Pith. sign in

REVIEW 4 major objections 4 minor 59 references

Mitigating Hidden Confounding by Progressive Confounder Imputation via Large Language Models

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that an LLM can progressively generate and impute hidden confounders until a conditional-independence test passes, at which point standard estimators recover treatment effects without assuming unconfoundedness.

desk verdict A real idea with a load-bearing proof bug and a circular validation test; the recipe is worth engaging but the claims as stated don't hold. read the letter →

arxiv 2507.02928 v1 pith:RLAXLTCE submitted 2025-06-26 cs.CL cs.AI

classification cs.CLcs.AI
keywords causalinferencetreatmenteffectestimationhiddenconfoundinglargelanguagemodelsconfounderimputationunconfoundednesskernelconditionalindependencetestcounterfactualreasoning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

ProCI is a framework for treatment-effect estimation from observational data when some confounders are unobserved. It asks a large language model to propose a missing confounder, impute its values by first naming a distribution family and then estimating unit-level parameters, and then test whether conditioning on the augmented covariate set restores conditional independence between treatment and the potential outcomes, using counterfactual outcomes imputed by the same LLM. The generation-and-test loop repeats until a kernel-based conditional independence test no longer rejects independence, at which point ProCI hands the augmented dataset to any standard CATE estimator. The paper's thesis is that this procedure constructs a sufficient adjustment set without observing the true confounders, and its experiments report consistent improvements across four LLMs and several base estimators on the Twins and Jobs benchmarks. If true, it would make the unconfoundedness assumption replaceable by a semantic, knowledge-driven search for missing confounders.

What carries the argument

The load-bearing object is the progressive confounder-imputation loop together with the imputation-based KCIT validation. KCIT (kernel-based conditional independence test) computes a test statistic from kernel matrices of the variables; the paper's Theorem 1 asserts that when LLM imputations are accurate enough, the KCIT statistic on imputed variables converges in probability to the KCIT statistic on the true variables: $\mathrm{KCIT}((\hat Y^0,\hat Y^1), T \mid X,\hat U) = \mathrm{KCIT}((Y^0,Y^1), T \mid X,U) + o_p(1)$. This identity is what converts a heuristic 'did the LLM find a confounder?' step into a stopping rule: the loop terminates exactly when the generated set $X^{(k)}$ passes the test, which the theorem says is asymptotically the same as conditioning on the true hidden confounders. The other piece is distributional reasoning, which avoids collapsed LLM outputs by sampling imputed values from an LLM-chosen distribution family (e.g., Bernoulli or Gaussian) with LLM-inferred per-unit parameters rather than asking for direct point values.

What would settle it

A concrete check: on a synthetic dataset where a hidden confounder genuinely affects both T and Y but is semantically undetectable from the observed features (e.g., a random hidden label with no lexical cue), run ProCI. If the KCIT-based loop terminates with a passing test while the imputed confounders are uncorrelated with the true hidden variable and the estimated ATE remains biased, then passing the test would not certify a sufficient adjustment set.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that hidden confounding can be mitigated by treating the LLM as a confounder-discovery engine with a built-in validity check. Given observed covariates $X$, treatment $T$, and outcome $Y$, ProCI prompts the LLM to name a new confounder $\hat U$ that plausibly affects both $T$ and $Y$, imputes individual values via distributional reasoning (identify the distribution family, then infer per-unit parameters and sample), and appends it to form $X^{(k+1)}$. It then prompts the LLM to impute the missing potential outcomes $\hat Y^0,\hat Y^1$ and runs KCIT on $(\hat Y^0,\hat Y^1) \perp\!\!\perp T \mid X^{(k+1)}$; if the test rejects, another confounder is generated, and the process stops at the first $k$ that passes. The paper proves that, under kernel and imputation-accuracy conditions, the KCIT statistic computed from imputed variables equals the KCIT statistic computed from the true latent variables plus $o_p(1)$, which is the formal basis for treating a non-rejection as evidence that the adjustment set is sufficient. The reported experiments then show that augmenting five base estimators with these generated confounders improves in-sample and out-of-sample ATT/ATE/PEHE metrics on Jobs and Twins, with representation-based estimators gaining most.

Load-bearing premise

The load-bearing premise is that the LLM's imputed confounders and counterfactual outcomes are accurate enough that the kernel matrices in the test converge to those of the true variables (so the KCIT statistic on imputed variables tracks the test on true hidden variables), and the authors' own limitations section adds that this test is only asymptotically valid for large samples.

