Pith. sign in

REVIEW 3 major objections 4 minor 59 references

This paper claims that selective, role-conditioned LLM semantic augmentation, allocated by the estimator's own uncertainty, improves finite-sample CATE estimation across ten host learners on four benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 12:36 UTC pith:42EUEOPI

load-bearing objection CURL is a genuinely novel plug-in for CATE that is honestly scoped and broadly tested, but its central uncertainty-allocation premise rests on thin evidence and most gains are within noise. the 3 major comments →

arxiv 2607.26599 v1 pith:42EUEOPI submitted 2026-07-29 cs.LG

Uncertainty-Guided LLM Semantic Augmentation for Heterogeneous Treatment Effect Estimation

Jialu Xu , Mengkun Liang , Guannan Liu , Xiaojie Mao , Junjie Wu This is my paper
classification cs.LG
keywords heterogeneous treatment effectsCATElarge language modelsuncertainty-guided allocationMC dropoutrepresentation learningsemantic augmentationplug-in adapter
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that a common cause of poor finite-sample treatment-effect estimates is representational lossiness: the raw numeric and categorical encoding of pre-treatment covariates leaves semantic relations and higher-order interactions implicit, so even a correctly identified CATE is hard to learn. To fix it, the authors propose CURL, a plug-in adapter that uses the host estimator's own MC-dropout uncertainty—separately for the propensity component and the CATE component—to select the most locally unstable units, queries a frozen LLM with two role-conditioned prompts, and routes the resulting assignment- and heterogeneity-oriented embeddings into separated pathways of the host. On four benchmarks and ten host learners, CURL improves the host in most settings, and ablations indicate the gains come from the uncertainty-guided allocation and role-aware routing rather than from raw LLM capacity. The paper is explicit that CURL reorganizes observed covariates only; it does not address unmeasured confounding or overlap violations.

Core claim

CURL's central claim is that selective, role-conditioned semantic augmentation can reduce finite-sample CATE error without changing the observed information set, the estimand, or the identification assumptions. The mechanism: compute two uncertainty scores by MC dropout—propensity-side instability and CATE-side instability—sum their percentile ranks to choose a fixed fraction of units each round, build an assignment-oriented embedding and a heterogeneity-oriented embedding from a frozen LLM for each newly selected unit, add the first to the shared representation while concatenating the second into the effect pathway so it can never influence propensity prediction, then iterate diagnosis and

What carries the argument

The central object is the dual-channel, uncertainty-gated semantic adapter. Two MC-dropout standard deviations—s_e from the propensity head and s_tau from the CATE output—are converted to percentile ranks and summed to rank units (Eq. 3); the top rho fraction receives LLM embeddings z_A and z_H from separate role-conditioned prompts. z_A is added residually to the host's shared representation (Eq. 9); z_H is concatenated as a separate block consumed only by the effect operator (Eq. 10), so heterogeneity semantics cannot leak into assignment modeling. A cache ensures each unit is queried at most once, and progressive rounds recompute uncertainty and reallocate. The asymmetry of residual-add v

Load-bearing premise

The whole selection mechanism rests on MC-dropout predictive variance being a faithful, role-aligned measure of which units the host estimator is actually getting wrong; if that variance mostly reflects architecture noise or miscalibrated confidence rather than estimation error, the allocation collapses toward random augmentation and the gains would not replicate.

