WorksComputer Science
Regularizing Treatment Leakage for LLM-based Uplift Estimation
Abstract
Large language models (LLMs) provide semantic features for uplift estimation, but treatment-revealing text can induce prediction shortcuts that do not transfer to deployment. We propose LLM-based Uplift Leakage Regularization (LULR) to address this risk. With the LLM encoder frozen, LULR penalizes dependence between the causal head's trainable shared representation and treatment while retaining separate outcome heads. The penalty uses a variational approximation to the contrastive log-ratio upper bound (CLUB) of mutual information, targeting treatment dependence rather than generic parameter shrinkage. Controlled proxy experiments provide exploratory evidence of improved ranking scores under injected treatment information relative to no regularization and the tested weight-decay setting. Structured-only benchmarks show modest, dataset-dependent effects. On an industrial pricing randomized controlled trial, mean offline [email protected] is 0.258 for LULR, compared with 0.106 for the same-backbone baseline. A separate one-week online A/B test of the complete deployed pricing policy records 5.45% higher sales volume and 1.91% higher gross merchandise value than the control arm. These short-term policy-level gains do not isolate the regularizer's effect.
Introduction
In marketing, recommendation, and pricing, decision makers need to understand not only the average effect of an intervention but also which units respond to it, so that budgets can be allocated accordingly. This is uplift modeling, i.e., estimation of the conditional average treatment effect (CATE) [1], [2]. Classical solutions (meta-learners, causal forests) operate on structured covariates and are mature [3], [4]. A newer line brings text into causal estimation, letting language models encode covariates, treatments, or confounders expressed in natural language [5]–[7].
Industrial pricing is a natural fit: a product’s title carries demand-relevant semantics (“essential daily item,” “niche specialty good”) that structured features miss. Product text can also carry semantic treatment leakage if its recorded content reflects the intervention itself. For example, “clearance” added after a discount decision could reveal that decision rather than describe pre-existing demand. Such content may improve factual prediction through a shortcut without helping the model learn how demand semantics moderate the treatment effect. The shortcut may fail when future decisions are still pending or text associations change. We study this potential form of shortcut learning [8], rather than assume that promotional vocabulary by itself establishes leakage. The controlled crossover experiment (§ 4-B) probes the risk through an injected proxy whose outcome association changes between training and evaluation.
We address this risk with LULR, a treatment-leakage regularizer for text-enhanced uplift estimation. A frozen LLM encodes product text, and a trainable causal head combines these features with structured covariates. An information-bottleneck-inspired penalty targets between the head’s shared representation and treatment. We instantiate it with the contrastive log-ratio upper bound (CLUB) of mutual information [9]. Its variational form supplies a differentiable surrogate for minimizing representation–treatment dependence. Our contribution is this treatment-targeted regularization within an existing uplift architecture, not a new mutual-information estimator or language model.
The evaluation proceeds from controlled proxy behavior to public-data reference comparisons, industrial offline ranking, and online policy performance. Each stage addresses a different question. In particular, the proxy retains treatment information at evaluation, so its score changes are exploratory behavioral evidence; the online A/B test instead evaluates the complete deployed pricing policy.
Contributions.
We formulate LULR to address potential treatment-information shortcuts by penalizing in the trainable causal head, with frozen LLM features (§ 3).
We characterize ranking-score behavior under an injected treatment-dependent proxy, including strength sweeps and a weight-decay comparison (§ 4-B).
We report structured-only reference comparisons and industrial offline results, followed by descriptive one-week online A/B results: sales volume +5.45% and GMV +1.91% relative to the control arm (§ 4-E).
Related Work
Uplift / CATE estimation. Meta-learners (S/T/X/R/DR) reduce CATE estimation to supervised sub-problems [1], [10]–[12], often built on double/debiased machine learning [13]; causal forests and generalized random forests provide nonparametric, honest estimators [3, 4], and recent reviews and tutorials survey this space [2, 14, 15]. Neural approaches such as DragonNet [16] and CFRNet [17] share a representation across potential-outcome heads; we adopt a DragonNet-style head as our backbone and place the regularizer on its shared representation.
