Pith. sign in

REVIEW 4 major objections 6 minor 76 references

Large Language Models for Statistical Inference: Context Augmentation with Applications to the Two-Sample Problem and Regression

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that valid frequentist inference is possible directly on raw text by treating LLM-generated contexts as latent variables, giving calibrated two-sample t-statistics and asymptotically normal text-on-text regression.

desk verdict A genuinely new LLM-for-inference framework with a strong empirical core, but the headline rate result in Result 3 is not proven and the key identification assumption is untestable. read the letter →

arxiv 2506.23862 v1 pith:KVR446R5 submitted 2025-06-30 stat.ME

Marc Ratkovic This is my paper
classification stat.ME MSC 62F0362F1262G2062R07
keywords contextaugmentationlargelanguagemodelstwo-sampletesttextregressionfrequentistinferencecross-fittingnegativecontrolssemanticversussyntactic
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

Context augmentation treats the textual environment around each observed string as a latent variable: a large language model generates contexts under different conditioning events, and the likelihood of the string inside each context becomes the raw material for estimation. The paper claims that this yields valid frequentist inference directly on text, with two-sample t-statistics that are calibrated under the null and regression coefficients that are asymptotically normal. It also provides a way to attribute predictive power to semantic versus syntactic dimensions using negative controls built by masking, shuffling, or replacing content words with nonsense. The practical stakes are that text data can enter hypothesis testing and regression through standard tools, rather than through fixed embeddings or topic models that discard contextual uncertainty.

What carries the argument

The argument is carried by the clause function $\mathrm{Cl}(s,c,E,\hat M)$, which returns a score for how likely an observed string $s$ is within a generated context $c$ under conditioning event $E$, using scoring model $\hat M$. Contexts are generated separately under each event, aggregated within a string by operator $A$, and mapped to the target parameter by operator $T$. Identification rests on three assumptions: Overlap, Weak Ignorability (the clause score depends on the event only through the generated context), and Injectivity. Estimation uses repeated cross-fitting, where strings are evaluated against out-of-fold contexts and the resulting statistics are averaged over many splits; pivotality of the test statistic then lets a second-order Edgeworth expansion cancel the leading bias term.

What would settle it

A direct test of Weak Ignorability is to hold the generated context fixed while changing only the event label, then compare the clause scores $\mathrm{Cl}(s,c_E,E,\hat M)$ and $\mathrm{Cl}(s,c_E,E',\hat M)$ across many strings; any systematic difference falsifies the identification lemma. A second check is to generate contexts from identical groups while deliberately injecting the group label into the prompt, and observe whether the cross-fit t-statistic drifts away from its null distribution.

Watch

Extended reading notes

Core claim

The central claim is that LLM-generated contexts can serve as latent variables that make ordinary frequentist tools applicable to raw text. In the two-sample problem, each string is scored by its log-probability under contexts generated from its own group versus the other group, and the cross-fit averaged t-statistic is shown to be null-calibrated and powerful on synthetic data. In text-on-text regression, contexts generated around the predictor string are treated as mediators, and the outcome's log-probability is regressed on the outcome's log-probability under masked, shuffled, and jabberwocky variants of the predictor; negative controls then separate semantic from syntactic and lexical contributions. The theoretical contribution is an influence-function decomposition that isolates sampling variation, context-level error, and self-referential bias, together with a pivotality result showing that repeated cross-fitting reduces the required context-generation rate from order $n^{1/2}$ to $n^{1/4}$, and bounds linking estimation error to context count and number of cross-fits.

Load-bearing premise

The load-bearing premise is Weak Ignorability: the language model's clause score for a string must depend on the conditioning event only through the generated context, never through the event label itself, and the paper does not verify this property for any real model.

Editorial extensions

If this is right

  • Two-sample tests on raw text can be run with calibrated p-values, without fixed embeddings or topic-model summaries.
  • Text-on-text regression becomes feasible with coefficients and standard errors, allowing researchers to attribute prediction to semantic, syntactic, or lexical content.
  • Context-generation cost can be traded against cheap repeated cross-fits: the required number of contexts per string shrinks while cross-fit repetitions grow.
  • The influence-function representation opens the way to text-based treatment effects and causal estimands using standard semiparametric machinery.