What would settle it

On a host with known ground-truth effects, compute the rank correlation between MC-dropout uncertainty (s_e, s_tau) and per-unit absolute CATE error across test units; if correlations are near zero, or if replacing the uncertainty ranking with random selection under the same 25% budget yields the same PEHE/AUQC on a new benchmark, the allocation mechanism is not doing the claimed work.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If CURL's claim holds, pretrained semantic knowledge becomes a general plug-in inductive bias: any CATE host—meta-learner, balanced-representation, or latent-variable—can be wrapped without changing its objective or identification assumptions.
  • The separation of assignment- and heterogeneity-oriented channels implies that injecting a single undifferentiated LLM embedding is the wrong design; future augmentation methods should match the representational channel to the statistical role it serves.
  • Because gains persist under training-set fractions from 30% to 100%, the benefit is not only a small-sample patch; semantic reorganization helps even with full benchmark samples.
  • The progressive-refinement dynamics—cache saturation and falling uncertainty—suggest the adapter's effect is to stabilize the estimator locally, which would make downstream uses such as individual-level targeting or treatment-effect ranking more reliable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to test whether the uncertainty signal itself is the active ingredient by swapping MC dropout for deep ensembles or conformal predictive intervals; if the gains persist, the principle generalizes beyond the specific uncertainty estimator.
  • The two-role decomposition suggests a design template for other conditional estimands—for example, conditional quantile treatment effects or heterogeneous policy value—where one channel carries adjustment structure and the other carries response heterogeneity.
  • The corruption experiments imply a boundary condition: CURL's benefit should shrink on datasets where feature names carry no semantics (e.g., anonymized or artificially encoded covariates), so deployment value depends on the semantic alignment between the LLM's pretraining and the domain's variables.
  • The paper leaves calibration under domain shift open; if test-time covariates fall outside the training uncertainty distribution, the fixed pointwise threshold in Eq. (16) may over- or under-augment, a testable risk when applying CURL to new populations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. Xu et al. introduce CURL, a plug-in adapter that wraps an arbitrary CATE estimator. The paper formalizes a notion of 'representational lossiness' — a finite-sample bottleneck in exploiting semantic relations left implicit by raw covariate encodings — and argues that local instability can be diagnosed via MC-dropout uncertainty of propensity and CATE predictions (Eq. 2). A fixed proportion ρ of units is selected for LLM-based semantic augmentation through percentile-ranked uncertainty scores (Eq. 3). Two frozen-LLM channels, assignment-oriented and heterogeneity-oriented, are fused asymmetrically (Eqs. 9–10) and routed so that heterogeneity information cannot reach the propensity head. Training proceeds in refinement rounds, and at test time a threshold calibrated on final training uncertainty determines augmentation pointwise (Eq. 16). On IHDP, Adult, Jobs, and Hillstrom, CURL is reported to improve ten host learners in most settings, with ablations, route-reassignment, probing, and semantic-corruption analyses supporting the channel design.

Significance. The empirical scope is substantial: ten host learners spanning meta-learners, balanced-representation methods, and deep latent-variable models; external held-out evaluation with semi-synthetic ground truth on IHDP/Adult and uplift metrics on Jobs/Hillstrom; and careful controls such as random-embedding corruption, anonymized feature names, and replayed query plans. If the central mechanism holds, CURL is a practical recipe for injecting pretrained semantic inductive bias into CATE estimation without changing the observed information set, estimand, or identifying assumptions. The paper does not ship machine-checked proofs or code, but the prompt templates, routing details, and instantiation tables are unusually complete. The main risks are evidentiary: the uncertainty-guided allocation is directly validated on only one host/configuration, many headline improvements are within one standard deviation over five seeds, and the test-time calibration rule does not obviously match the cumulative training-time augmentation rate. These are fixable within the manuscript's scope, but they are load-bearing for the paper's central attribution claims.

major comments (3)
  1. [§5.2, Tables 1 and 2] The central claim that CURL improves its host 'in most settings' rests on point estimates from five seeds. Several reported gains are much smaller than the standard deviations: Jobs CFRNet AUUC Δ=+0.38% (816.48±79.80 vs 819.61±123.81), Hillstrom R-Learner AUUC Δ=+1.07% (0.205±0.049 vs 0.207±0.049), Adult CFRNet PEHE Δ=+0.67% (0.947±0.270 vs 0.941±0.128), and Adult TEDVAE PEHE Δ=+1.63%. Since all methods use identical splits per seed, paired tests or paired confidence intervals are feasible and should be reported. Without them, the 'most settings' claim is not statistically established, and the later discussion overstates the consistency of the gains.
  2. [§4.2, Eq. (2), Fig. 1; §5.3, Fig. 3] The load-bearing premise is that s_e and s_tau are faithful, locally separable signals of assignment-side and heterogeneity-side estimation error across all ten hosts. Direct evidence is a single Q1–Q4 study on an S-Learner with an auxiliary propensity head on IHDP (Fig. 1). The random-allocation ablation in Fig. 3 covers only S/T/TARNet/CFRNet and compares the combined score with random selection; it does not measure whether MC-dropout variances correlate with local error in VAE-family hosts, where MC dropout is confounded with native stochasticity, or in balanced-representation hosts. This is not a circularity problem — the evaluation is external — but the attribution of the observed gains to uncertainty-guided allocation is under-supported. Please add per-host calibration diagnostics (e.g., binned local error vs. score correlation, or an oracle-selection comparison) or temper the caus
  3. [§4.4.2, Eq. (16)] The test-time criterion uses the (1−ρ) quantile κρ of the final training ω distribution, so under distributional similarity it will augment roughly ρ of test units. During training, however, because the top-ρ sets overlap across rounds, the cumulative cache |C(r)|/N grows beyond ρ; Fig. 10 reports final ratios of 51.2–82.5% after five rounds on IHDP/Jobs, and the ratio already exceeds ρ with the default R=3. The final wrapped estimator is therefore trained with a substantially larger augmented fraction than the test-time rule provides. The statement that 'the same ρ calibrates this threshold to the training-time selection budget' is true only for the per-round budget, not for the actual training-time augmentation coverage. Please either set κρ from the final cache ratio |C(R−1)|/N, or justify why the smaller test-time augmentation rate preserves the training-time input distribution.
minor comments (4)
  1. [§4.1] Typo in the heading: 'Role-A ware Prediction Routing' should be 'Role-Aware Prediction Routing'.
  2. [Figure 7] The legend repeats 'R7: Drop zA on T' and 'R8: Swap on Y' twice; one duplicate pair should be removed.
  3. [§5.4] The sensitivity analysis for the allocation ratio ρ and refinement rounds R is reported only on Jobs; a compact version on a second dataset would help establish that the chosen defaults are not dataset-specific.
  4. [§4.4.2] The sentence 'when the test uncertainty distribution is similar' is a real assumption; the paper should state explicitly that the main experiments are i.i.d. and that domain shift is left to future work (it is mentioned in the conclusion, but the assumption should also appear where Eq. (16) is introduced).