Editorial extensions

If this is right

  • Observational studies can relax the unconfoundedness assumption: instead of requiring all confounders to be measured, a practitioner can let an LLM propose and impute plausible missing confounders until the empirical independence test passes.
  • Because ProCI is model-agnostic, any existing CATE estimator—meta-learners, propensity-score matching, or representation-based networks—can be augmented, and the paper reports gains across all three families.
  • The stopping rule gives a principled defense against both under-adjustment and over-adjustment: it stops at the smallest covariate set that restores conditional independence, avoiding redundant overlapping confounders.
  • The approach is not tied to a single proprietary model: GPT-4o, DeepSeek-R1, LLaMA-3-8B, and Qwen2.5-7B all improve estimates, with the strongest-reasoning model giving the largest corrections.
  • Generated confounders carry information beyond the observed covariates, as shown by conditional mutual information values orders of magnitude above a random baseline, so the performance gains are not merely added noise.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A corollary the authors leave implicit: the method's ceiling is the LLM's knowledge coverage—a true confounder that is neither lexically nameable nor inferable from the observed variables would not be generated, and the KCIT pass could then certify an insufficient set; a separate audit would be needed to measure this coverage gap.
  • The distributional-reasoning trick is a general remedy for collapsed LLM imputation and could transfer to other missing-data and data-augmentation tasks, not just confounder imputation.
  • Because the same LLM generates confounders and imputes counterfactual outcomes, any systematic bias in the LLM's counterfactual predictions is shared by both the adjustment set and the validation test; evaluating ProCI on datasets with ground-truth counterfactuals (like Twins) partially isolates this, but a fully neutral test would need human or domain-model counterfactuals.
  • The same loop could be applied to continuous or multi-valued treatments and to treatments defined in unstructured text, where the semantic advantage of an LLM over tabular-only methods should be largest.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. This manuscript proposes ProCI, an iterative LLM-based procedure for mitigating hidden confounding in observational CATE estimation. In Phase 1, an LLM generates a candidate hidden confounder and imputes unit-level values through distribution identification and parameter inference; in Phase 2, the same LLM imputes counterfactual outcomes and a kernel conditional independence test (KCIT) decides whether the augmented covariate set is sufficient. The procedure iterates until KCIT passes, then feeds the augmented covariates into standard CATE estimators such as S-Learner, PSM, TARNet, CFR-Wass, and ESCFR. The authors claim asymptotic validity via Theorem 1 and report improvements on the Jobs and Twins benchmarks across four LLMs (GPT-4o, DeepSeek-R1, LLaMA 3-8B, Qwen2.5-7B).

Significance. The idea of using LLM world knowledge to propose latent confounders is timely and practically motivated, and the paper contains useful prompt templates and experiments with several LLM backbones. However, the central methodological contribution—the imputation-based unconfoundedness test and its asymptotic justification—is not supported. The proof of Theorem 1 has a scaling error, the key assumptions are unverified and partly circular, and the headline empirical claim of consistent improvement is contradicted by Table 1. Therefore the paper does not meet the publication bar in its current form.

major comments (4)
  1. [Appendix B, Lemma 2 and Eq. (16)] The proof of Lemma 2 contains a scaling error that invalidates the o_p(1) conclusions in Eq. (16). From Lemma 1 the authors obtain ||K̃_Ŷ - E K̃_Ŷ|| = o_p(n), and the lemma's assumption gives ||E K̃_Ŷ - K̃_Y|| = o_p(1), so ||K̃_Ŷ - K̃_Y|| = o_p(n). Since the bounded RBF kernel matrix satisfies ||K̃_X|| = O(n) in both operator and Frobenius norms, the bound (1/n)||K̃_X|| · ||K̃_Ŷ - K̃_Y|| is o_p(n), not o_p(1). The same issue affects the other displays in Eq. (16) and therefore the inverse-matrix conclusion and Theorem 1. The stated o_p(1) convergence would require ||E K̃_Ŷ - K̃_Y|| = o_p(1/n) or a normalization that is not present; neither follows from the LLM prompting scheme in Eqs. (4)-(7). Consequently, the asymptotic validity of the imputation-based unconfoundedness test is not established.
  2. [Section 3.4, Rationality Analysis and Eqs. (7)-(9)] The validation phase is circular. The KCIT in Eq. (9) tests whether (Ŷ0, Ŷ1) ⊥⊥ T | X, Û using counterfactual outcomes imputed by the same LLM (Eq. (7)) that generated Û (Eqs. (4)-(6)). The Rationality Analysis posits that the LLM implicitly conditions on a latent U* satisfying (Y0,Y1) ⊥⊥ T | X, U*, which is exactly the existence of a sufficient adjustment set that ProCI claims to discover. The pass/fail outcome can therefore reflect the LLM's internal consistency rather than true unconfoundedness, and no independent evidence is given that the imputed potential outcomes are accurate enough for the test to be meaningful.
  3. [Table 1, Section 4.2] Table 1 contradicts the claim in Section 4.2 that 'ProCI-augmented estimators consistently outperform base models.' On Jobs, S-Learner+ProCI-4o worsens in-sample ϵ_ATT from 0.0491±0.0011 to 0.0793±0.0043, and TARNet+ProCI-4o worsens both in-sample ϵ_ATT (0.0424±0.0023 vs 0.0191±0.0002) and out-sample ϵ_ATT (0.3223±0.0098 vs 0.1466±0.0026). The 'consistently' claim is therefore not supported even for one of the two main LLM backbones, and the paper offers no explanation for these degradations.
  4. [Section 4.3, Figure 3] The CMI comparison does not establish that imputed U captures confounding information. Because the value-imputation prompt in Eq. (6) and Appendix E.4 provides the LLM with each unit's T and Y values, Uhat is generated as a function of T and Y, so high I(U,T|X) and I(U,Y|X) can be an artifact of that construction rather than evidence that Uhat is a confounder. Comparing against random noise does not control for this.
minor comments (4)
  1. [Section 3.4, Eq. (9)] The notation 'I[KCIT(...) > α] = 1' defines a decision rule rather than a test result; the p-value threshold should be stated separately and the decision rule written as an explicit step.
  2. [Appendix B] The theorem proved in the appendix is labeled Theorem 2 but is referred to as Theorem 1 in the main text; the numbering should be aligned.
  3. [Section 4.1 and Appendix D.2] Section 4.1 fixes the LLM temperature at 0.7, while Appendix D.2 reports temperature sweeps; the relationship between the tuned temperature and the fixed experimental setting should be clarified.
  4. [Section 4, Eq. (11)] The KCIT significance level α and the maximum number of generated confounders are not reported, even though the stopping rule in Eq. (11) depends on them; the reader cannot assess how the termination criterion was applied.