Text-based causal inference. A growing body of work encodes text for treatment-effect estimation [5, 6, 7, 18, 19, 20, 21, 22, 23, 24]. Most target confounding expressed in text. We instead target treatment leakage in text under a randomized design, where treatment randomization does not by itself prevent model inputs from containing intervention-related information.
Treatment leakage. Recent work diagnoses text leakage and mitigates it via distillation, filtering, or sensitivity analysis [25, 26, 27]; related deconfounding methods learn representations robust to weak overlap [28]. Our contribution differs in kind: a differentiable information-theoretic constraint (CLUB upper bound of ) optimized in the causal head with frozen LLM features, rather than a post-hoc diagnostic or discrete text filter.
Dependence minimization. Several differentiable penalties can suppress representation–treatment dependence: MINE [29] estimates a mutual-information lower bound (well-suited to maximization but not to the minimization we need), HSIC [30] measures kernel dependence without a direct information-theoretic interpretation, and adversarial discriminators require a delicate min–max balance. We adopt the CLUB upper-bound construction [9] as a minimization objective, while accounting for estimation limitations [31]. Its variational approximation conditions are discussed in § 3.
Foundation causal estimators. In-context estimators such as CausalPFN [32] amortize CATE estimation across tasks; they are complementary to our focus on textual leakage under a fixed backbone.
Policy learning and uplift evaluation. Downstream of CATE estimation, policy learning selects whom to treat under a budget via empirical welfare maximization and doubly robust objectives [33, 34, 35], and learning-to-rank formulations optimize the treatment ordering directly [36]. Evaluating uplift rankings is itself subtle on RCT data [37, 38]; we report a held-out lift@ and an IPW policy value under a three-way split (§ 2) to avoid winner’s-curse bias.
Synthetic and controlled evaluation. Because ground-truth CATE is unobservable, controlled and semi-synthetic constructions are standard for isolating a mechanism [15, 39]. Our injected leakage proxy (§ 4-B) follows this tradition by comparing ranking-score changes with and without an injected artifact.
Problem Setup
We work under a randomized controlled design (RCT). Each unit has valid pre-treatment covariates (structured attributes and any text recorded before treatment assignment), a binary treatment (discount vs. baseline price) assigned independently of , and an outcome (next-day gross merchandise value, GMV). Under the potential-outcomes framework the target is the CATE
and uplift modeling seeks a scoring rule that ranks units by so that the top- responders can be selected for treatment. Because in an RCT, is identified without confounding adjustment.
Valid covariates versus recorded inputs. Let denote the text actually supplied to the model. When recorded before treatment assignment, it may be included in ; when it contains intervention-related updates, the RCT condition does not imply . This distinction separates the CATE identification setting from the potential input-contamination risk addressed by LULR. The industrial data are not assumed to contain such contamination merely because they include product titles. The controlled study explicitly introduces a treatment-dependent proxy, and its evaluation therefore has a different interpretation (§ 4-B).
Prediction target and ranking score. In the neural runs, the factual outcome target is . The two outcome heads predict this transformed outcome; their difference is used directly for ranking. We treat as a surrogate for raw-outcome CATE ranking, not as a direct estimate of : the nonlinear transformation need not preserve treatment-effect rankings.
Evaluation. We report the normalized Qini coefficient (QINI_NORM), defined as the Qini area divided by . Random ranking thus scores approximately zero. For heavy-tailed outcomes with unstable normalization, we also report the decision metric lift@. Let contain the top- fraction of units ranked by . If selection uses only pre-treatment information, the difference between the two RCT arms within estimates its treatment effect. We report the excess over the global ATE:
Under these conditions, positive lift@ estimates above-average responsiveness; sampling uncertainty still applies. Neural runs use a three-way split: the epoch is selected on validation data and the selected model is evaluated on held-out test data. The historical tree-based reference runs use their own train/test splits and are not paired comparisons with these neural runs. We also report the IPW policy value of the induced decision rule. Neural normalized Qini uses , whereas the reported lift@ and policy values use the original outcome scale. The QINI_NORM normalizer degrades only when the ATE is near zero relative to outcome scale; on the public datasets the ATE is a non-negligible fraction of the outcome standard deviation, so QINI_NORM is well-behaved, whereas on industrial GMV (highly zero-inflated, extreme tail) strong prefix gains dwarf and the metric becomes unreliable; there we rely on [email protected].