Reading between the lines

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

  • If Weak Ignorability holds only approximately, cross-fitting may remove sampling and context noise but leave a bias proportional to the LLM's reliance on event labels; a sensitivity analysis that varies prompt emphasis would bound this bias.
  • The jabberwocky, shuffled, and masked negative controls could serve as a general diagnostic for whether any text-to-text model's predictions are semantics-driven or syntax-driven, not just in dialogue.
  • Because the framework conditions on the scoring model, different LLMs could be compared on the same estimand, turning model choice into a sensitivity check that the paper does not develop.
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 / 6 minor

Summary. The paper introduces "context augmentation," a method for frequentist inference on text data that uses LLM-generated contexts as latent variables. Observed strings are scored against contexts generated by an LLM, and the paper develops a two-sample t-test and a text-on-text regression framework. It claims asymptotic normality of the resulting estimators, an influence-function decomposition that isolates sampling, context, and self-referential bias, and a higher-order efficiency result: repeated cross-fitting of a pivotal statistic reduces the required number of contexts per string from order n^{1/2} to n^{1/4}. Empirical sections demonstrate null calibration and power in synthetic two-sample data, a regression with negative controls distinguishing semantic from syntactic contributions, a replication of Egami et al. in which context augmentation yields a much larger test statistic than a topic-model baseline, and a dialogue analysis supporting the integrative-repair hypothesis. The theoretical core is Result 3 in Section 5.4, whose proof is the subject of the main concerns below.

Significance. If the theoretical claims held, the paper would offer a practically relevant way to use LLMs as inference engines while retaining standard frequentist guarantees, and the computational guidance on trading context generation against cross-fitting repetitions would be valuable. The empirical pipeline is clearly described, with pseudocode and prompt templates, and the authors are transparent about computational costs and tuning choices. The replication and dialogue applications are interesting and suggest that context-based scores can detect differences that topic models miss. However, the central theoretical contribution—the n^{1/2}-to-n^{1/4} rate reduction under pivotality—is not established. The proof in Appendix A.4 contains unsupported identifications and an invalid application of a classical empirical-process bound. The identification assumption underlying the general theory is not satisfied in the regression setting as presented. These issues are load-bearing because the paper's primary novelty is the claimed theoretical guarantee, and the empirical results, while suggestive, cannot compensate for a missing or incorrect derivation.