Circularity Check

0 steps flagged

No significant circularity: the claimed gains are evaluated on held-out targets and no prediction in the paper reduces by construction to a fitted input.

full rationale

The paper's central claim is an empirical one: CURL, as a plug-in adapter, improves ten host CATE estimators on four benchmarks. The evaluation is external — IHDP and Adult use ground-truth PEHE/epsilon_ATE/epsilon_ATT, and Jobs/Hillstrom use uplift metrics computed from held-out outcomes. No parameter in the method (uncertainty scores in Eq. (2), selection rule in Eq. (3), gates in Eq. (8), test-time threshold in Eq. (16), or objective weights in Eq. (12)) is fitted to any evaluation target. The MC-dropout variances are operational allocation signals computed from the wrapped estimator, not fitted values whose prediction is then reported as a result; the paper explicitly treats them as 'operational allocation signals rather than evidence of missing causal variables.' The motivating study in Figure 1 is correlational evidence that local instability separates into assignment- and heterogeneity-side channels; it is not an equation-level reduction of CATE-PEHE to the uncertainty scores. Ablations (random allocation, channel removal, route reassignment, semantic corruption) are independent empirical controls, not restatements of the construction. There is no load-bearing self-citation: the references to prior uncertainty-based causal-effect work and LLM estimators are external, and no uniqueness theorem or ansatz is imported from the authors' own prior work. The skeptic's concern — that the uncertainty signal's fidelity is validated only on IHDP/S-Learner and may not transfer to VAE-family hosts — is a correctness and external-validity concern, not circularity, because nothing in the derivation forces the outcome. Accordingly, the paper does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The method introduces no new identification assumptions beyond standard unconfoundedness/overlap. Its load-bearing premises are: (i) MC-dropout variance is a task-relevant instability signal; (ii) the signal separates into assignment vs heterogeneity channels; (iii) LLM hidden states of textualized covariates carry useful semantics; (iv) test uncertainty resembles training uncertainty. Only (iii) has dedicated in-paper evidence (corruption experiments, Appendix G); (i)–(ii) rest on the single IHDP/S-Learner motivating study plus internal ablations; (iv) is stated and untested. No physically invented entities; the two semantic channels are internal design constructs without independent evidence.