Method: Treatment-Leakage Regularization
Motivation: potential shortcuts in text inputs
A frozen LLM encodes the recorded text , and its features are fused with structured covariates into . The causal head learns the shared representation and treatment-specific predictions of . If contains intervention-related cues, may rely on them for factual prediction rather than capture transferable effect heterogeneity. This motivates penalizing representation–treatment dependence while retaining separate outcome heads and . The penalty does not require these heads to agree: it targets implicit treatment cues in , not the treatment contrast needed to estimate an effect. Figure 1 illustrates this design hypothesis.

Figure 1. Potential treatment-information shortcut and the LULR objective, starting from the same recorded text. CLUB penalizes representation–treatment dependence; the dashed arrow indicates the intended deployment behavior.
Representation-level treatment constraint
Inspired by the information bottleneck [40], we fit the factual outcome while constraining how much treatment information the trainable shared representation may carry. For causal-head parameters , the objective is
The explicit treatment argument selects the corresponding outcome head; the constraint applies only to . Under an RCT, this is a regularization objective, not a confounding adjustment. It does not turn post-treatment inputs into valid pre-treatment covariates.
Differentiable realization via the CLUB upper bound
To implement Eq. (3), we use the contrastive log-ratio upper-bound construction (CLUB) [9], subject to the estimation conditions discussed below [31]. A lightweight auxiliary classifier approximates the treatment conditional. Its variational CLUB estimate on a batch of size is
The first term scores the observed pairs ; the second pairs each with every treatment label in the batch and averages their scores. The contrast measures how much better
the auxiliary classifier scores observed pairs than re-paired samples. Minimizing it discourages treatment-predictive cues in the shared representation. Training alternates, as in CLUB: each step first updates to fit on the detached representation (maximizing the log-likelihood term), then updates the main network to minimize the resulting estimate, yielding the Lagrangian objective
where is the RCT grouped factual loss on (each arm updates its own outcome head), is the DragonNet propensity term, and weights the CLUB penalty. We retain the baseline’s propensity branch and loss when adding CLUB. Unlike the detached auxiliary fit, the propensity loss updates to predict , so it can oppose the dependence penalty. The weights and control this trade-off between treatment prediction and dependence suppression.
Approximation conditions. The variational estimate depends on how well approximates ; a poorly fitted classifier can violate the upper-bound condition. Updating before the main network supports this approximation but does not certify an upper bound at every step.
Where the penalty acts. A key design point: CLUB must act on the causal head’s trainable shared representation , not on the raw fused input . If applied to a non-trainable input the penalty has zero gradient to the model and is inert. Figure 2 shows the architecture.

Figure 2. LULR architecture. Solid arrows show ranking inference; dashed arrows and the shaded panel mark training-only branches. CLUB regularizes the trainable shared representation ; the optional LLM encoder remains frozen.
Applicability and theoretical scope
RCT independence concerns valid pre-treatment covariates, whereas the penalty targets empirical dependence in learned representations. Finite-sample correlations can cause useful predictive directions to be penalized along with unwanted treatment cues. Thus trades dependence suppression against potential over-balancing; neither finite-sample nor asymptotic removal of only leakage is guaranteed. The regularizer also does not extend RCT identification to a treatment-dependent proxy. Its empirical scope is assessed through the separate evaluations below.