major comments (4)
  1. [Section 5.4, Result 3 and Appendix A.4] The rate-reduction claim is not proven. First, the repeated cross-fit estimator is asserted to be "a symmetric U-statistic of order two in the sample indices" without a derivation; the estimator defined in Appendix B averages over R random splits, and such an average over random partitions is not an average over all pairs (i, j), so the U-statistic representation is not immediate. Second, the statement that "pivotality removes the leading r_n^{-1} term in both the pivot and the estimator" is an assertion; no Edgeworth expansion of the cross-fitted estimator under the nonparametric nuisance is written down, and pivotality of the limit distribution does not by itself control a nonparametric bias term of order ||hat F - F||_infty appearing in Result 1, Eq. (5). Third, the factorization of the self-referential bias into ||hatF - F||_infty times (1/M) sum R^(m) is introduced without a derivation and the R^(m) are not defined from the estimation problem. Consequently, the conclusion that the nuisance rate relaxes from n^{-1/4} to n^{-1/8}, and hence that the context count requirement drops from order n^{1/2} to n^{1/4}, is not established.
  2. [Appendix A.4] The application of the Dvoretzky-Kiefer-Wolfowitz bound to ||hat F - F||_infty is invalid in this setting. DKW is a statement about the empirical CDF of real-valued random variables and requires a total order on the sample space. Here F is the distribution of LLM-generated contexts, which are text strings; no order or metric on the context space is defined in the paper. Without such an order, the sup-norm ||hat F - F||_infty is not well defined, and even if an arbitrary order were imposed, the DKW rate O_p(sqrt(log n_c / n_c)) would not follow for a distribution over a high-dimensional discrete space. The bound for the context estimation error is therefore unsupported, and the subsequent bias bound in Result 3 lacks a valid foundation.
  3. [Section 5.1 vs. Section 7] The Weak Ignorability assumption, Cl(s, c_E, E, M) = Cl(s, c_E, E', M), is not satisfied by the regression clause function introduced in Section 7, which is defined as Cl(s_y, s_x^nu, c_j;i, M) = log Pr(s_y | s_x^nu, c_j;i, M). Here the event E is operationalized through the predictor variant s_x^nu, which appears explicitly in the clause-function input. Thus changing the event changes the clause function, violating the stated assumption. The paper presents the identification lemma for the general framework and does not flag this limitation, so either the identification theory does not cover the regression application or the assumption needs to be reformulated to match the actual clause function used.
  4. [Section 5.3] The claim after Result 2 that cross-fitting reduces the nuisance rate requirement from o_p(n^{-1/2}) to o_p(n^{-1/4}) is not supported by the displayed decomposition. In Eq. (5), the self-referential bias term is linear in the sup-norm ||hatF - F||_infty; standard cross-fitting arguments achieve bias reduction for product terms of two n^{-1/4}-rate nuisances, but no such product structure is shown in Result 1. The factorization in Appendix A.4 that would yield a product is not derived from Eq. (5) and conflicts with the linear form. This n^{-1/4} rate is a prerequisite for the pivotality-based relaxation in Result 3, so it needs to be proven or explicitly assumed rather than stated.
minor comments (6)
  1. [Section 1] The phrase "reduce the order of generated contexts from order n^{1/2} to n^{1/4}" is ambiguous: the number of contexts per string n_c is a user-chosen integer, not of order n^{1/2} in any formula in the paper. The text appears to refer to the bias rate or the required relationship (n_c M)^{1+epsilon} >> n^{2 gamma}; please restate this in terms of the derived inequality in Result 3.
  2. [Section 6.1 and Figure 2] Null calibration is reported only as a QQ plot of p-values. Please also report the empirical type I error rate at conventional levels (e.g., 0.05) and the distribution of the cross-fit t-statistic under the null, since the headline claim is that the t-statistics "exhibit the expected null behaviour."
  3. [Section 8] The replication reports a z-score of 37.24 for context augmentation, which is extraordinarily large. Given that the theoretical null distribution is under scrutiny, please provide a permutation-based or split-sample null check in this application to verify that the reported t-statistic is not over-calibrated.
  4. [Table 5] The entries n_c = 0.5 and n_c = 0.25 are described as "one context sampled from every other and fourth string," but n_c is defined elsewhere as the number of contexts per string. Using non-integer values of n_c is confusing; please define the sampling procedure precisely.
  5. [Appendix B] The definition of the t-statistic in each direction (t_{I1->I2} and t_{I2->I1}) is not given. Please write out the formula for the standard error used in the average and the rationale for the sqrt(2) rescaling, as these affect the validity of the reported p-values.
  6. [Throughout] There are several typographical and formatting issues, such as "the method'st-statistics" in the abstract, missing section reference "Appendix contains pseudocode" in Section 2, and inconsistent notation for the cross-fit estimator across Section 5.4 and Appendix A.4. A careful copyedit is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: theoretical claims are conditional on stated assumptions and empirical claims are checked against synthetic data and external replications.

full rationale

The paper does not exhibit a circular derivation chain. The central estimators (t-statistics and regression coefficients) are computed from LLM-generated contexts and then evaluated for null calibration and power on synthetic data; they are not fitted to reproduce those evaluation targets. The replication against Egami et al. (2022) and the DeliData analysis provide external, non-circular benchmarks. The theoretical results are conditional on explicitly stated assumptions, most notably Weak Ignorability and Assumption 3 (Pivotal Limit). These assumptions are strong and their validity for real LLMs is not established, and the paper candidly concedes in Section 10 that 'it is not clear whether the tail-conditions or approximation error rates needed for valid inference obtain.' That is a correctness and external-validity concern, not a circularity: the paper does not define its estimand in terms of the estimator, does not relabel a fitted parameter as a prediction, and does not rely on a self-citation chain to force its conclusions. The proof of Result 3 in Appendix A.4 is terse and arguably under-derived, but the rate claim is a stated consequence of an assumed pivotality property rather than an identity smuggled in through definitions. No circular step satisfying the required evidentiary standard can be exhibited.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a small set of free parameters (context count, cross-fit count, generation hyperparameters) and on several domain assumptions about the behavior of the LLM, especially Weak Ignorability and Overlap. These assumptions are untestable in the paper and are not supported by external evidence. The paper introduces no new physical entities, but the latent context construct is essential and lacks independent falsifiability.