free parameters (6)
  • semantic-allocation ratio ρ = 0.25
    Per-round fraction of units selected for LLM semantic augmentation (Eq. 3); stated in §5.1 as part of the configuration, with sensitivity shown only on Jobs (Figure 5a).
  • MC-dropout passes M = 30
    Number of stochastic forward passes for the uncertainty scores (Eq. 2); fixed in §5.1 without calibration analysis.
  • refinement rounds R = 3
    Number of diagnosis–allocation–integration cycles; sensitivity on Jobs (Figure 5b) but not on the other three benchmarks.
  • auxiliary loss weights λ_prop, λ_ortho = 1.0, 0.1
    Weights in the objective (Eq. 12); selected on validation; no sensitivity analysis reported for these weights.
  • test-time augmentation threshold κ_ρ = empirical (1−ρ)-quantile of training ω_i
    Converts the relative training ranking into a pointwise rule (Eq. 16); calibrated to the training uncertainty distribution under an assumed training/test similarity.
  • LLM backbone per dataset = Qwen2.5-7B or Llama-3-8B (six backbones compared)
    Backbone selected per benchmark by validation (§5.1, Appendix A.2); dataset-dependent choice adds a model-selection degree of freedom not available to fixed-configuration baselines.
axioms (5)
  • domain assumption Conditional unconfoundedness {Y(0),Y(1)} ⊥ T | X and overlap ε ≤ e(x) ≤ 1−ε
    Stated at the start of §3; identifies τ(x)=μ1(x)−μ0(x). Standard for CATE and explicitly scoped out of the paper's contribution.
  • domain assumption MC-dropout predictive variance (Eq. 2) reflects task-relevant local estimation instability
    The allocation signal of the whole method; supported only by the Figure 1 IHDP/S-Learner motivating study and by downstream ablations, not by an independent calibration analysis.
  • ad hoc to paper Assignment-side and heterogeneity-side instability are partially separable and remain role-aligned across all ten hosts
    Figure 1 shows Q2/Q3/Q4 separation for one host on IHDP; §4.2 zeroes s_e in round 0 for hosts without a native propensity head, so the separation premise is only partially exercised on other architectures.
  • ad hoc to paper Test-time uncertainty distribution resembles the training uncertainty distribution
    §4.4.2 (Eq. 16) applies final-training percentiles and κ_ρ pointwise; the text concedes the rule 'gives a comparable augmentation rate when the test uncertainty distribution is similar' — an untested distributional assumption.
  • domain assumption LLM hidden states of textualized covariates carry transferable covariate semantics without leaking outcome or test information
    Supported by the Appendix G corruption experiments (anonymized/shuffled/random all degrade relative to CURL). The no-leak property holds by construction of ψ (Appendix C.3), though LLM pretraining could still echo dataset-level associations.
invented entities (2)
  • Assignment-oriented semantic channel z_A no independent evidence
    purpose: LLM-derived representation injected as a residual correction into the shared encoder state feeding the propensity head (Eq. 9).
    A method construct defined by the adapter itself; its 'assignment role' is validated only by in-paper probe and route-reassignment analyses (Figures 7–8), with no externally falsifiable handle.
  • Heterogeneity-oriented semantic channel z_H no independent evidence
    purpose: LLM-derived representation concatenated as a separate feature block available only to the CATE operator (Eq. 10).
    Same status as z_A: the role is enforced by the routing design rather than by any externally measurable property; the orthogonality penalty (Eq. 14) partially discourages redundancy between the two channels.

pith-pipeline@v1.3.0-daily-deepseek · 5334 in / 5948 out tokens · 226115 ms · 2026-08-01T12:36:12.071896+00:00 · methodology

0 comments
read the original abstract

Estimating heterogeneous treatment effects is central to targeted interventions, such as personalized promotions and precision medicine. We focus on the conditional average treatment effect (CATE), a standard estimand for characterizing such heterogeneity. Even under standard identification conditions, finite-sample CATE estimation requires learning the nuisance structure for covariate adjustment and treatment-effect heterogeneity, often together with an effective representation of X. Raw numerical and categorical encodings can leave semantic relations and higher-order interactions implicit, making this joint task locally unstable. A motivating study further shows that this instability appears through partially separable assignment- and heterogeneity-side channels. Building on this observation, we propose CURL (Causal Uncertainty-guided Representation Learning), a plug-in adapter that uses estimator uncertainty to allocate pretrained semantic capacity to locally unstable units. CURL queries a frozen LLM through two role-conditioned prompts, constructs assignment- and heterogeneity-oriented representations from the observed covariates, and routes them through separated pathways. On four benchmarks, CURL improves ten host learners in most settings, while ablation, refinement-dynamics, route-reassignment, and probe analyses support the intended design and roles of the two channels.

Figures

Figures reproduced from arXiv: 2607.26599 by Guannan Liu, Jialu Xu, Junjie Wu, Mengkun Liang, Xiaojie Mao.

Figure 1
Figure 1. Figure 1: Motivating study on IHDP using an S-Learner with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of CURL. Uncertainty-guided diagnosis identifies samples that receive semantic augmentation; [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation results across four host learners on IHDP [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity of CURL to different LLM backbones on Hillstrom (AUQC) and IHDP (PEHE) across multiple host learners. Overall, CURL is not highly sensitive to either hyperparameter around the selected configuration. 5.5 Progressive Refinement Dynamics We further examine how semantic allocation evolves across refine￾ment rounds. Setting 𝑅 = 5, we track the cumulative semantic￾cache ratio |C(𝑟) |/𝑁 and the mean … view at source ↗
Figure 7
Figure 7. Figure 7: Route-reassignment validation. The original rout [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Probe-based role validation. The assignment [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Representative CURL instantiations for (a) S-Learner, (b) TARNet, (c) R-Learner, and (d) TEDVAE. Across hosts, the propensity pathway consumes 𝑢𝐴, whereas the host-specific effect operator consumes 𝑢𝐻 . The insertion point varies with the host architecture, while the role-aware interface remains unchanged. Meanwhile, both 𝑠 (𝑟) 𝑒 and 𝑠 (𝑟) 𝜏 exhibit an overall downward trend, with only minor intermediate f… view at source ↗
Figure 10
Figure 10. Figure 10: Progressive-refinement dynamics on IHDP and Jobs. For each host, the left panel reports the cumulative semantic [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional validation of representational lossi [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Walkthrough of a representative locally unstable Hillstrom unit. The observed profile, aggregate covariate-shift hint, [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 1 canonical work pages