Experiments
Datasets and protocol
We report results on one industrial dataset, two public datasets, and one constructed semi-synthetic dataset (Table I). The industrial data contain product titles; semi_text contains constructed text. The neural results reported for hillstrom, x5, and semi_text use structured features only, whereas the industrial neural runs use LLM text features. On industrial we use a temporal split aligned with the deployment platform (train before a cutoff, evaluate on a later window). A temporal split alone does not establish that semantic treatment leakage is present. The main neural comparisons use five paired seeds; reported SDs describe across-seed dispersion with divisor . For text-enabled neural runs, text is encoded by a frozen Qwen2.5-0.5B-Instruct backbone [41], whose pooled representation is fused with the structured features before the causal head.
| Dataset | Size | Dim. | Text | Note |
| industrial | 9 | ✓ | product-pricing RCT, GMV (zero-inflated, heavy-tailed) | |
| semi_text | 8K | 2 | ✓ | text carries the signal |
| hillstrom | 64K | 5 | — | structured, email marketing |
| x5 | 200K | 11 | — | structured, low-dim weak signal |
Table I. Datasets. “Dim.” counts structured covariates; “Text” denotes availability, not use by every model. The reported neural runs use text only on industrial; semi_text is a constructed semi-synthetic dataset.
Baselines. We select controls to assess the regularizer’s incremental contribution and provide external reference performance. The primary control, labeled “Baseline (MLP head)” in Tables IV and V, uses the same DragonNet-style causal head [16], factual and propensity losses, input features, seed-specific splits, and validation-based model selection as LULR, but omits the CLUB penalty and its auxiliary classifier. It is therefore a same-backbone control, not a separately chosen generic MLP. The public comparison uses structured features only; both industrial methods additionally use the same frozen LLM encoder. S-learner and X-learner+tfidf [1], both with gradient-boosted-tree base learners, provide structured and text-feature reference points, respectively. Their historical runs use different protocols, so they contextualize performance rather than isolate the effect of CLUB. Finally, the no-regularization and weight-decay settings in Table III compare the treatment-targeted penalty with generic shrinkage under the same injected-proxy setting.
Controlled proxy experiment
This experiment asks how the regularizer changes model behavior when a treatment-dependent proxy is introduced. We construct a semi-synthetic crossover on hillstrom: the injected
feature remains correlated with treatment, while its outcome association is changed between training and evaluation. The main experiment (Table II) fixes the injection strength at ; we then sweep (Figure 3) to show the finding is not tied to this value. We compare CLUB on () vs. off (), under injected leakage (leak) and without it (clean), over five seeds with a paired -test. The construction diagnostics confirm the intended association shift: across seeds, the proxy–outcome correlation is 0.741–0.743 in training and 0.053–0.065 in the later evaluation window, while the full-sample proxy–treatment correlation remains 0.700–0.701.
| Setting | CLUB effect | ||
| Leakage present | +0.165 | 4.59 | 0.010 |
| Leakage absent | −0.001 | −0.03 | 0.977 |
| Interaction | +0.166 | 2.91 | 0.044 |
Table II. Controlled proxy experiment (5 seeds, paired -test on Qini_norm). CLUB effect is the score at minus that at ; interaction is the difference between effects with and without leakage. These are descriptive score changes under the constructed proxy.

Figure 3. Leakage-strength sweep: mean Qini interaction across 5 seeds, except , which averages two independent 5-seed runs. Lines guide the eye; no uncertainty intervals are shown.
Table II reports a positive interaction in the computed ranking score (+0.166, nominal =0.044). The CLUB effect is positive under injected leakage (+0.165, =0.010), while no statistically detectable difference is observed without it (−0.001, =0.977). The latter is not an equivalence test and does not establish harmlessness. These exploratory results cover five seeds on one injected dataset, with three related tests and no multiplicity correction.
Interpretation. The injected feature is a scalar proxy, not a semantic direction inside an LLM embedding. It still contains the realized treatment at evaluation. Selection by scores using this proxy need not preserve randomized comparability within ranked subsets, so the reported Qini and lift changes cannot be read as causal-ranking gains. The experiment characterizes behavior under the constructed artifact; it does not establish semantic leakage removal.
Variability across seeds. We repeated the experiment on an independent seed set (6–10). The interaction remained positive (+0.100 vs. +0.166 on seeds 1–5; leak effect +0.095, clean effect ), i.e. the direction reproduces while the magnitude varies run-to-run.
