acceptodds
Under review as a conference paper at ICLR 2027

MASI: MECHANISM-AWARE INTERVAL INFERENCE WITH LLM SEMANTIC COVARIATES

Abstract

Insufficient sample size is a primary obstacle to precise inference, especially when observations pair text covariates with scalar outcomes. Unlike structured covariates with prespecified coordinates and interpretable perturbations, text lacks a canonical low-dimensional representation, and routine edits need not preserve outcome-relevant semantics. Modern large language models can generate source-conditioned texts, creating a new possibility for addressing this difficulty. Yet distribution shift and within-source dependence can make naive use biased and overconfident. We introduce Model-assisted Semantic Inference (MaSI), which uses generated texts to construct a more precise confidence interval for the parameter of interest while controlling the inferential effects of LLM-induced distribution shift under the stated calibration conditions. Specifically, the first set is generated under a scheme that preserves source content while varying expression style; together with real labeled data, it is used to learn outcome-relevant content coordinates and complementary style coordinates. A separate evaluation set constructs a calibrated auxiliary statistic in those coordinates. Density-ratio calibration addresses distribution shift, real-label residual correction preserves the population target, and uncertainty is estimated across original sources. Under explicit representation, calibration, and covariance-control conditions, MaSI provides consistent variance estimation and asymptotically valid confidence intervals. Simulations and empirical-data experiments show shorter intervals with near-nominal coverage in the primary settings, establishing a principled route to more precise inference from scarce text–outcome pairs.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
prediction-powered inference, uncertainty quantification, content–style representation learning, LLM semantic augmentation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_HcmKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors develop MaSI to estmate a population mean using labeled text and LLM-generated covariates. The method learns semantic features, reweights generated samples, and applies a residual correction. Simulations and text experiments report shorter confidence intervals.

Critical strengths

The main strength is the treatment of dependence among texts generated from the same source; the negative controls and variance sensitivity analysis make it easy to assess how much of the reported precision gain survives that dependence.

Critical weaknesses
  1. Accounting for the observed positive covariance increases interval-width ratios from about 0.93 to 0.96–0.98 (Table F.20). This really reduces the practical gains..

  2. The interval guarantee assumes that errors in the learned weights become negligible, but this is not established for the estimator used. The simulation also gives the generator true population parameters; please explain what comparable information supports the gains in real data.

Questions for authors

Please report intervals that account for source covariance and ratio estimation. What assumption connecting the generator to the target population establishes the required calibration rate and supports the remaining precision gain?

6 comments
AuthorsOpenReview ↗

Response to Reviewer HcmK [1/4]

1. How does density-ratio estimation error enter MaSI? What conditions does the interval guarantee impose on weight learning?

The theoretical goal of this paper is to connect learning and inference through statistical error conditions on the final fitted quantities, providing a common framework for different representation and weight-learning rules and directly analyzing the final MaSI statistic with estimated weights. Weight error enters the statistic through a centered sampling term and a population prediction-mean bias: the former is controlled by L2L_2 consistency of the weights under cross-fitting, whereas the latter requires βw,n=op(n−1/2)\beta_{w,n}=o_p(n^{-1/2}). Appendix DD further gives sufficient conditions for controlling model approximation bias, a first-order parameter effect, and a second-order remainder. Thus, under all the inference conditions in this paper, learning rules meeting these error requirements have the corresponding first-order expansion and interval guarantee; the density-ratio function class, regularization, and solver can be chosen for the data. For the implementation used in this paper, Response 22 reports empirical coverage and interval width using estimated weights and including source cross-covariance, and evaluates the finite-sample effect of weight estimation through paired replacement with true weights in the simulation.

(1) Centered sampling error and weight consistency.

Fix the representation and outcome predictor learned in each fold. Using estimated weights and the true density ratio gives the exact decomposition

μ^(w^)−μ^(w∘)=∑kπk(Qk,n−Q)(φ^k−φk∘)+Rw,n,\widehat\mu(\widehat w)-\widehat\mu(w^\circ) = \sum_k\pi_k(\mathbb Q_{k,n}-Q) (\widehat\varphi_k-\varphi_k^\circ) + R_{w,n},

where φ^k=w^kh^k\widehat\varphi_k=\widehat w_k\widehat h_k and φk∘=wk∘h^k\varphi_k^\circ=w_k^\circ\widehat h_k, with all functions evaluated in the content coordinates of the corresponding fold; Qk,n\mathbb Q_{k,n} averages first within each source and then across held-out sources.