Circularity Check

3 steps flagged · score 6.0 of 10

ProCI's 'empirical unconfoundedness test' validates the LLM's own imputed counterfactuals under a posited sufficient latent U*, so the sufficiency claim is partly circular; Theorem 1's proof also assumes the kernel consistency it needs to show.

  1. self definitional [Section 3.4, Rationality Analysis (after Eq. (8))]
    "Since LLMs embed extensive world knowledge that potentially covers all relevant hidden confounders U ∗, they can be viewed as approximately unbiased counterfactual estimators under the unconfoundedness assumption. That is, when conditioned on the full context accessible to the LLM—including both structured and unstructured information—the potential outcomes (Y 0, Y1) become independent of the treatment T ."

    This posits that the LLM already implicitly conditions on a sufficient latent set U*, which is exactly the property ProCI is supposed to discover and validate through progressive confounder imputation and the KCIT stopping rule. The empirical unconfoundedness test is justified by assuming that the LLM's counterfactual imputations are unbiased because the LLM already knows a sufficient set of confounders. Whether such a sufficient adjustment set exists, and which observed or generated variables form it, is the very question the framework claims to answer. The conclusion needed to make the test valid is therefore asserted as a premise rather than derived.

  2. fitted input called prediction [Section 3.4, Eqs. (7)-(9)]
    "ˆy1−ti i =LLM(P out(xi, ui, ti, yi)). (7) ... After obtaining the imputed outcomes ˆY 0 and ˆY 1, we can statistically test the unconfoundedness assumption using a non-parametric independence test. Specifically, we adopt the Kernel-based Conditional Independence Test (KCIT) ... KCIT evaluates whether the treatment T is conditionally independent of the imputed outcomes( ˆY 0, ˆY 1)given the observed and generated confoundersXand ˆU: I[KCIT(( ˆY 0, ˆY 1), T|X, ˆU)>α]=1. (9)"

    The test target is not the true potential outcomes (Y0,Y1) but the LLM's imputed values, produced by prompting the same LLM with the observed xi, ti, yi and the LLM-generated ui. The same observed treatment and outcome data are used both to impute the candidate confounder U and to impute the counterfactual outcomes that are subsequently tested for conditional independence. A passing KCIT can therefore certify only internal consistency of the LLM's imputation procedure, not actual unconfoundedness of X∪U. The 'empirical validation' is a fitness check of the LLM's own outputs, i.e., a fitted input renamed as a prediction.

1 more flagged steps
  1. other [Appendix B, Lemma 2 and its proof (Eqs. (14)-(16))]
    "Lemma 2. Let ˆYi, ˆZi be samples from LC classes LC(µY ,Σ Y i, ω)and LC(µZ, Σ Zi, ω)respectively, such that ∥E ˜K ˆY − ˜KY ∥=o p(1) and ∥E ˜K ˆZ − ˜KZ∥=o p(1). Then, for Gaussian RBF such that σ→ ∞asn→ ∞,we have ∥ ˆΣ ˆYT − ˆΣYT ∥=o p(1) ..."