1) Robustness across leakage strengths: To examine sensitivity to proxy strength, we sweep the injected leakage strength (Figure 3). Two findings are stable. First, the CLUB interaction effect is positive at every leakage strength (qini interaction 0.13–0.21), and the no-leakage (clean) effect stays throughout. This is a descriptive pattern, not an equivalence guarantee. Second, the effect is not strictly monotone in : the point at is reproducibly lower (0.133 and 0.124 over two independent 5-seed runs, mean 0.128) than its neighbours. We therefore report consistency of the observed score-change direction across the tested strengths, not a causal or monotone dose–response.
| Regularization | QINI_NORM | [email protected] |
| (A) none | ||
| (B) weight decay () | ||
| (C) CLUB min |
Table 3. REGULARIZATION UNDER INJECTED LEAKAGE (; MEANSD ACROSS 5 SEEDS). BOLD MARKS THE LARGEST MEAN PER METRIC.
2) Comparison with weight decay: To compare the treatment-targeted penalty with generic shrinkage, we evaluate three regularization settings in the leak condition (, 5 seeds): (A) no regularization, (B) optimizer weight decay, and (C) CLUB. Weight decay serves as a generic control that does not target . Both QINI_NORM and [email protected] are computed on the same runs (Table III). In this separate rerun, the main optimizer uses weight decay in (B) and explicitly sets it to 0 in (A) and (C). These arm means are not pooled with the runs in Table II.
Under the tested leakage setting, weight decay changes QINI_NORM by relative to no regularization, whereas CLUB changes it by . The gap between CLUB and this weight-decay setting is , with the same direction on [email protected]. Thus, the observed score improvement is not reproduced by this particular generic regularizer. This comparison does not establish superiority over all regularization alternatives, nor does agreement between the two metrics remove the treatment-dependent selection concern.
Reference comparisons on public datasets
We next ask whether the regularizer adds value to the same neural backbone outside the injected-proxy setting. Table IV separates this matched comparison from historical tree references. Relative to the same MLP causal head without CLUB, LULR increases mean QINI_NORM by 0.011 on and 0.016 on semi_text; the change on hillstrom is . Thus, the same-backbone results show modest mean improvements on two datasets, rather than a uniform gain.
| Method (backbone) | hillstrom | x5 | semi_text |
| Same-backbone comparison, structured features only | |||
| Baseline (MLP head) | 0.051 ± 0.021 | 0.143 ± 0.049 | 0.079 ± 0.006 |
| LULR (MLP head) | 0.049 ± 0.014 | 0.154 ± 0.040 | 0.095 ± 0.014 |
| vs. baseline | −0.002 ± 0.015 | +0.011 ± 0.043 | +0.016 ± 0.011 |
| Historical reference scores (different protocols) | |||
| S-learner (GBDT) | 0.072 | 0.244 | 0.105 |
| X-learner+tfidf (GBDT) | 0.030 | 0.249 | 0.399 |
Table IV. Public and semi-synthetic QINI_NORM. Neural entries are mean±SD over 5 paired seeds; is LULR minus the baseline. Tree rows are historical, unpaired references: S-learner on Hillstrom/X5 uses archived five-seed summary means; the remaining tree entries are single-run scores.
All neural entries use structured features only, including semi_text; they do not measure an LLM text gain or LLM numerical-processing ability. The tf-idf reference uses serialized structured features, supplemented with constructed text on semi_text. Tree scores provide context, not a paired test of backbone superiority.
Industrial product pricing
The industrial comparison asks whether the regularizer improves held-out ranking when the model uses real product text. We rank responses to a price discount versus a baseline price under the temporal split. GMV combines price and volume: increased sales can be partly offset by lower unit prices, so volume gains need not translate one-for-one into GMV gains. For this heavy-tailed, zero-inflated outcome, we report [email protected] rather than rely on the unstable normalized Qini score (Table V).
| Method | [email protected] | vs. baseline |
| Baseline (MLP head) | 0.106 ± 0.262 | — |
| LULR | 0.258 ± 0.361 | +0.152 ± 0.444 |
Table V. Industrial [email protected] (mean±SD across 5 paired seeds). is LULR minus the baseline. Both methods use the same frozen LLM encoder.