free parameters (3)
  • number of contexts per string (n_c) = 10 in synthetic two-sample; 5 to 0.1 contexts per string in replication
    User-chosen; the theoretical bounds relate n_c to n, but no data-driven selection is provided and results vary with n_c (Table 5).
  • number of repeated cross-fits (R) = 25
    Chosen by hand to stabilize the two-sample t-statistic; the paper does not analyze sensitivity to R.
  • generation hyperparameters (temperature, top_k, do_sample) = temperature=0.8, top_k=50, do_sample=True
    Set manually for all experiments; the paper acknowledges tuning complexity but does not ablate these choices.
assumptions (5)
  • domain assumption Overlap: supp(F_c|s,E) = supp(F_c|s,E') for events E,E'
    Section 5.1, Assumption 1. Requires the LLM to generate the same support of contexts under different events; untested.
  • domain assumption Weak Ignorability: Cl(s,c_E,E,M) = Cl(s,c_E,E',M)
    Section 5.1, Assumption 2. The clause function must be event-ignorable given the context; this is the key identifying assumption and is not validated.
  • domain assumption Injectivity: the mapping T∘A is injective
    Section 5.1, Assumption 3. Needed to turn a difference in θ into a difference in string distributions.
  • standard math Standard semiparametric regularity: Hadamard differentiability, Lyapunov moments, stochastic equicontinuity, uniform nuisance error bound
    Section 5.3, Assumptions 1 and 2. These are standard but not checked for the LLM-generated contexts.
  • standard math Pivotal limit and Cramér condition: the statistic has a known limiting distribution with 4+ε moments
    Section 5.4, Assumption 3. Needed for the higher-order Edgeworth argument.
invented entities (1)
  • LLM-generated latent contexts (c_E)
    purpose: Serve as auxiliary latent variables connecting observed strings to events; the generative model for contexts is the key object of inference.
    Contexts are generated by the LLM and are not directly observable or falsifiable outside the model. The method's validity depends on these contexts satisfying the identification assumptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Models for Statistical Inference: Context Augmentation with Applications to the Two-Sample Problem and Regression." pith.science (2026). https://pith.science/paper/KVR446R5

@misc{pith2026250623862,
  author       = {Pith},
  title        = {Pith review of: Large Language Models for Statistical Inference: Context Augmentation with Applications to the Two-Sample Problem and Regression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KVR446R5}},
  note         = {Machine review of arXiv:2506.23862}
}
read the original abstract

We introduce context augmentation, a data-augmentation approach that uses large language models (LLMs) to generate contexts around observed strings as a means of facilitating valid frequentist inference. These generated contexts serve to reintroduce uncertainty, incorporate auxiliary information, and facilitate interpretability. For example, in the two-sample test, we compare the log-probability of strings under contexts from its own versus the other group. We show on synthetic data that the method's t-statistics exhibit the expected null behaviour while maintaining power and, through a replication, that the method is powerful and interpretable. We next introduce text-on-text regression. Contexts generated around the predictor string are treated as mediating variables between the predictor and outcome strings. Using negative controls, we then distinguish between semantic and syntactic dimensions of prediction. Analysis of real-world dialogic data illustrates behaviour predicted from a psycholinguistic framework. Theoretically, we provide identification conditions, derive an influence-function decomposition, and show that repeated cross-fitting of a pivotal statistic yields higher-order efficiency. We derive bounds linking estimation error, context count, and number of cross-fits. Taken together, context augmentation offers the ability to connect LLMs with longstanding statistical practice.

Figures

Figures reproduced from arXiv: 2506.23862 by the authors.

Figure 1
Figure 1. An LLM is used to generate multiple contexts around each observed string. For example, given the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. Intuitive overview of our context-augmentation approach. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Null calibration. QQ-plot of p-values from within-category comparisons. Alignment with the diagonal illustrates valid null calibration. while retaining lexical items; and jabberwocky, where content words are replaced with invented nonsense words, preserving an approximate syntactic scaffold. Let {(sx,i, sy,i)} n i=1 be n observations of predictor–outcome text pairs. For each pair, the LLM generates a set of latent c… view at source ↗
Figures from the paper (1 more)
Figure 3
Figure 3. Figure 3: Power. Violin plots of pairwise t-statistics for cross-category comparisons. Shaded regions represent non-significant t-statistics (|t| < 2). The distributions highlight clear differentiation among semantic categories. Predictor Strings Informative Outcome Strings Plac…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 52 canonical work pages

  1. [1]

    G. G. Agarwal and W. J. Studden. Asymptotic integrated mean square error using least squares and bias minimizing splines. Annals of Statistics, 8 0 (6): 0 1307--1325, Nov. 1980. doi:10.1214/aos/1176345203

    arXiv 1980

  2. [2]