    Theorem 1's proof establishes the test-statistic convergence only under the lemma's assumptions that the expected kernel matrices of the imputed variables converge to the true kernel matrices. Those assumptions are the very imputation-fidelity property the framework needs to justify, and no argument links the LLM prompting scheme of Eqs. (4)-(7) to them. Moreover, the proof's scaling is erroneous: from Lemma 1, ∥K_Yhat−EK_Yhat∥=o_p(n), so the trace bound gives (1/n)·O(n)·o_p(n)=o_p(n), not o_p(1). Thus the asymptotic validity of the unconfoundedness test is not established, and the claimed sufficiency result reduces to an unverified kernel-consistency assumption equivalent to the desired conclusion.

full rationale

ProCI's final CATE estimates are evaluated against external ground-truth outcomes on the Twins and Jobs benchmarks, and the ablation studies provide independent evidence of practical improvements. However, the paper's central theoretical claim—that ProCI 'adaptively constructs a sufficient adjustment set'—rests on the empirical unconfoundedness test of Section 3.4. That test is partially circular. First, the counterfactual outcomes used as test targets are themselves LLM imputations (Eq. (7)) generated from the same observed X, T, Y and the same LLM that produced the candidate confounders, so a pass can indicate internal consistency of the imputation procedure rather than true unconfoundedness. Second, the Rationality Analysis justifies those imputations by positing that the LLM already conditions on a sufficient latent U*, which is precisely the property ProCI is meant to discover. The formal Theorem 1 does not resolve this: Lemma 2 assumes kernel-consistency of the imputations (||E K_Yhat - K_Y|| = o_p(1)) rather than deriving it from the prompting scheme, and the proof's trace-bound reasoning yields o_p(n), not o_p(1), leaving the theorem unproven. Because this test is the stopping rule that licenses the 'sufficient adjustment set' conclusion, the sufficiency claim is partially circular or at least rests on an assumption equivalent to the conclusion. The empirical gains over base estimators remain independent evidence of practical value, so the overall score is moderate rather than maximal.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claim depends on the existence of a sufficient hidden set U* inside the LLM, on kernel-consistency of LLM imputations, and on a set of LLM outputs (distribution families, parameters) that are fitted using the observed treatment and outcome. These are substantial assumptions with no independent support.

free parameters (4)
  • KCIT significance level alpha = not reported
    Eq. (9) thresholds the test p-value; the paper does not state the chosen alpha, making the stopping rule in Eq. (11) underspecified.
  • LLM generation temperature = 0.7
    Section 4.1 fixes temperature at 0.7 for all LLM calls; temperature controls the diversity of generated confounders and imputation noise.
  • Number of imputed confounders = stopped by KCIT p-value > alpha
    The stopping rule in Eq. (11) determines the count per dataset and seed; the actual count is not reported.
  • Per-unit distribution parameters theta_i = elicited from LLM per individual
    Eq. (6) samples confounder values from LLM-inferred distributions parameterized per unit; these parameters are fitted using observed T, Y, X, so they are not independent of the outcome.
assumptions (4)
  • domain assumption The LLM implicitly conditions on a sufficient latent variable U* with (Y0,Y1) independent of T given X and U*.
    Section 3.4 Rationality Analysis states this to justify using LLM-imputed counterfactuals in the unconfoundedness test.
  • domain assumption LLM imputed potential outcomes and imputed confounders are kernel-consistent: ||E K_Yhat - K_Y|| = o_p(1) and ||E K_Zhat - K_Z|| = o_p(1).
    Appendix B Lemma 2 assumes these convergences; Theorem 1 depends on them. No evidence, theoretical or empirical, is given that LLM imputations satisfy them.
  • domain assumption The candidate distribution families F include the true distribution of the missing confounder.
    Section 3.3 asks the LLM to pick a distribution type from a limited set; a wrong choice misspecifies the imputed confounder.
  • standard math KCIT satisfies 'standard regularity conditions' and the kernel is chosen so the test is valid on the augmented data.
    Theorem 1 and Appendix B invoke regularity conditions without specification.
invented entities (2)
  • Latent variable U*
    purpose: Conceptual sufficient set of hidden confounders supposedly encoded in LLM world knowledge; used to argue LLM counterfactual imputations are unbiased.
    Introduced in Section 3.4; not observed, not measured, and no falsifiable handle is provided.
  • LLM-generated confounders Uhat
    purpose: Imputed variables appended to X to restore unconfoundedness before CATE estimation.
    Their values are sampled from distributions inferred by the LLM using T, Y, X; there is no external data validating that they correspond to real confounding factors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Hidden Confounding by Progressive Confounder Imputation via Large Language Models." pith.science (2026). https://pith.science/paper/RLAXLTCE