LULR improves the mean [email protected] from 0.106 to 0.258. It outperforms the baseline in 4/5 seeds. The paired differences for seeds 1–5 are , , , , and , respectively. The negative result contributes to the large variance; the difference is not statistically significant. The estimated policy values are nearly tied: 7.10 for LULR and 7.15 for the baseline. Thus, the directional ranking gain does not establish a decision-value improvement. These offline results motivate a separate assessment of deployed policy performance (Section V-E), rather than imply a business gain from ranking metrics alone.
Online A/B test
To assess business performance beyond offline ranking metrics, the LULR-based pricing policy was evaluated in an online A/B experiment on the production pricing platform. Here the A/B arms denote policy assignments, whereas in the offline task denotes whether a product receives a discount. The online comparison therefore assesses the complete deployed policy rather than the isolated CLUB penalty. Over a one-week observation window, the experimental arm recorded 5.45% higher sales volume and 1.91% higher GMV relative to the control arm. These are relative changes, not percentage-point changes of offline uplift scores. We report aggregate relative results as point estimates.
The joint increase in sales volume and GMV is consistent with the business explanation that additional sales offset the reduction in unit prices. This is an aggregate interpretation, not a separately identified causal decomposition of price and volume effects.
Discussion and limitations
Structured-data boundary. On , the MLP baseline without CLUB (0.143) and LULR (0.154) both fall below the reported GBDT S-learner reference (0.244). Within the explored neural hyper-parameter grid, scores reach approximately 0.166, with similar configuration averages at (0.146) and (0.145). These observations suggest that backbone choice matters, but differences in reference protocols prevent attributing the entire gap to one factor. Since these neural runs omit text, they neither measure an LLM text gain nor demonstrate that trees cannot use textual features.
Evidence scope. The controlled study uses a scalar treatment-dependent proxy to characterize model behavior, not to establish unbiased causal ranking or semantic leakage removal. The industrial offline comparison is statistically inconclusive, with near-tied estimated policy values. The online experiment offers complementary policy-level evidence: the positive one-week aggregate results assess the complete pricing policy, do not isolate the regularizer, and do not establish profit improvement or long-term performance. These evaluations address distinct outcomes and cannot substitute for one another.
Conclusion
We presented LULR, which regularizes treatment dependence in the trainable causal head of an uplift model with a frozen LLM. The evaluation separates exploratory proxy behavior, references using only structured inputs, uncertain industrial offline gains, and deployed policy outcomes. In a one-week online A/B experiment, sales volume rises by 5.45% and GMV by 1.91% relative to the control arm. These results provide descriptive short-term evidence for the complete pricing policy. Our contribution is a regularization method with a scoped empirical assessment. It does not guarantee semantic leakage removal or superiority across estimators. Future work will isolate the regularizer’s contribution to online policy gains and assess longer-term deployment.
References
- [1]S. R. Künzel, J. S. Sekhon, P. J. Bickel, and B. Yu, “Meta-learners for estimating heterogeneous treatment effects using machine learning,” arXiv preprint arXiv:1706.03461, 2017.DOI
- [2]P. Gutierrez and J.-Y. Gérardy, “Causal inference and uplift modelling: a review of the literature,” in International Conference on Predictive Applications and APIs, 2017, pp. 1–13.