The first term is centered sampling error. Cross-fitting separates held-out sources from the training fits while allowing dependence among texts from the same source. Appendix DD proves that this term, multiplied by n\sqrt n, is Op(εw,n)O_p(\varepsilon_{w,n}), where εw,n\varepsilon_{w,n} is the fold-weighted root-mean-square L2L_2 error of the density ratio. Weight consistency therefore suffices to control this component.

(2) Calibration rate for the population prediction mean.

Appendix DD defines the prediction-direction calibration error βw,n\beta_{w,n} and proves, under the local regularity conditions of Proposition D.2D.2, that

∣Rw,n∣≤βw,n≤aβ,n+χw,nrγ,n+CwMhrγ,n2.|R_{w,n}| \le \beta_{w,n} \le a_{\beta,n} +\chi_{w,n}r_{\gamma,n} +C_wM_h r_{\gamma,n}^{2}.

Here, aβ,na_{\beta,n} measures model approximation bias in the prediction direction, rγ,nr_{\gamma,n} measures parameter estimation error, and χw,n\chi_{w,n} measures the first-order sensitivity of the prediction mean to parameter changes. The theorem requires this upper bound to be op(n−1/2)o_p(n^{-1/2}). Ordinary weight consistency and the prediction-mean calibration accuracy required for inference are thus distinct requirements, and cross-fitting alone does not eliminate population calibration error. Together with the remaining conditions on representation learning and outcome fitting, these controls yield the fixed-center first-order expansion in Theorem 33; the moment and covariance conditions in Appendix EE then give the corresponding interval guarantee. The theory applies when the final weights satisfy these learning and calibration conditions.

2. How much precision gain does MaSI retain after accounting for source covariance? How does density-ratio estimation affect these results?

After accounting for source cross-covariance, MHS intervals remain 3.53.5\\% shorter than Classical on average. A paired simulation with matched marginal distributions fixes the samples, representations, and predictors, replaces estimated weights with true weights, and computes Main and Block under each. It shows similar width gains under the two weighting schemes, with gains retained after including the cross term in this simulation setting.

(1) Covariance correction and remaining gains in MHS.

Table F.20F.20 reports Main and Block for the same MaSI point estimate: both retain dependence among generated texts from the same source, and Block additionally includes cross-covariance between the real residual and the synthetic-group mean prediction. At nominal 9090\\% coverage, the average width ratio across five sample sizes changes from 0.9330.933 to 0.9650.965, and average empirical coverage changes from 89.789.7\\% to 90.490.4\\%. Appendix F.4F.4, Table F.8F.8 shows that the estimated cross term relative to additive variance decreases from 0.1130.113 at n=100n=100 to 0.0550.055 at n=800n=800; in Table F.20F.20, the mean width increase of Block relative to Main also decreases from about 4.94.9\\% to 2.72.7\\%.

AuthorsOpenReview ↗

Response to Reviewer HcmK [2/4]

The theory considers covariance tending to 00 under the corresponding conditions, with Main and Block becoming closer. Main presents the inference guarantee, precision criterion, and variance budget under the structure studied; when the cross term is zero or very small, omitting its estimation may also reduce finite-sample variance-estimation noise. With the same point estimate and critical value, including positive covariance makes Block wider than Main, while negative covariance makes it narrower. The positive cross term in MHS explains the widening there; the paired simulation below also includes cases with narrower Block intervals.

The true density ratio in MHS is unknown, so replacement with true weights cannot isolate the effect of weight learning. We therefore add the following paired simulation with a known true density ratio to distinguish weight estimation from covariance treatment.

(2) Paired weight and covariance comparisons under matched marginal distributions.

The simulated text vectors contain 55 dimensions of score-related content, 3535 dimensions of style factors unrelated to the outcome, and measurement noise. The outcome is a nonlinear function of content, including an interaction between two content factors and independent label noise. Pure style factors have large marginal variances, so many high-variance directions in the raw vectors reflect expression factors unrelated to the score. This tests the role of extracting score-related content under strong style interference.

Each source generates two groups of 1010 observations: the first preserves content and changes style for representation learning; the second introduces same-source variation along outcome-related directions to construct the synthetic-group mean prediction while matching the real-data population marginal covariate distribution. The true density ratio is therefore w=1w=1. We use five-fold cross-fitting and 200200 paired repetitions at each of n=200,400,800n=200,400,800. Within each repetition:

  • Fit the content representation and outcome predictor once using the same real sample, generated data, and fold split.