@misc{pith2026250702928,
  author       = {Pith},
  title        = {Pith review of: Mitigating Hidden Confounding by Progressive Confounder Imputation via Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RLAXLTCE}},
  note         = {Machine review of arXiv:2507.02928}
}
read the original abstract

Hidden confounding remains a central challenge in estimating treatment effects from observational data, as unobserved variables can lead to biased causal estimates. While recent work has explored the use of large language models (LLMs) for causal inference, most approaches still rely on the unconfoundedness assumption. In this paper, we make the first attempt to mitigate hidden confounding using LLMs. We propose ProCI (Progressive Confounder Imputation), a framework that elicits the semantic and world knowledge of LLMs to iteratively generate, impute, and validate hidden confounders. ProCI leverages two key capabilities of LLMs: their strong semantic reasoning ability, which enables the discovery of plausible confounders from both structured and unstructured inputs, and their embedded world knowledge, which supports counterfactual reasoning under latent confounding. To improve robustness, ProCI adopts a distributional reasoning strategy instead of direct value imputation to prevent the collapsed outputs. Extensive experiments demonstrate that ProCI uncovers meaningful confounders and significantly improves treatment effect estimation across various datasets and LLMs.

Figures

Figures reproduced from arXiv: 2507.02928 by the authors.

Figure 1
Figure 1. Illustration of how LLMs help mitigate hidden confounding. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of the ProCI framework. ProCI alternates between two phases: (1) Confounder [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. (Left) CMI measures dependence between LLM-generated confounders and treatment/out￾come, compared to random confounders. (Right) Confounders generated in Jobs demonstrate their semantic relevance and impact on treatment/outcome. Original RM-1 RM-2 RM-3 RM-4 RM-5 RM-6 # Removal Confounders 0.02 0.04 0.06 0.08 0.10 Performance Comparison over In-sample ATT CFR-Wass ESCFR ProCI-4o ProCI-R1 Original RM-1 RM-2 RM-3 RM-4 … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Robustness of ProCI in CATE estimation with varying hidden confounder removal. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Conditional mutual information (CMI) values across different temperatures for confounders generated by ProCI-La (LLaMA 3–8B), ProCI-Qw (Qwen2.5–7B), and a Random baseline. The left plot shows I(U, Y | X), and the right plot shows I(U, T | X), with all values on a log s…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 40 canonical work pages

  1. [33]

    Discovery of the hidden world with large language models, 2024

    Chenxi Liu, Yongqiang Chen, Tongliang Liu, Mingming Gong, James Cheng, Bo Han, and Kun Zhang. Discovery of the hidden world with large language models, 2024

  2. [1]

    Extracting Self-Consistent Causal Insights from Users Feedback with LLMs and In-context Learning

    Sara Abdali, Anjali Parikh, Steve Lim, and Emre Kiciman. Extracting self-consistent causal insights from users feedback with llms and in-context learning.arXiv preprint arXiv:2312.06820, 2023

  3. [2]

    Estimating conditional average treatment effects.Journal of Business & Economic Statistics, 33(4):485–505, 2015

    Jason Abrevaya, Yu-Chin Hsu, and Robert P Lieli. Estimating conditional average treatment effects.Journal of Business & Economic Statistics, 33(4):485–505, 2015

  4. [3]

    The costs of low birth weight.The Quarterly Journal of Economics, 120(3):1031–1083, 2005

    Douglas Almond, Kenneth Y Chay, and David S Lee. The costs of low birth weight.The Quarterly Journal of Economics, 120(3):1031–1083, 2005

  5. [4]

    Concentration of kernel matrices with application to kernel spectral clustering.The Annals of Statistics, 49(1):531–556, 2021

    Arash A Amini and Zahra S Razaee. Concentration of kernel matrices with application to kernel spectral clustering.The Annals of Statistics, 49(1):531–556, 2021

  6. [5]

    Cv Ananth and E. F. Schisterman. Hidden biases in observational epidemiology: the case of unmeasured confounding.BJOG: An International Journal of Obstetrics & Gynaecology, 125:644 – 646, 2018

  7. [6]

    Counterfactual representation learning with balancing weights

    Serge Assaad, Shuxi Zeng, Chenyang Tao, Shounak Datta, Nikhil Mehta, Ricardo Henao, Fan Li, and Lawrence Carin. Counterfactual representation learning with balancing weights. In International Conference on Artificial Intelligence and Statistics, pages 1972–1980. PMLR, 2021

  8. [7]