    L. P. Argyle, E. C. Busby, N. Fulda, J. R. Gubler, C. Rytting, and D. Wingate. Out of one, many: Using language models to simulate human samples. Political Analysis, 31 0 (3): 0 337–351, 2023. doi:10.1017/pan.2023.2

    doi:10.1017/pan.2023.2 2023

  3. [3]

    Athey, J

    S. Athey, J. Tibshirani, and S. Wager. Generalized random forests. Annals of Statistics, 47 0 (2): 0 1148--1178, Apr. 2019

    work page 2019

  4. [4]

    P. D. Bartolomeis, J. Abad, G. Wang, K. Donhauser, R. M. Duch, F. Yang, and I. J. Dahabreh. Efficient randomized experiments using foundation models. arXiv preprint arXiv:2502.04262, 2025. Preprint

    arXiv 2025

  5. [5]

    Belloni, V

    A. Belloni, V. Chernozhukov, and C. Hansen. Inference on treatment effects after selection among high-dimensional controls. Review of Economic Studies, 81 0 (2): 0 608--650, 2014

    work page 2014

  6. [6]

    J. O. Berger, B. Liseo, and R. L. Wolpert. Integrated likelihood methods for eliminating nuisance parameters. Statistical Science, 14 0 (1): 0 1--28, February 1999. doi:10.1214/ss/1009211804

    arXiv 1999

  7. [7]

    Bhattacharya and R

    R. Bhattacharya and R. R. Rao. Normal Approximation and Asymptotic Expansions, volume 64 of Classics in Applied Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2010. ISBN 978-0-89871-644-3. doi:10.1137/1.9780898719895. Originally published by Wiley, 1976

    doi:10.1137/1.9780898719895 2010

  8. [8]

    P. J. Bickel. On adaptive estimation. Annals of Statistics, 10 0 (3): 0 647--671, 1982

    work page 1982

Show all 76 references
  1. [9]

    D. M. Blei and J. D. McAuliffe. Supervised topic models. In Advances in Neural Information Processing Systems (NIPS), volume 20, pages 121--128, 2007

    2007

  2. [10]

    D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation. Journal of Machine Learning Research, 3: 0 993--1022, 2003

    2003

  3. [11]

    Brown, B

    T. Brown, B. Mann, N. Ryder, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33: 0 1877--1901, 2020

    1901

  4. [12]

    E. J. Cand\` e s, Y. Fan, L. Janson, and J. Lv. Panning for gold: ``model--x’’ knockoffs for high-dimensional controlled variable selection. Journal of the Royal Statistical Society: Series B, 80 0 (3): 0 551--577, 2018. doi:10.1111/rssb.12265

    2018 doi

  5. [13]

    B. Chen, Z. Zhang, N. Langren\'e, and S. Zhu. Unleashing the potential of prompt engineering for large language models. Patterns, 6 0 (6): 0 101260, 2025. doi:10.1016/j.patter.2025.101260. Creative Commons Open Access

    2025

  6. [14]

    S. Chen, Y. Li, S. Lu, H. Van, H. J. W. L. Aerts, G. K. Savova, and D. S. Bitterman. Evaluating the chatgpt family of models for biomedical reasoning and classification. Journal of the American Medical Informatics Association: JAMIA, 31 0 (4): 0 940--948, 2024. doi:10.1093/jam...

    2024 doi

  7. [15]

    Chernozhukov, I

    V. Chernozhukov, I. Fern \'a ndez-Val, and A. E. Kowalski. Quantile regression with censoring and endogeneity. Journal of Econometrics, 186 0 (1): 0 201--221, 2015

    2015

  8. [16]

    Chernozhukov, D

    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, 21 0 (1): 0 C1--C68, 2018

    2018

  9. [17]

    Chernozhukov, J

    V. Chernozhukov, J. C. Escanciano, H. Ichimura, W. K. Newey, and J. M. Robins. Locally robust semiparametric estimation. Econometrica, 90 0 (4): 0 1501--1535, 2022. doi:10.3982/ECTA16294

    2022 doi

  10. [18]

    Chung, L

    H. Chung, L. Hou, S. Longpre, B. Zoph, Y. Tay, W. Fedus, E. Li, X. Wang, M. Dehghani, S. Brahma, A. Webson, S. Gu, Z. Dai, M. Suzgun, X. Chen, A. Chowdhery, S. Narang, G. Mishra, A. Yu, V. Zhao, Y. Huang, A. Dai, H. Yu, S. Petrov, E. Chi, J. Dean, J. Devlin, A. Roberts, D. Zho...