  1. [1]

    Improved algorithms for linear stochastic bandits

    Yasin Abbasi-Yadkori, Dávid Pál, and Csaba Szepesvári. Improved algorithms for linear stochastic bandits. InAdvances in Neural Information Processing Systems, volume 24, pages 2312–2320, 2011

    2011

  2. [2]

    Alaa and Mihaela van der Schaar

    Ahmed M. Alaa and Mihaela van der Schaar. Bayesian inference of individualized treatment effects using multi-task Gaussian processes. InAdvances in Neural Information Processing Systems, 2017

    2017

  3. [3]

    Uncertainty- based offline reinforcement learning with diversified Q-ensemble

    Gaon An, Seungyong Moon, Jang-Hyun Kim, and Hyun Oh Song. Uncertainty- based offline reinforcement learning with diversified Q-ensemble. InAdvances in Neural Information Processing Systems, volume 34, pages 7436–7447, 2021

    2021

  4. [4]

    Pessimistic bootstrapping for uncertainty-driven offline reinforcement learning.arXiv preprint arXiv:2202.11566, 2022

    Chenjia Bai, Lingxiao Wang, Zhuoran Yang, Zhihong Deng, Animesh Garg, Peng Liu, and Zhaoran Wang. Pessimistic bootstrapping for uncertainty-driven offline reinforcement learning.arXiv preprint arXiv:2202.11566, 2022

    Pith/arXiv arXiv 2022

  5. [5]

    LLM-driven causal discovery via harmonized prior.IEEE Transactions on Knowledge and Data Engineering, 2025

    Taiyu Ban, Lyuzhou Chen, Derui Lyu, Xiangyu Wang, Qinrui Zhu, and Huanhuan Chen. LLM-driven causal discovery via harmonized prior.IEEE Transactions on Knowledge and Data Engineering, 2025

    2025

  6. [6]

    Weight uncertainty in neural networks

    Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. InInternational Conference on Machine Learning, pages 1613–1622. PMLR, 2015

    2015

  7. [7]

    Chen, Rohit Bhattacharya, and Katherine A

    Jacob M. Chen, Rohit Bhattacharya, and Katherine A. Keith. Proximal causal inference with text data. InAdvances in Neural Information Processing Systems, volume 37, 2024

    2024

  8. [8]

    Unveiling causal reasoning in large language models: Reality or mirage? InAdvances in Neural Information Processing Systems, volume 37, 2024

    Haoang Chi, He Li, Wenjing Yang, Feng Liu, Long Lan, Xiaoguang Ren, Tongliang Liu, and Bo Han. Unveiling causal reasoning in large language models: Reality or mirage? InAdvances in Neural Information Processing Systems, volume 37, 2024

    2024

  9. [9]

    Krishnan, and Chris J

    Nikita Dhawan, Leonardo Cotta, Karen Ullrich, Rahul G. Krishnan, and Chris J. Maddison. End-to-end causal effect estimation from unstructured natural lan- guage data. InAdvances in Neural Information Processing Systems, volume 37, 2024

    2024

  10. [10]

    Causal discovery through synergizing large language model and data-driven reasoning

    Huaming Du, Yujia Zheng, Baoyu Jing, Yu Zhao, Gang Kou, Guisong Liu, Tao Gu, Weimin Li, and Carl Yang. Causal discovery through synergizing large language model and data-driven reasoning. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, 2025

    2025

  11. [11]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Maria Florina Balcan and Kilian Q. Weinberger, editors,Proceedings of The 33rd International Conference on Machine Learning, volume 48 ofProceedings of Machine Learning Research, pages 1050–1059, New York, New York, USA, 20–22 Jun 2016...

    2016

  12. [12]

    Learning disentangled representations for counterfactual regression

    Negar Hassanpour and Russell Greiner. Learning disentangled representations for counterfactual regression. InInternational Conference on Learning Represen- tations, 2020

    2020

  13. [13]

    Improving treatment effect estimation with llm-based data augmentation

    Nicolas Huynh, Julianna Piskorz, Jeroen Berrevoets, Max Ruiz Luyten, and Mi- haela van der Schaar. Improving treatment effect estimation with llm-based data augmentation. In1st ICML Workshop on Foundation Models for Structured Data, 2025

    2025

  14. [14]

    Imbens and Donald B

    Guido W. Imbens and Donald B. Rubin.Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. Cambridge University Press, New York, NY, USA, 2015. ISBN 9780521885881

    2015

  15. [15]

    Identifying causal- effect inference failure with uncertainty-aware models.Advances in Neural Information Processing Systems, 33:11637–11649, 2020

    Andrew Jesson, Sören Mindermann, Uri Shalit, and Yarin Gal. Identifying causal- effect inference failure with uncertainty-aware models.Advances in Neural Information Processing Systems, 33:11637–11649, 2020

    2020

  16. [16]

    Cladder: Assessing causal reasoning in language models

    Zhijing Jin, Yuen Chen, Felix Leeb, Luigi Gresele, Ojasv Kamal, Zhiheng LYU, Kevin Blin, Fernando Gonzalez Adauto, Max Kleiman-Weiner, Mrinmaya Sachan, and Bernhard Schölkopf. Cladder: Assessing causal reasoning in language models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Sy...