- [3]S. Wager and S. Athey, “Estimation and inference of heterogeneous treatment effects using random forests,” arXiv preprint arXiv:1510.04342, 2015.arxiv.org/abs/1510.04342
- [4]S. Athey, J. Tibshirani, and S. Wager, “Generalized random forests,” arXiv preprint arXiv:1610.01271, 2016.arxiv.org/abs/1610.01271
- [5]V. Veitch, D. Sridhar, and D. M. Blei, “Adapting text embeddings for causal inference,” arXiv preprint arXiv:1905.12741, 2019.arxiv.org/abs/1905.12741
- [6]R. Pryzant, D. Card, D. Jurafsky, V. Veitch, and D. Sridhar, “Causal effects of linguistic properties,” arXiv preprint arXiv:2010.12919, 2020.arxiv.org/abs/2010.12919
- [7]Y.-F. Zhang, H. Zhang, Z. C. Lipton, L. E. Li, and E. P. Xing, “Exploring transformer backbones for heterogeneous treatment effect estimation,” arXiv preprint arXiv:2202.01336, 2022.arxiv.org/abs/2202.01336
- [8]R. Geirhos, J.-H. Jacobsen, C. Michaelis, R. Zemel, W. Brendel, M. Bethge, and F. A. Wichmann, “Shortcut learning in deep neural networks,” Nature Machine Intelligence, vol. 2, no. 11, pp. 665–673, 2020.DOI
- [9]P. Cheng, W. Hao, S. Dai, J. Liu, Z. Gan, and L. Carin, “CLUB: A contrastive log-ratio upper bound of mutual information,” in International Conference on Machine Learning (ICML), 2020, pp. 1779–1788.arxiv.org/abs/2006.12013
- [10]D. Jacob, “CATE meets ML: The conditional average treatment effect and machine learning,” arXiv preprint arXiv:2104.09935, 2021.arxiv.org/abs/2104.09935
- [12]E. H. Kennedy, “Towards optimal doubly robust estimation of heterogeneous causal effects,” Electronic Journal of Statistics, vol. 17, no. 2, pp. 3008–3049, 2023.arxiv.org/abs/2004.14497
- [13]V. Chernozhukov, D. Chetverikov, M. Demirer, E. Duflo, C. Hansen, W. Newey, and J. Robins, “Double/debiased machine learning for treatment and structural parameters,” The Econometrics Journal, vol. 21, no. 1, pp. C1–C68, 2018.
- [14]E. Sverdrup, M. Petukhova, and S. Wager, “Estimating treatment effect heterogeneity in psychiatry: A review and tutorial with causal forests,” arXiv preprint arXiv:2409.01578, 2024.
- [15]A. Curth, A. Hüüyük, and M. van der Schaar, “Adaptive identification of populations with treatment benefit in clinical trials: Machine learning challenges and solutions,” arXiv preprint arXiv:2208.05844, 2022.arxiv.org/abs/2208.05844
- [16]C. Shi, D. M. Blei, and V. Veitch, “Adapting neural networks for the estimation of treatment effects,” Advances in Neural Information Processing Systems (NeurIPS), vol. 32, 2019.arxiv.org/abs/1906.02120
- [17]U. Shalit, F. D. Johansson, and D. Sontag, “Estimating individual treatment effect: generalization bounds and algorithms,” in International Conference on Machine Learning (ICML), 2017, pp. 3076–3085.
- [18]K. A. Keith, D. Rice, and B. O’Connor, “Text as causal mediators: Research design for causal estimates of differential treatment of social groups via language aspects,” arXiv preprint arXiv:2109.07542, 2021.arxiv.org/abs/2109.07542
- [19]H. Arno, P. Rabaey, and T. Demeester, “From text to treatment effects: A meta-learning approach to handling text-based confounding,” arXiv preprint arXiv:2409.15503, 2024.arxiv.org/abs/2409.15503
- [20]V. Lin, L.-P. Morency, and E. Ben-Michael, “Text-transport: Toward learning causal effects of natural language,” arXiv preprint arXiv:2310.20697, 2023.arxiv.org/abs/2310.20697
- [21]J. Kaddour, Y. Zhu, Q. Liu, M. J. Kusner, and R. Silva, “Causal effect inference for structured treatments,” arXiv preprint arXiv:2106.01939, 2021.arxiv.org/abs/2106.01939
