out":5737,"duration_ms":46977,"significance":"The problem is practically relevant, and the idea of using contrastive learning to isolate the causal components of high-dimensional treatments is interesting and potentially useful. If the theoretical results were sound, the paper would offer a principled representation learning procedure for causal inference. However, the paper's own structural causal model renders the target estimand degenerate, and the proof of the central bias claim contradicts the model's structural equations. The empirical evaluation, while suggestive, is not anchored to a well-defined causal estimand and cannot rescue the theoretical failure. The manuscript therefore does not currently establish its central claims.","major_comments":[{"comment":"Under the SCM in Section 3, with Y = f(TC, X, εY) and T = m(TC, TnC), intervening on T via do(T=t) replaces only the equation for T and leaves TC, X, and εY unchanged. Consequently E[Y | do(T=t), X=x] = E[Y | X=x] for all t, and the CATE τ(T,T',x) defined in Eq. (2) is identically zero. The paper's theorems discuss unbiasedness relative to the latent CATE τ(TC,T'C,X), which is a different estimand from the one introduced in Section 2. This is a fundamental mismatch: the paper's motivating problem, bias in the effect of the observed treatment T, is not well-defined under the proposed model.","section":"Section 3 (structural equations) and Section 2 (Eq. (2))"},{"comment":"The proof constructs an alternative outcome model Y = αTC + σ²βTnC + σ²X + εY, which makes Y depend directly on TnC and therefore violates the structural equation Y = f(TC, X, εY) stated at the start of Section 3. In the paper's own linear DGP, E[Y | T=(tc,tnc), X=x] = ρtc + δx, which is independent of tnc; hence backdoor adjustment with T correctly identifies the conditional average treatment effect of TC. The claimed counterexample does not occur under the stated model, and the proof's appeal to 'a possible solution to regressing Y on T and X' conflates the true conditional expectation with an arbitrary alternative model that is not compatible with the SCM.","section":"Theorem 1 (proof)"},{"comment":"The statement is largely definitional. The paper defines unbiased CATE estimation as the condition that τ(T,T',X) reproduces τ(TC,T'C,X) in Section 3. Under that definition, the 'only if' direction essentially restates that a representation that reproduces the latent CATE cannot depend on TnC. The proof's key step—from equality of conditional expectations E[Y|ψ(T),X] = E[Y|ψ(T'),X] to the conclusion that ψ is constant on each TTC—assumes that ψ(T) itself is the quantity being intervened on, which is not established. The theorem therefore does not provide the claimed necessary-and-sufficient characterization, and it does not confront the degenerate estimand of Eq. (2).","section":"Theorem 2"},{"comment":"The proof consists of an invocation of Theorem 4.2 of Von Kügelgen et al. (2021) rather than a derivation. The augmentation argument requires exact equality of X and Y values to conclude TC = T'C via invertibility of f; the practical relaxation with thresholds δ and ε in Section 4 is not proved to preserve block identification. As a result, the claim that the contrastive algorithm 'provably' block-identifies the causal latents is not supported for the actual algorithm evaluated in Section 6.","section":"Theorem 3 and Section 4 (algorithm)"}],"minor_comments":[{"comment":"In the Introduction, 'we cloud improve product recommendation' should be 'we could improve'; please correct this typo.","section":"Introduction"},{"comment":"'V on Kügelgen et al.' should be 'Von Kügelgen et al.'; the same typo appears in the References entry for Von Kügelgen et al. (2021).","section":"Section 4 and References"},{"comment":"The PEHE formula in Eq. (1) omits the expectation or averaging convention; as written it is not clear whether the average is over a fixed test set or the population.","section":"Section 6, Eq. (1)"},{"comment":"The caption of Figure 2b, 'Performance under perturbations to y', does not match the text, which describes differences in predictions under perturbed non-causal treatment components.","section":"Figure 2b"},{"comment":"The notation f is used both for the outcome structural equation and for the treatment mixture T = f(TC,TnC) in Section 3; this is confusing and should be disambiguated.","section":"Section 3"},{"comment":"The standard errors are formatted inconsistently (e.g., '0 .74' should be '0.74'); please correct the spacing throughout the tables.","section":"Tables 1 and 2"}],"recommendation":"reject","confidential_remarks":"This manuscript does not meet the bar for publication because the central theoretical claim is internally inconsistent. A major revision would require redefining the SCM and the causal estimand, which would change the paper's scope. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper has a genuinely useful intuition — learn a treatment representation that isolates causally relevant latents via contrastive pairs — but the theoretical scaffolding around it is broken. The SCM in Section 3 makes the observed treatment T causally inert: Y = f(TC, X, εY), so an intervention on T has no effect on Y. That means the CATE defined in Section 2, τ(T,T′,x), is identically zero under their own model. The paper wants to estimate the effect of the latent TC but writes the estimand in terms of T, and that mismatch undermines the whole motivation.