    2023

  17. [17]

    What uncertainties do we need in bayesian deep learning for computer vision?Advances in neural information processing systems, 30, 2017

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision?Advances in neural information processing systems, 30, 2017

    2017

  18. [18]

    Varia- tional autoencoders and nonlinear ica: A unifying framework

    Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Varia- tional autoencoders and nonlinear ica: A unifying framework. In Silvia Chiappa and Roberto Calandra, editors,Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, volume 108 ofProceedings of Machine Learning Research, pages 2207–2217. ...

    2020

  19. [19]

    Metalearners for estimating heterogeneous treatment effects using machine learning.Proceedings of the National Academy of Sciences, 116(10):4156–4165, 2019

    Sören R Künzel, Jasjeet S Sekhon, Peter J Bickel, and Bin Yu. Metalearners for estimating heterogeneous treatment effects using machine learning.Proceedings of the National Academy of Sciences, 116(10):4156–4165, 2019

    2019

  20. [20]

    Causal reasoning and large language models: Opening a new frontier for causality

    Emre Kıcıman, Robert Ness, Amit Sharma, and Chenhao Tan. Causal reasoning and large language models: Opening a new frontier for causality. 2024. URL https://arxiv.org/abs/2305.00050

    Pith/arXiv arXiv 2024

  21. [21]

    Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in neural information processing systems, 30, 2017

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in neural information processing systems, 30, 2017

    2017

  22. [22]

    Self-distilled disentangled learning for counterfactual prediction

    Xinshu Li, Mingming Gong, and Lina Yao. Self-distilled disentangled learning for counterfactual prediction. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1667–1678, 2024

    2024

  23. [23]

    Large language models and causal inference in collaboration: A comprehensive survey

    Xiaoyu Liu, Paiheng Xu, Junda Wu, Jiaxin Yuan, Yifan Yang, Yuhang Zhou, Fuxiao Liu, Tianrui Guan, Haoliang Wang, Tong Yu, Julian McAuley, Wei Ai, and Furong Huang. Large language models and causal inference in collaboration: A comprehensive survey. In Luis Chiruzzo, Alan Ritter, and Lu Wang, editors, Findings of the Association for Computational Linguisti...

    2025

  24. [24]

    Causal effect inference with deep latent-variable models

    Christos Louizos, Uri Shalit, Joris Mooij, David Sontag, Richard Zemel, and Max Welling. Causal effect inference with deep latent-variable models. InProceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 6449–6459, Red Hook, NY, USA, 2017. Curran Associates Inc. ISBN 9781510860964

    2017

  25. [25]

    David J.C. MacKay. A practical Bayesian framework for backpropagation net- works.Neural Computation, 4(3):448–472, 1992

    1992

  26. [26]

    Deep disentangled represen- tation network for treatment effect estimation.arXiv preprint arXiv:2507.06650, 2025

    Hui Meng, Keping Yang, Xuyu Peng, and Bo Zheng. Deep disentangled represen- tation network for treatment effect estimation.arXiv preprint arXiv:2507.06650, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Quasi-oracle estimation of heterogeneous treat- ment effects.Biometrika, 108(2):299–319, 2021

    Xinkun Nie and Stefan Wager. Quasi-oracle estimation of heterogeneous treat- ment effects.Biometrika, 108(2):299–319, 2021. doi: 10.1093/biomet/asaa076

    doi:10.1093/biomet/asaa076 2021

  28. [28]

    A critical look at the consistency of causal estimation with deep latent variable models

    Severi Rissanen and Pekka Marttinen. A critical look at the consistency of causal estimation with deep latent variable models. InAdvances in Neural Information Processing Systems, volume 34, 2021. 9 Jialu Xu, Mengkun Liang, Guannan Liu, Xiaojie Mao, and Junjie Wu

    2021

  29. [29]

    Rosenbaum and Donald B

    Paul R. Rosenbaum and Donald B. Rubin. The central role of the propensity score in observational studies for causal effects.Biometrika, 70(1):41–55, 1983. doi: 10.1093/biomet/70.1.41

    doi:10.1093/biomet/70.1.41 1983

  30. [30]