  • Fix these fits and use either the density ratio estimated on the training folds or the known true density ratio.

  • Compute the MaSI point estimate under each weighting scheme; different weights may give different point estimates. Within each scheme, construct Main and Block, which retains source cross-covariance, around that scheme's same point estimate.

Each cell below gives empirical coverage / mean interval-width ratio relative to Classical, at nominal 9090\\% coverage.

nnDensity ratioMainBlock
200200Estimated density ratio88.588.5\\% / 0.76488.088.0\\% / 0.750
200200Oracle density ratio90.090.0\\% / 0.76489.089.0\\% / 0.751
400400Estimated density ratio90.090.0\\% / 0.75289.089.0\\% / 0.749
400400Oracle density ratio90.590.5\\% / 0.75289.589.5\\% / 0.749
800800Estimated density ratio91.591.5\\% / 0.75091.591.5\\% / 0.750
800800Oracle density ratio93.593.5\\% / 0.75093.593.5\\% / 0.750

At all three sample sizes, replacing the weights with the true density ratio leaves interval widths essentially unchanged, and the reported Block widths do not exceed Main widths. For example, at n=800n=800, both weighting schemes give a Block width ratio of about 0.7500.750, approximately 2525\\% shorter than Classical, with coverage of 91.591.5\\% and 93.593.5\\%, respectively. Estimated weights already yield coverage close to nominal and about 2525\\% interval shortening; Oracle replacement gives no clear width improvement. This provides empirical sensitivity evidence under matched marginal distributions: shortening in this simulation does not require advance knowledge of the true weights and persists after including source cross-covariance. Table F.20F.20 directly evaluates the remaining gains in MHS.

AuthorsOpenReview ↗

Response to Reviewer HcmK [3/4]

3. Known population parameters in the simulation can guide favorable generation. When these parameters are unknown in real applications, how can generation be guided and the corresponding quality of generated texts be assessed?

Real applications use the task definition and source context to specify semantic generation rules, then use real labels and source-pairing diagnostics to assess statistical usefulness, without first recovering the population parameters in the simulation. Good data must serve each channel's purpose: the style channel changes expression while keeping rating cues relatively stable, and the evaluation channel provides predictive local semantic variation. Fluency and textual diversity alone do not establish these properties.

(1) Task-specific semantic rules and the scope of generation.

In MHS, hostility, attack targets, author stance, irony, and quotation attribution may affect real ratings. Even without their numerical coefficients, the task definition allows the permitted changes to be specified. For QstyleQ_{\mathrm{style}}, instructions change sentence structure and wording while preserving semantic cues that determine the rating. For example, when a comment quotes offensive language to refute it, the rewrite must preserve quotation attribution and the author's opposing stance. For QevalQ_{\mathrm{eval}}, instructions allow local semantic variation around hostility or ambiguity cues in the source while preserving targets, stance, facts, and context, and avoiding unsupported new attack targets or facts. These rules are specified in advance; the generator receives no real labels, numerical ratings, or target-population parameters.

(2) Style distinguishability and prediction drift.

Table F.11F.11, Panel A jointly diagnoses expression changes and the stability of outcome-related content: classification AUC between originals and style rewrites is 0.6430.643 in the raw embeddings and decreases to about 0.4950.495--0.4980.498 in the learned content coordinates; style-class identification accuracy in the remaining expression coordinates is 0.6030.603, above chance at 0.2000.200. This supports retention of expression differences with reduced distinguishability between originals and rewrites in the content coordinates. The absolute mean prediction drift is 0.0140.014--0.0220.022, further assessing how rewriting affects outcome prediction.

(3) Real-label prediction, source-group permutation, and calibration diagnostics.

We assess task-related information in QevalQ_{\mathrm{eval}} using real labels rather than relying solely on LLM judgments of text plausibility. Appendix F.4F.4, Table F.11F.11, Panel B keeps the synthetic text pool fixed, permutes only the correspondence between complete synthetic groups and real sources, and compares prediction on held-out real labels. Predictive R2R^2 is higher under correct pairing in every configuration, with differences of 0.04070.0407--0.12110.1211, showing that the predictive cues relate to the true outcomes of their specific sources. Prediction-mean differences and weight-stability diagnostics further assess whether these texts support stable calibration. These checks use real labels and source relationships to assess the practical value of semantic changes and provide empirical evidence for generation quality.