    arXiv

  11. [19]

    A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39 0 (1): 0 1--38, 1977

    1977

  12. [20]

    M. J. Denny and A. Spirling. Text preprocessing for unsupervised learning: Why it matters, when it misleads, and what to do about it. Political Analysis, 26 0 (2): 0 168--189, 2018

    2018

  13. [21]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL-HLT, pages 4171--4186, 2019

    2019

  14. [22]

    Egami, C

    N. Egami, C. J. Fong, J. Grimmer, and M. E. Roberts. How to make causal inferences using texts. Science Advances, 8 0 (6): 0 eabg2652, 2022

    2022

  15. [23]

    J. C. Escanciano and J. R. Terschuur. Machine learning inference on inequality of opportunity, 2022

    2022

  16. [24]

    M. H. Farrell, T. Liang, and S. Misra. Deep neural networks for estimation and inference. Econometrica, 89 0 (1): 0 181--213, 2021. doi:10.3982/ECTA16901

    2021 doi

  17. [25]

    Garrod and M

    S. Garrod and M. J. Pickering. Joint action, interactive alignment, and dialog. Topics in Cognitive Science, 1 0 (2): 0 292--304, 2009. doi:10.1111/j.1756-8765.2009.01020.x

    2009

  18. [26]

    Gelman and E

    A. Gelman and E. Loken. The statistical crisis in science. American Scientist, 102 0 (6): 0 460--465, 2014

    2014

  19. [27]

    Gentzkow, B

    M. Gentzkow, B. Kelly, and M. Taddy. Text as data. Journal of Economic Literature, 57 0 (3): 0 535--574, 2019. doi:10.1257/jel.20181020

    2019 doi

  20. [28]

    Gretton, K

    A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Sch \"o lkopf, and A. J. Smola. A kernel two-sample test. Journal of Machine Learning Research, 13: 0 723--773, 2012

    2012

  21. [29]

    C. Gu. Smoothing Spline ANOVA Models, volume 297 of Springer Series in Statistics. Springer, 2013

    2013

  22. [30]

    P. Hall. The Bootstrap and Edgeworth Expansion. Springer Series in Statistics. Springer, 1992. ISBN 978-0-387-97742-7. doi:10.1007/978-1-4612-4384-9

    1992 doi

  23. [31]

    Z. S. Harris. Distributional structure. Word, 10 0 (2-3): 0 146--162, 1954

    1954

  24. [32]

    N. E. Helwig and P. Ma. Smoothing spline anova for super-large samples: Scalable computation via rounding parameters. Statistics and Its Interface, 9 0 (4): 0 433--444, 2016

    2016

  25. [33]

    M. A. Hern\'an and J. M. Robins. Causal Inference: What If. CRC Press (Taylor & Francis Group), Boca Raton, FL, 1st edition, July 2025. ISBN 9781420076165

    2025

  26. [34]

    K. Imai, L. Keele, and D. Tingley. A general approach to causal mediation analysis. Psychological methods, 15 0 (4): 0 309, 2010

    2010

  27. [35]

    Karadzhov, T

    G. Karadzhov, T. Stafford, and A. Vlachos. Delidata: A dataset for deliberation in multi-party problem solving. arXiv preprint arXiv:2108.05271, 2023

    2023 arXiv

  28. [36]

    group‑conditional

    J. Kim, S. O'Hagan, and V. Rocková. Adaptive uncertainty quantification for generative ai. arXiv preprint arXiv:2408.08990, 2024. Includes tree‑based “group‑conditional” split‑conformal guarantees

    2024 arXiv

  29. [37]

    Kirchler, S

    M. Kirchler, S. Khorasani, M. Kloft, and C. Lippert. Two-sample testing using deep learning. In S. Chiappa and R. Calandra, editors, Proceedings of the 23rd International Conference on Artificial Intelligence and Statistics (AISTATS), volume 108 of Proceedings of Machine Learn...