    Bellemare, Jeffrey R

    Marc F. Bellemare, Jeffrey R. Bloem, and Noah Wexler. The paper of how: Estimating treatment effects using the front-door criterion*.Oxford Bulletin of Economics and Statistics, 2024

Show all 59 references
  1. [8]

    Some practical guidance for the implementation of propensity score matching.Journal of economic surveys, 22(1):31–72, 2008

    Marco Caliendo and Sabine Kopeinig. Some practical guidance for the implementation of propensity score matching.Journal of economic surveys, 22(1):31–72, 2008

  2. [9]

    Quantifying and mitigating uni- modal biases in multimodal large language models: A causal perspective.arXiv preprint arXiv:2403.18346, 2024

    Meiqi Chen, Yixin Cao, Yan Zhang, and Chaochao Lu. Quantifying and mitigating uni- modal biases in multimodal large language models: A causal perspective.arXiv preprint arXiv:2403.18346, 2024

  3. [10]

    Causal evaluation of language models.arXiv preprint arXiv:2405.00622, 2024

    Sirui Chen, Bo Peng, Meiqi Chen, Ruiqi Wang, Mengying Xu, Xingyu Zeng, Rui Zhao, Shengjie Zhao, Yu Qiao, and Chaochao Lu. Causal evaluation of language models.arXiv preprint arXiv:2405.00622, 2024

  4. [11]

    End-to-end causal effect estimation from unstructured natural language data.arXiv preprint arXiv:2407.07018, 2024

    Nikita Dhawan, Leonardo Cotta, Karen Ullrich, Rahul G Krishnan, and Chris J Maddison. End-to-end causal effect estimation from unstructured natural language data.arXiv preprint arXiv:2407.07018, 2024

  5. [12]

    Franks, Alexander D’Amour, and Avi Feller

    Alexander M. Franks, Alexander D’Amour, and Avi Feller. Flexible sensitivity analysis for observational studies without observable implications.Journal of the American Statistical Association, 115:1730 – 1746, 2018

  6. [13]

    Fulcher, Ilya Shpitser, Stella Marealle, and Eric J

    Isabel R. Fulcher, Ilya Shpitser, Stella Marealle, and Eric J. Tchetgen Tchetgen. Robust inference on population indirect causal effects: the generalized front door criterion.Journal of the Royal Statistical Society: Series B (Statistical Methodology), 82, 2017

  7. [14]

    Causal inference in sociological research.Annual review of sociology, 36(1):21– 47, 2010

    Markus Gangl. Causal inference in sociological research.Annual review of sociology, 36(1):21– 47, 2010. 10

  8. [15]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

  9. [16]

    A review of instrumental variables estimation of treatment effects in the applied health sciences.Health Services and Outcomes Research Methodology, 7:159–179, 2007

    Paul Grootendorst. A review of instrumental variables estimation of treatment effects in the applied health sciences.Health Services and Outcomes Research Methodology, 7:159–179, 2007

  10. [17]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

  11. [18]

    Causal agent based on large language model.arXiv preprint arXiv:2408.06849, 2024

    Kairong Han, Kun Kuang, Ziyu Zhao, Junjian Ye, and Fei Wu. Causal agent based on large language model.arXiv preprint arXiv:2408.06849, 2024

  12. [19]

    Combining observational and randomized data for estimating heterogeneous treatment effects

    Tobias Hatt, Jeroen Berrevoets, Alicia Curth, Stefan Feuerriegel, and Mihaela van der Schaar. Combining observational and randomized data for estimating heterogeneous treatment effects. ArXiv, abs/2202.12891, 2022

  13. [20]

    Bayesian nonparametric modeling for causal inference.Journal of Computa- tional and Graphical Statistics, 20(1):217–240, 2011

    Jennifer L Hill. Bayesian nonparametric modeling for causal inference.Journal of Computa- tional and Graphical Statistics, 20(1):217–240, 2011

  14. [21]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

  15. [22]

    Reasoning elicitation in language models via counterfactual feedback.arXiv preprint arXiv:2410.03767, 2024

    Alihan Hüyük, Xinnuo Xu, Jacqueline Maasch, Aditya V Nori, and Javier González. Reasoning elicitation in language models via counterfactual feedback.arXiv preprint arXiv:2410.03767, 2024

  16. [23]

    Causal representation learning with generative artificial intelligence: Application to texts as treatments, 2024

    Kosuke Imai and Kentaro Nakamura. Causal representation learning with generative artificial intelligence: Application to texts as treatments, 2024

  17. [24]

    Instrumental variables: An econometrician’s perspective.Political Methods: Quantitative Methods eJournal, 2014

    Guido Imbens. Instrumental variables: An econometrician’s perspective.Political Methods: Quantitative Methods eJournal, 2014