    Estimating individual treat- ment effect: generalization bounds and algorithms

    Uri Shalit, Fredrik D Johansson, and David Sontag. Estimating individual treat- ment effect: generalization bounds and algorithms. InProceedings of the 34th International Conference on Machine Learning, pages 3076–3085. PMLR, 2017

    2017

  31. [31]

    Adapting neural networks for the estimation of treatment effects

    Claudia Shi, David Blei, and Victor Veitch. Adapting neural networks for the estimation of treatment effects. InAdvances in Neural Information Processing Systems, volume 32, 2019

    2019

  32. [32]

    Improving data-driven heterogeneous treatment effect estimation under structure uncertainty

    Christopher Tran and Elena Zheleva. Improving data-driven heterogeneous treatment effect estimation under structure uncertainty. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1787–

  33. [33]

    DoubleLingo: Causal estimation with large language models

    Marko Veljanovski and Zach Wood-Doughty. DoubleLingo: Causal estimation with large language models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies (Volume 2: Short Papers), pages 799–807. ACL, 2024

    2024

  34. [34]

    CE-RCFR: Robust counterfactual regression for consensus-enabled treatment effect estimation

    Fan Wang, Chaochao Chen, Weiming Liu, Tianhao Fan, Xinting Liao, Yanchao Tan, Lianyong Qi, and Xiaolin Zheng. CE-RCFR: Robust counterfactual regression for consensus-enabled treatment effect estimation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3013–3023. Association for Computing Machinery, 2024. doi: 1...

    arXiv 2024

  35. [35]

    Treatment effect estimation with adjustment feature selection

    Haotian Wang, Kun Kuang, Haoang Chi, Longqi Yang, Mingyang Geng, Wan- rong Huang, and Wenjing Yang. Treatment effect estimation with adjustment feature selection. InProceedings of the 29th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining, pages 2290–2301. Association for Computing Machinery, 2023. doi: 10.1145/3580305.3599531

    arXiv 2023

  36. [36]

    Progressive generalization risk reduction for data-efficient causal effect estimation

    Hechuan Wen, Tong Chen, Guanhua Ye, Li Kheng Chai, Shazia Sadiq, and Hongzhi Yin. Progressive generalization risk reduction for data-efficient causal effect estimation. InProceedings of the 31st ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining, Volume 1, pages 1575–1586. Association for Computing Machinery, 2025. doi: 10.1145/3690624.3709305

    arXiv 2025

  37. [37]

    From supervised to generative: A novel paradigm for tabular deep learning with large language models

    Xumeng Wen, Han Zhang, Shun Zheng, Wei Xu, and Jiang Bian. From supervised to generative: A novel paradigm for tabular deep learning with large language models. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discov- ery and Data Mining, pages 3323–3333. Association for Computing Machinery,

  38. [38]

    Learning decomposed representations for treatment effect estimation.IEEE Transactions on Knowledge and Data Engineering, 35(5):4989– 5001, 2023

    Anpeng Wu, Junkun Yuan, Kun Kuang, Bo Li, Runze Wu, Qiang Zhu, Yueting Zhuang, and Fei Wu. Learning decomposed representations for treatment effect estimation.IEEE Transactions on Knowledge and Data Engineering, 35(5):4989– 5001, 2023. doi: 10.1109/TKDE.2022.3150807

    arXiv 2023

  39. [39]

    𝛽-intact-VAE: Identifying and estimat- ing causal effects under limited overlap

    Pengzhou Abel Wu and Kenji Fukumizu. 𝛽-intact-VAE: Identifying and estimat- ing causal effects under limited overlap. InInternational Conference on Learning Representations, 2022

    2022

  40. [40]

    Uncertainty weighted actor-critic for offline reinforcement learning.arXiv preprint arXiv:2105.08140, 2021

    Yue Wu, Shuangfei Zhai, Nitish Srivastava, Joshua Susskind, Jian Zhang, Ruslan Salakhutdinov, and Hanlin Goh. Uncertainty weighted actor-critic for offline reinforcement learning.arXiv preprint arXiv:2105.08140, 2021

    Pith/arXiv arXiv 2021

  41. [41]

    Neural contextual bandits with deep representation and shallow exploration

    Pan Xu, Zheng Wen, Handong Zhao, and Quanquan Gu. Neural contextual bandits with deep representation and shallow exploration. InInternational Conference on Learning Representations, 2022. URL https://openreview.net/for um?id=xnYACQquaGV

    2022

  42. [42]

    Causal inference with conditional front-door adjustment and identifiable variational autoencoder

    Ziqi Xu, Debo Cheng, Jiuyong Li, Jixue Liu, Lin Liu, and Kui Yu. Causal inference with conditional front-door adjustment and identifiable variational autoencoder. In B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun, editors, International Conference on Learning Representations, volume 2024, pages 18692– 18713, 2024. URL https://proceeding...