- [22]N. Dhawan, L. Cotta, K. Ullrich, R. G. Krishnan, and C. J. Maddison, “End-to-end causal effect estimation from unstructured natural language data,” arXiv preprint arXiv:2407.07018, 2024.arxiv.org/abs/2407.07018
- [23]S. Guo, M. G. Marmarelis, F. Morstatter, and K. Lerman, “Estimating causal effects of text interventions leveraging large language models,” arXiv preprint arXiv:2410.21474, 2024.arxiv.org/abs/2410.21474
- [24]O. Feldman, A. Venugopal, J. Spiess, and A. Feder, “Causal effect estimation with latent textual treatments,” arXiv preprint arXiv:2602.15730, 2026.arxiv.org/abs/2602.15730
- [25]M. Neubrander, G. Tierney, and A. Volfovsky, “The confounder trap: Treatment-encoding representations in causal inference with text,” arXiv preprint arXiv:2607.26309, 2026.arxiv.org/abs/2607.26309
- [26]A. D’Amour and A. Franks, “Deconfounding scores: Feature representations for causal effect estimation with weak overlap,” arXiv preprint arXiv:2104.05762, 2021.arxiv.org/abs/2104.05762
- [27]A. Daoud et al., “Detecting and mitigating treatment leakage in text-based causal inference: Distillation and sensitivity analysis,” arXiv preprint arXiv:2601.02400, 2025.arxiv.org/abs/2601.02400
- [28]O. Clivio, A. D’Amour, A. Franks, D. Bruns-Smith, C. Holmes, and A. Feller, “Deconfounding scores and representation learning for causal effect estimation with weak overlap,” arXiv preprint arXiv:2604.00811, 2026.
- [29]M. I. Belghazi, A. Baratin, S. Rajeshwar, S. Ozair, Y. Bengio, A. Courville, and D. Hjelm, “Mutual information neural estimation,” in International Conference on Machine Learning (ICML), 2018, pp. 531–540.
- [30]A. Gretton, O. Bousquet, A. Smola, and B. Schölkopf, “Measuring statistical dependence with Hilbert-Schmidt norms,” International Conference on Algorithmic Learning Theory (ALT), pp. 63–77, 2005.
- [31]D. McAllester and K. Stratos, “Formal limitations on the measurement of mutual information,” in International Conference on Artificial Intelligence and Statistics (AISTATS), 2020, pp. 875–884.
- [32]V. Balazadeh, H. Kamkari, V. Thomas, B. Li, J. Ma, J. C. Cresswell, and R. G. Krishnan, “CausalPFN: Amortized causal effect estimation via in-context learning,” arXiv preprint arXiv:2506.07918, 2025.arxiv.org/abs/2506.07918
- [33]T. Kitagawa and A. Tetenov, “Who should be treated? empirical welfare maximization methods for treatment choice,” Econometrica, vol. 86, no. 2, pp. 591–616, 2018.
- [34]S. Athey and S. Wager, “Policy learning with observational data,” Econometrica, vol. 89, no. 1, pp. 133–161, 2021.arxiv.org/abs/1702.02896
- [35]M. Dudík, J. Langford, and L. Li, “Doubly robust policy evaluation and learning,” in International Conference on Machine Learning (ICML), 2011.arxiv.org/abs/1103.4601
- [36]F. Devriendt, T. Guns, and W. Verbeke, “Learning to rank for uplift modeling,” arXiv preprint arXiv:2002.05897, 2020.arxiv.org/abs/2002.05897
- [37]B. Bokelmann and S. Lessmann, “Improving uplift model evaluation on RCT data,” arXiv preprint arXiv:2210.02152, 2022.arxiv.org/abs/2210.02152
- [38]J. Crabbé, A. Curth, I. Bica, and M. van der Schaar, “Benchmarking heterogeneous treatment effect models through the lens of interpretability,” arXiv preprint arXiv:2206.08363, 2022.arxiv.org/abs/2206.08363
- [39]Z. Wood-Doughty, I. Shpitser, and M. Dredze, “Generating synthetic text data to evaluate causal inference methods,” arXiv preprint arXiv:2102.05638, 2021.
- [40]N. Tishby, F. C. Pereira, and W. Bialek, “The information bottleneck method,” arXiv preprint physics/0004057, 2000.
- [41]Q. Team, “Qwen2.5 technical report,” arXiv preprint arXiv:2412.15115, 2025.