  18. [25]

    Quantifying ignorance in individual-level causal-effect estimates under hidden confounding

    Andrew Jesson, Sören Mindermann, Yarin Gal, and Uri Shalit. Quantifying ignorance in individual-level causal-effect estimates under hidden confounding. InInternational Conference on Machine Learning, pages 4829–4838. PMLR, 2021

  19. [26]

    Cladder: Assessing causal reasoning in language models.Advances in Neural Information Processing Systems, 36:31038–31065, 2023

    Zhijing Jin, Yuen Chen, Felix Leeb, Luigi Gresele, Ojasv Kamal, Zhiheng Lyu, Kevin Blin, Fernando Gonzalez Adauto, Max Kleiman-Weiner, Mrinmaya Sachan, et al. Cladder: Assessing causal reasoning in language models.Advances in Neural Information Processing Systems, 36:31038–31065, 2023

  20. [27]

    Removing hidden confounding by experimental grounding.ArXiv, abs/1810.11646, 2018

    Nathan Kallus, Aahlad Puli, and Uri Shalit. Removing hidden confounding by experimental grounding.ArXiv, abs/1810.11646, 2018

  21. [28]

    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

  22. [29]

    Evaluating the econometric evaluations of training programs with experi- mental data.The American economic review, pages 604–620, 1986

    Robert J LaLonde. Evaluating the econometric evaluations of training programs with experi- mental data.The American economic review, pages 604–620, 1986

  23. [30]

    Removing hidden confounding in recommendation: A unified multi-task learning approach

    Haoxuan Li, Kunhan Wu, Chunyuan Zheng, Yanghao Xiao, Hao Wang, Zhi Geng, Fuli Feng, Xiangnan He, and Peng Wu. Removing hidden confounding in recommendation: A unified multi-task learning approach. InNeural Information Processing Systems, 2023

  24. [31]

    Optimizing language models for human preferences is a causal inference problem.arXiv preprint arXiv:2402.14979, 2024

    Victoria Lin, Eli Ben-Michael, and Louis-Philippe Morency. Optimizing language models for human preferences is a causal inference problem.arXiv preprint arXiv:2402.14979, 2024

  25. [32]

    Text-transport: Toward learning causal effects of natural language.arXiv preprint arXiv:2310.20697, 2023

    Victoria Lin, Louis-Philippe Morency, and Eli Ben-Michael. Text-transport: Toward learning causal effects of natural language.arXiv preprint arXiv:2310.20697, 2023. 11

  26. [34]

    Causal effect inference with deep latent-variable models.Advances in neural information processing systems, 30, 2017

    Christos Louizos, Uri Shalit, Joris M Mooij, David Sontag, Richard Zemel, and Max Welling. Causal effect inference with deep latent-variable models.Advances in neural information processing systems, 30, 2017

  27. [35]

    Causal inference with large language model: A survey.arXiv preprint arXiv:2409.09822, 2024

    Jing Ma. Causal inference with large language model: A survey.arXiv preprint arXiv:2409.09822, 2024

  28. [36]

    Answering causal questions with augmented llms

    Nick Pawlowski, James Vaughan, Joel Jennings, and Cheng Zhang. Answering causal questions with augmented llms. 2023

  29. [37]

    Causal inference in statistics: An overview.Statistics Surveys, 3:96–146, 2009

    Judea Pearl. Causal inference in statistics: An overview.Statistics Surveys, 3:96–146, 2009

  30. [38]

    Basic books, 2018

    Judea Pearl and Dana Mackenzie.The book of why: the new science of cause and effect. Basic books, 2018

  31. [39]

    Sutherland, and Arthur Gretton

    Roman Pogodin, Antonin Schrab, Yazhe Li, Danica J. Sutherland, and Arthur Gretton. Practical kernel tests of conditional independence, 2024

  32. [40]

    Causal inference and counterfactual prediction in machine learning for actionable healthcare.Nature Machine Intelligence, 2(7):369–375, 2020

    Mattia Prosperi, Yi Guo, Matt Sperrin, James S Koopman, Jae S Min, Xing He, Shannan Rich, Mo Wang, Iain E Buchan, and Jiang Bian. Causal inference and counterfactual prediction in machine learning for actionable healthcare.Nature Machine Intelligence, 2(7):369–375, 2020

  33. [41]

    Qwen2.5 technical report, 2025

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, ...