    2024

  43. [43]

    Treatment effect estimation with dis- entangled latent factors.Proceedings of the AAAI Conference on Artificial Intel- ligence, 35(12):10923–10930, May 2021

    Weijia Zhang, Lin Liu, and Jiuyong Li. Treatment effect estimation with dis- entangled latent factors.Proceedings of the AAAI Conference on Artificial Intel- ligence, 35(12):10923–10930, May 2021. doi: 10.1609/aaai.v35i12.17304. URL https://ojs.aaai.org/index.php/AAAI/article/view/17304

    work page doi:10.1609/aaai.v35i12.17304 2021

  44. [44]

    Lui, and Hang Li

    Xiaoying Zhang, Junpu Chen, Hongning Wang, Hong Xie, Yang Liu, John C.S. Lui, and Hang Li. Uncertainty-aware instance reweighting for off-policy learning. InAdvances in Neural Information Processing Systems, volume 36, 2023

    2023

  45. [45]

    DESCN: Deep entire space cross networks for individual treatment effect estimation

    Kailiang Zhong, Fengtong Xiao, Yan Ren, Yaorong Liang, Wenqing Yao, Xiaofeng Yang, and Ling Cen. DESCN: Deep entire space cross networks for individual treatment effect estimation. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4612–4620. Association for Computing Machinery, 2022. doi: 10.1145/3534678.3539198

    arXiv 2022

  46. [46]

    Neural contextual bandits with UCB-based exploration

    Dongruo Zhou, Lihong Li, and Quanquan Gu. Neural contextual bandits with UCB-based exploration. InInternational Conference on Machine Learning, pages 11492–11502. PMLR, 2020

    2020

  47. [47]

    Causal effect estimation with mixed latent confounders and post-treatment variables

    Yaochen Zhu, Jing Ma, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. Causal effect estimation with mixed latent confounders and post-treatment variables. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=qe1CsfnN1W. A Dataset and Implementation Details This appendix provides additional benc...

    2025

  48. [50]

    Socio-economic constraints rooted in the past

  49. [51]

    Fundamental health/financial status established prior to this event

  50. [52]

    [Goal] Compress these pre-existing confounders into a single representation

    Historical behavioral patterns. [Goal] Compress these pre-existing confounders into a single representation. [Trigger] Pre-treatment Confounding Context: Heterogeneity-oriented prompt𝑃 𝐻 for IHDP, Jobs, and Hillstrom. Prompt𝐻 : General Response-Side Heterogeneity Inference [System] You are a Causal Analyst specializing in effect heterogeneity. [Constraint...

  51. [53]

    12 Uncertainty-Guided LLM Semantic Augmentation for Heterogeneous Treatment Effect Estimation

    Psychological resilience or sensitivity (Stable Personality). 12 Uncertainty-Guided LLM Semantic Augmentation for Heterogeneous Treatment Effect Estimation

  52. [54]

    Biological or structural receptivity

  53. [55]

    [Goal] Compress these heterogeneity modifiers into a single representation

    Risk tolerance levels. [Goal] Compress these heterogeneity modifiers into a single representation. [Trigger] Effect Modification Context: Specialized𝑃𝐻 for Adult.Adult uses a domain-specific variant because its simulated heterogeneity is primarily associated with employment structure, education, and socio-economic mobility. Prompt𝐻 for Adult: Socio-Econom...

  54. [56]

    Social mobility potential based on their current class and education

  55. [57]

    [Goal] Compress these heterogeneity modifiers into a single representation

    Human capital accumulation potential. [Goal] Compress these heterogeneity modifiers into a single representation. [Trigger] Socio-economic Effect Modification Context Semantic extraction.For each prompt, the frozen LLM first gen- erates a deterministic textual response. We then perform a second forward pass over the concatenated prompt and generated respo...

    1974

  56. [58]

    This suggests a certain socio-economic standing but does not provide detailed information about income, education, or employment status

    Socio-Economic Constraints: - Historical Value Segment: The customer is categorized in the '6) $750 - $1,000' segment, indicating a moderate level of historical value with the brand. This suggests a certain socio-economic standing but does not provide detailed information about income, education, or employment status. - Location: The customer is located i...

  57. [59]

    However, as a new customer, they may still be relatively sensitive to new brand strategies or product changes

    Psychological resilience or sensitivity - Psychological resilience: Since this customer’s most recent interaction with the brand occurred 9 months ago, this may suggest that they have some capacity to adapt to changes related to the brand. However, as a new customer, they may still be relatively sensitive to new brand strategies or product changes. - Sens...

  58. [1797]

    doi: 10.1145/3534678.3539444

    Association for Computing Machinery, 2022. doi: 10.1145/3534678.3539444

    arXiv 2022

  59. [2024]

    doi: 10.1145/3637528.3671975

    arXiv