\n\nThe good part: the contrastive pairing idea is simple and plausible. Positive pairs with same X,Y and different T should share TC if f is invertible; negative pairs with same X and different Y should have different TC. Algorithmically that's a reasonable recipe, and the empirical results on synthetic data, molecules, and the Coat recommender show a consistent drop in PEHE compared to the two baselines. The authors also clearly position the work against SIN, GraphITE, and the text-causal-effects literature.\n\nNow the soft spots, which are load-bearing. Theorem 1 is meant to show backdoor adjustment with T is biased, but the proof constructs an alternative model Y = αTC + σ²βTnC + σ²X + εY that adds a direct TnC→Y edge, which violates the DAG in Figure 1. Worse, in the paper's own linear example, E[Y|T,X] = ρTC + δX, independent of TnC, so backdoor adjustment with X actually identifies the effect of TC (not the zero effect of T). The claimed counterexample doesn't exist. Theorem 2 is mostly definitional: \"unbiased\" is defined as matching the CATE of TC, so the equivalence is not surprising. Theorem 3 imports von Kügelgen et al.'s result, but the practical algorithm uses thresholds on g(X) and Y, and the Taylor expansion argument is too loose to be a proof. The experiments also have issues — the SIN baseline is modified (MLP instead of GNN), and several PEHE estimates come with large standard errors.\n\nWho gets value from this? Readers working on causal representation learning for treatments will find the contrastive recipe worth trying, but they should not rely on the theoretical guarantees. The paper is a useful example of a promising heuristic with an unfortunately flawed formal apparatus. For peer review, I'd send it to a competent referee because the topic is timely and the errors are instructive, but I'd reject in its current form. The authors need to re-define the estimand in terms of TC, provide a valid bias proof, and tighten the gap between the theory and the thresholded algorithm.","headline":"A promising contrastive recipe undermined by an internally inconsistent SCM; the bias theorem's proof violates the paper's own model.","tokens_in":11575,"tokens_out":6163,"would_cite":false,"duration_ms":51951,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"To estimate causal effects of high-dimensional treatments, discard non-causal latents.","keywords":["causal inference","high-dimensional treatments","contrastive learning","treatment representation","conditional average treatment effect","backdoor adjustment","disentanglement","structured treatments"],"falsifier":"Construct a synthetic dataset with the same structure as Figure 1 but add a direct edge from T to Y, e.g., Y = f(TC, X) + γ·T + εY, and estimate the CATE using the contrastive representation ψ(T) after training on this data; if the resulting estimate is biased, the paper's claim that ψ(T) discards all non-causal information and preserves CATE is falsified.","tokens_in":10357,"feed_emoji":"🎯","tokens_out":2731,"duration_ms":23559,"temperature":0.7,"pith_summary":"This paper addresses causal effect estimation when treatments are high-dimensional structured objects such as text, video, or molecules. The observed treatment T is assumed to be a mixture of latent causal factors TC and non-causal factors TnC, with the outcome Y depending only on TC and covariates X. The paper shows that standard back-door adjustment using T directly is biased, even with all confounders observed, whenever TnC correlates with confounders. It proves that using a learned representation ψ(T) that contains no information about TnC is necessary and sufficient for unbiased estimation, and provides a contrastive algorithm that provably block-identifies TC. Empirical results on synthetic, molecule, and recommender datasets support the claim that contrastive representations yield lower bias than existing methods.","feed_headline":"Contrastive trick debiases causal effects for complex treatments","feed_subtitle":"Learning a representation that discards non-causal latents makes back-door adjustment unbiased for text, video, and molecules.","key_machinery":"The contrastive pair construction is the mechanism that carries the argument: positive pairs are samples with X ≈ X′ and Y ≈ Y′ but T ≠ T′, which forces TC to be approximately equal while TnC varies; negative pairs have X ≈ X′ but Y not close, forcing TC to differ. Under smooth invertible functions, this augmentation scheme block-identifies TC by appealing to an existing theorem on self-supervised contrastive learning that isolates content from style. The resulting representation ψ(T) is then used as the treatment in back-door adjustment, which Theorem 2 shows is unbiased exactly when ψ(T) discards TnC.","core_discovery":"The central claim is that causal effect estimation from high-dimensional treatments is unbiased if and only if the treatment representation ψ(T) contains no information about non-causal latent variables TnC. Under structural equations X = l(εX), TC = g(X, εTC), TnC = h(X, εTnC), T = m(TC, TnC), and Y = f(TC, X, εY), the outcome depends on the treatment only through TC. The paper constructs a contrastive learning method that creates positive pairs from samples sharing the same X and Y (hence the same TC, since f is invertible) and negative pairs with different Y (hence different TC). Applying a known block-identification result, the method provably yields a representation ψ(T) = φ(TC) that isolates the causal latents. This representation preserves interventions on TC and consequently gives unbiased CATE estimates, whereas using T directly fails when TnC acts as a proxy for confounders.","pith_inferences":["If the outcome depends on T directly (violating Y = f(TC, X, εY)), the contrastive representation may not preserve the intervention on T, and the unbiasedness guarantee could fail; a modified estimand or additional assumptions would be needed.","The practical algorithm uses thresholds δ and ε to construct pairs; the bias introduced by these approximations remains unquantified, so a sensitivity analysis relative to δ and ε is a natural next step.","The block-identification result suggests the method could extend to settings with multiple treatment views or where partially observed causal latents exist, potentially enabling semi-supervised causal representation learning.","The authors' approach could be tested on text sentiment examples (e.g., tone vs. style) to verify empirically that the learned representation isolates tone and discards style, as the motivating example suggests."],"forward_implications":["Causal effect estimation for text, video, audio, and molecular treatments can be made unbiased by learning a contrastive representation that isolates the causally relevant latents.","Recommendation systems and drug discovery could benefit: if the causal aspects of a product or molecule are identified, similar causal profiles can be matched more efficiently.","The paper's Theorem 2 provides a necessary and sufficient condition for unbiasedness, giving practitioners a clear criterion for when a treatment representation is valid.","Existing methods for high-dimensional treatment effect estimation that use the treatment directly, such as SIN, are shown to be biased in this setting, motivating adoption of contrastive pre-training.","The fully non-parametric nature of the approach means it can be applied without specifying a parametric model of the treatment-outcome relationship."],"supporting_citations":[{"why":"Supplies the block-identification theorem used to prove that the contrastive augmentation scheme isolates TC from TnC.","marker":"V on Kügelgen et al. [2021]"},{"why":"Provides the structural causal model framework and the back-door adjustment formula that the paper's bias analysis relies on.","marker":"Pearl [2009]"},{"why":"The SIN method is the primary baseline in experiments, and the paper demonstrates that it yields biased estimates in this setting.","marker":"Kaddour et al. [2021]"},{"why":"Introduces contrastive predictive coding, the foundation for the contrastive learning approach used in the algorithm.","marker":"Oord et al. [2018]"},{"why":"Defines the PEHE metric used to evaluate the error in treatment effect estimation in the experiments.","marker":"Hill [2011]"}],"fun_headline_variants":["Contrastive reps debias causal estimates for complex treatments","Contrastive learning isolates causal latents for unbiased effects","Drop non-causal latents: contrastive reps give unbiased CATE","Contrastive reps discard non-causal info, debias treatment effects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The outcome Y is assumed to depend on the treatment only through the latent causal components TC, not on the observed treatment T directly; if T itself has a direct causal effect on Y, the contrastive representation may not preserve the intervention and the unbiasedness guarantee collapses.","fun_headline_variants_meta":{"raw":{"variants":["Contrastive reps debias causal estimates for complex treatments","Contrastive learning isolates causal latents for unbiased effects","Drop non-causal latents: contrastive reps give unbiased CATE","Contrastive reps discard non-causal info, debias treatment effects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000544,"raw_usage":{"total_tokens":2570,"prompt_tokens":880,"completion_tokens":1690,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":1617}},"tokens_in":496,"tokens_out":1690,"duration_ms":10840,"temperature":1.0,"reasoning_tokens":1617,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:23:51.615458+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a synthetic dataset with the same structure as Figure 1 but add a direct edge from T to Y, e.g., Y = f(TC, X) + γ·T + εY, and estimate the CATE using the contrastive representation ψ(T) after training on this data; if the resulting estimate is biased, the paper's claim that ψ(T) discards all non-causal information and preserves CATE is falsified.","supporting_citations":[{"cited_title":"Causality (2nd edition)","cited_arxiv_id":null,"evidence_quote":"Provides the structural causal model framework and the back-door adjustment formula that the paper's bias analysis relies on."}],"review_version":1}