    2020

  30. [38]

    Kumar, C

    B. Kumar, C. Lu, G. Gupta, A. Palepu, D. Bellamy, R. Raskar, and A. Beam. Conformal prediction with large language models for multi-choice question answering. In Proceedings of the ICML 2023 Workshop on Machine Learning for Health (ML4H), 2023. arXiv:2305.18404

    2023 arXiv

  31. [39]

    Le Mens, B

    G. Le Mens, B. Kov \'a cs, M. T. Hannan, and G. Pros. Uncovering the semantics of concepts using gpt-4. Proceedings of the National Academy of Sciences, 120 0 (49): 0 e2309350120, 2023

    2023

  32. [41]

    L. Li, E. T. Tchetgen, A. van der Vaart, and J. M. Robins. Higher order inference on a treatment effect under low regularity conditions. Statistics & probability letters, 81 0 (7): 0 821--828, Jul 2011. ISSN 0167-7152. doi:10.1016/j.spl.2011.02.030

    2011 doi

  33. [42]

    C. F. Manski. Identification of endogenous social effects: The reflection problem. The Review of Economic Studies, 60 0 (3): 0 531--542, 1993

    1993

  34. [43]

    Mendelberg, C

    T. Mendelberg, C. F. Karpowitz, and J. B. Oliphant. Gender inequality in deliberation: Unpacking the black box of interaction. Perspectives on Politics, 12 0 (1): 0 18--44, 2014

    2014

  35. [44]

    S. A. Murphy and A. W. van der Vaart. On profile likelihood. Journal of the American Statistical Association, 95 0 (450): 0 449--465, 2000. doi:10.1080/01621459.2000.10474219

  36. [45]

    Noy and W

    S. Noy and W. Zhang. Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381 0 (6654): 0 187--192, July 2023. doi:10.1126/science.adh2586

    2023 doi

  37. [46]

    GPT -4 Technical Report , Mar

    OpenAI. GPT -4 Technical Report , Mar. 2023. URL https://cdn.openai.com/papers/gpt-4.pdf

    2023

  38. [47]

    J. Pearl. The causal mediation formula: a guide to the assessment of pathways and mechanisms. Prevention Science, 13 0 (4): 0 426--436, 2012

    2012

  39. [48]

    Pesarin and L

    F. Pesarin and L. Salmaso. Permutation Tests for Complex Data: Theory, Applications and Software. Wiley Series in Probability and Statistics. Wiley, 2010. ISBN 9780470516416. doi:10.1002/9780470689516

    2010 doi

  40. [49]

    M. J. Pickering and S. Garrod. Toward a mechanistic psychology of dialogue. Behavioral and Brain Sciences, 27 0 (2): 0 169--225, 2004

    2004

  41. [50]

    M. J. Pickering and S. Garrod. An integrated theory of language production and comprehension. Behavioral and Brain Sciences, 36 0 (4): 0 329--347, 2013

    2013

  42. [51]

    D. N. Politis, J. P. Romano, and M. Wolf. Subsampling. Springer Series in Statistics. Springer, New York, 1999

    1999

  43. [52]

    Pretorius and J

    C. Pretorius and J. W. H. Swanepoel. On the asymptotic theory of new bootstrap confidence bounds. The Annals of Statistics, 46 0 (1): 0 438--456, 2018. doi:10.1214/17-AOS1557

    2018 doi

  44. [53]

    Rauba, Q

    P. Rauba, Q. Wei, and M. van der Schaar. Distribution-based perturbation analysis for large language models. arXiv preprint arXiv:2412.00868, 2024. Statistical Foundations of LLMs and Foundation Models Workshop at NeurIPS 2024

    2024 arXiv

  45. [54]

    E. H. Rho, M. Harrington, Y. Zhong, and J. L. Eberhardt. Escalated police stops of black men are linguistically and psychologically distinct in their earliest moments. Proceedings of the National Academy of Sciences, 2023

    2023

  46. [55]

    M. E. Roberts, B. M. Stewart, D. Tingley, C. Lucas, J. Leder-Luis, S. K. Gadarian, B. Albertson, and D. G. Rand. Structural topic models for open-ended survey responses. American Journal of Political Science, 58 0 (4): 0 1064--1082, October 2014. doi:10.1111/ajps.12103

    2014 doi

  47. [56]

    M. E. Roberts, B. M. Stewart, and R. A. Nielsen. Adjusting for confounding with text matching. American Journal of Political Science, 64 0 (4): 0 887--903, October 2020. doi:10.1111/ajps.12526

    2020 doi

  48. [57]

    Robins, L

    J. Robins, L. Li, E. Tchetgen, and A. van der Vaart. ``higher order influence functions and minimax estimation of nonlinear functionals''. In Probability and Statistics: Essays in Honor of David A. Freedman, pages 335--421. Institute of Mathematical Statistics, 2008