  34. [42]

    Robins, Avshalom Caspi, and Terrie E

    Richard W. Robins, Avshalom Caspi, and Terrie E. Moffitt. Two personalities, one relationship: both partners’ personality traits shape the quality of their relationship.Journal of personality and social psychology, 79 2:251–9, 2000

  35. [43]

    Rosenbaum and Donald B

    Paul R. Rosenbaum and Donald B. Rubin. Assessing sensitivity to an unobserved binary covariate in an observational study with binary outcome.Journal of the royal statistical society series b-methodological, 45:212–218, 1983

  36. [44]

    Causal inference using potential outcomes: Design, modeling, decisions

    Donald B Rubin. Causal inference using potential outcomes: Design, modeling, decisions. Journal of the American statistical Association, 100(469):322–331, 2005

  37. [45]

    Front-door adjustment beyond markov equivalence with limited graph knowledge.ArXiv, abs/2306.11008, 2023

    Abhin Shah, Karthikeyan Shanmugam, and Murat Kocaoglu. Front-door adjustment beyond markov equivalence with limited graph knowledge.ArXiv, abs/2306.11008, 2023

  38. [46]

    Estimating individual treatment effect: generalization bounds and algorithms

    Uri Shalit, Fredrik D Johansson, and David Sontag. Estimating individual treatment effect: generalization bounds and algorithms. InInternational conference on machine learning, pages 3076–3085. PMLR, 2017

  39. [47]

    End-to-end cost-effective incentive recommendation under budget constraint with uplift modeling

    Zexu Sun, Hao Yang, Dugang Liu, Yunpeng Weng, Xing Tang, and Xiuqiang He. End-to-end cost-effective incentive recommendation under budget constraint with uplift modeling. In Proceedings of the 18th ACM Conference on Recommender Systems, pages 560–569, 2024

  40. [48]

    Causal abstraction for chain-of-thought reasoning in arithmetic word problems

    Juanhe TJ Tan. Causal abstraction for chain-of-thought reasoning in arithmetic word problems. InProceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pages 155–168, 2023

  41. [49]

    Causal inference in economics and marketing.Proceedings of the National Academy of Sciences, 113(27):7310–7315, 2016

    Hal R Varian. Causal inference in economics and marketing.Proceedings of the National Academy of Sciences, 113(27):7310–7315, 2016

  42. [50]

    Sense and sensitivity analysis: Simple post-hoc analysis of bias due to unobserved confounding.ArXiv, abs/2003.01747, 2020

    Victor Veitch and Anisha Zaveri. Sense and sensitivity analysis: Simple post-hoc analysis of bias due to unobserved confounding.ArXiv, abs/2003.01747, 2020. 12

  43. [51]

    Optimal transport for treatment effect estimation

    Hao Wang, Jiajun Fan, Zhichao Chen, Haoxuan Li, Weiming Liu, Tianqiao Liu, Quanyu Dai, Yichao Wang, Zhenhua Dong, and Ruiming Tang. Optimal transport for treatment effect estimation. InThirty-seventh Conference on Neural Information Processing Systems, 2023

  44. [52]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

  45. [53]

    Li, and Fei Wu

    Anpeng Wu, Kun Kuang, B. Li, and Fei Wu. Instrumental variable regression with confounder balancing. InInternational Conference on Machine Learning, 2022

  46. [54]

    Stable estimation of heteroge- neous treatment effects

    Anpeng Wu, Kun Kuang, Ruoxuan Xiong, Bo Li, and Fei Wu. Stable estimation of heteroge- neous treatment effects. InInternational Conference on Machine Learning, pages 37496–37510. PMLR, 2023

  47. [55]

    Integrative r-learner of heterogeneous treatment effects combining experimental and observational studies

    Lili Wu and Shu Yang. Integrative r-learner of heterogeneous treatment effects combining experimental and observational studies. InCLEaR, 2022

  48. [56]

    Representation learning for treatment effect estimation from observational data.Advances in neural information processing systems, 31, 2018

    Liuyi Yao, Sheng Li, Yaliang Li, Mengdi Huai, Jing Gao, and Aidong Zhang. Representation learning for treatment effect estimation from observational data.Advances in neural information processing systems, 31, 2018

  49. [57]

    Causal-cog: A causal-effect look at context generation for boosting multi-modal language models

    Shitian Zhao, Zhuowan Li, Yadong Lu, Alan Yuille, and Yan Wang. Causal-cog: A causal-effect look at context generation for boosting multi-modal language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13342–13351, 2024. 13 A I...

  50. [58]

    The unobserved counterfactual remains inaccessible

    Missing counterfactuals: For each individual, we only observe the outcome corresponding to the assigned treatment. The unobserved counterfactual remains inaccessible

  51. [59]

    To address these challenges, [38] proposed a two-stage framework

    Selection bias: Treatment assignment may depend on covariates related to the outcome, leading to systematic differences between treated and control groups. To address these challenges, [38] proposed a two-stage framework. The first stage,identification, aims to express causal ...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.