    2008

  49. [58]

    J. M. Robins and A. Rotznitzky. Semiparametric efficiency in multivariate regression models with missing data. Journal of the American Statistical Association, 90 0 (429): 0 122--129, 1995

    1995

  50. [59]

    Rodriguez, A

    P. Rodriguez, A. Spirling, and B. Stewart. Embedding regression: Models for context-specific description and inference. American Political Science Review, 117 0 (4): 0 1255--1274, Nov. 2023

    2023

  51. [60]

    Romano, M

    Y. Romano, M. Sesia, and E. J. Cand\` e s. Deep knockoffs: Controlled feature selection using generative models. Journal of the American Statistical Association, 115 0 (532): 0 1861--1872, 2019. doi:10.1080/01621459.2019.1660174

    2019

  52. [61]

    D. B. Rubin. Inference and missing data. Biometrika, 63 0 (3): 0 581--592, 1976

    1976

  53. [62]

    Sahlgren

    M. Sahlgren. The distributional hypothesis. Italian Journal of Linguistics, 20: 0 33--53, 01 2008

    2008

  54. [63]

    Sahoo, A

    P. Sahoo, A. K. Singh, S. Saha, V. Jain, S. Mondal, and A. Chadha. A systematic survey of prompt engineering in large language models: techniques and applications. arXiv preprint arXiv:2402.07927, 2024

    2024 arXiv

  55. [64]

    Schrab, I

    A. Schrab, I. Kim, M. Albert, B. Laurent, B. Guedj, and A. Gretton. MMD aggregated two-sample test. Journal of Machine Learning Research, 24: 0 1--81, 2023

    2023

  56. [65]

    N. H. Shah, D. Entwistle, and M. A. Pfeffer. Creation and adoption of large language models in medicine. JAMA, 330 0 (9): 0 866--869, 2023. doi:10.1001/jama.2023.14217

    2023

  57. [66]

    Shao and C

    J. Shao and C. F. J. Wu. A general theory for jackknife variance estimation. The Annals of Statistics, 17 0 (3): 0 1176--1197, 1989

    1989

  58. [67]

    Strobl, A.-L

    C. Strobl, A.-L. Boulesteix, A. Zeileis, and T. Hothorn. Bias in random forest variable importance measures: Illustrations, sources and a solution. BMC Bioinformatics, 8 0 (25), 2007. doi:10.1186/1471-2105-8-25

    2007 doi

  59. [68]

    M. Taddy. Multinomial inverse regression for text analysis. Journal of the American Statistical Association, 108 0 (503): 0 755--770, 2013

    2013

  60. [69]

    M. A. Tanner and W. H. Wong. The calculation of posterior distributions by data augmentation. Journal of the American Statistical Association, 82 0 (398): 0 528--540, 1987

    1987

  61. [70]

    van der Vaart

    A. van der Vaart. Higher order tangent spaces and influence functions. Statistical Science, 29 0 (4): 0 679--686, 2014

    2014

  62. [71]

    A. W. van der Vaart. Asymptotic Statistics, volume 3 of Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge, UK; New York, NY, USA, 1998

    1998

  63. [72]

    A. W. van der Vaart and J. A. Wellner. Weak convergence and empirical processes: With applications to statistics. 1996

    1996

  64. [73]

    Veitch, D

    V. Veitch, D. Sridhar, and D. M. Blei. Adapting text embeddings for causal inference. In Advances in Neural Information Processing Systems, volume 33, pages 16446--16460, 2020

    2020

  65. [74]

    Wager and S

    S. Wager and S. Athey. Estimation and inference of heterogeneous treatment effects using random forests. Journal of the American Statistical Association, 113 0 (523): 0 1228--1242, 2018. doi:10.1080/01621459.2017.1319839

    2018

  66. [75]

    Wang and B

    Q. Wang and B. G. Lindsay. Variance estimation of a general u-statistic with application to cross-validation. Statistica Sinica, 24 0 (3): 0 1117--1141, 2014

    2014

  67. [76]

    Wang and D

    Y. Wang and D. M. Blei. Counterfactual prediction using deep latent variable models. Journal of the American Statistical Association, 116 0 (532): 0 1555--1572, 2021

    2021

  68. [77]

    J. Xu, C. Qian, M. Liu, and F. Guo. A two‑sample test of text generation similarity. arXiv preprint arXiv:2505.05269, 2025

    2025 arXiv

Pith tools

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