Reviewer_kBDBOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MaSI, a framework for estimating population means from limited text–outcome data using LLM-generated texts as auxiliary information. It combines constrained representation learning, density-ratio calibration, and real-label residual correction to account for distribution shift and dependence among generated texts. Theoretical analysis and experiments suggest that MaSI can improve interval precision under certain conditions, while further clarification of its representation learning and second-generation design would help better establish its benefits.

Critical strengths

1.The paper addresses an interesting problem of using LLM-generated texts for statistical inference with limited labeled data. The proposed framework accounts for distribution shift and dependence among texts generated from the same source, providing a coherent approach to improving interval estimation. 2.The paper provides theoretical analysis of representation learning, estimator asymptotics, and confidence interval validity, supported by simulations and real-data experiments. The results show that MaSI can improve interval precision while maintaining reasonable coverage in several settings.

Critical weaknesses
  1. The representation ablations in Table F.18 and the budget-matched comparisons in Table F.20 are useful. However, the four representation constraints are removed jointly, leaving their individual contributions unclear. The benefit of representation learning also needs to be separated more clearly from that of density-ratio calibration through a PCA-based comparison with matched evaluation texts and downstream inference procedures.
  2. Although the two generation roles are described, concrete text examples are needed to show how Q_eval differs from ordinary style rewriting and what outcome-relevant structure it exposes. This would make the generation design and its practical implementation easier to understand.
  3. The primary additive variance estimator relies on residual–auxiliary covariance being zero or asymptotically negligible. The appendix provides relevant diagnostics and covariance-inclusive block intervals, but the practical motivation for preferring the additive estimator remains insufficiently established. Table F.20 shows that including covariance reduces the interval-width advantage, making the robustness of the precision gains an important consideration.
Questions for authors
  1. Please provide examples of an original text and its corresponding Q_style and Q_eval generations. What specific changes does Q_eval make, and what outcome-relevant information in the source becomes more accessible to the fitted predictor?
  2. Table F.18 removes all four representation constraints jointly. Please provide individual constraint ablations and a PCA-based comparison that matches the evaluation texts, outcome learner, density-ratio calibration, and source-level variance estimation. Please also clarify which data regimes require the proposed representation learning rather than a simpler calibrated baseline.
  3. Given the covariance diagnostics and covariance-inclusive block intervals already reported, why is the additive variance estimator preferred for the primary results? Please discuss how much of the precision gain remains when residual–auxiliary covariance is included, and how practitioners should assess whether the covariance-control condition is sufficiently accurate.
13 comments
AuthorsOpenReview ↗

Response to Reviewer kBDB [1/11]

1. Please provide examples of an original text and its corresponding QstyleQ_{\mathrm{style}} and QevalQ_{\mathrm{eval}} texts. What exactly does QevalQ_{\mathrm{eval}} change? Which outcome-related information becomes easier for the fitted predictor to use?

Thank you for this suggestion. QstyleQ_{\mathrm{style}} changes expression while retaining outcome-related content, supporting content representations that are stable to changes in expression. QevalQ_{\mathrm{eval}} allows local semantic changes near the original meaning, organizing source cues into task-related explanations and connections. In general, the predictor can further use outcome-related semantic attributes and their levels, combinations of multiple attributes, and how context affects their meaning. Through local semantic changes and elaboration near the original meaning, evaluation generation may add predictive information that a single encoding and finite-dimensional working representation of the original text do not fully retain. Real labels establish how these features relate to real ratings, and the MHS source-matching diagnostic further examines their incremental predictive value.

(1) Two generation channels for the same source. We retain one formal-review pair below; examples in all three expression styles are provided in Appendix G.1G.1 of the revised paper (pp. 7373–7474).

Original text: The service at this cafe is quite good, and the coffee tastes good too. There is a lot of wooden decor, and the setting is relatively quiet. The menu says that the coffee beans are freshly roasted. Overall, it feels quite good.

(1.1) QstyleQ_{\mathrm{style}}: formal review. Evaluating this cafe in terms of service quality and coffee taste, it performs well in both respects. The interior has wooden decor and a relatively quiet setting; the menu states that freshly roasted coffee beans are used. Taking these experiences together, my overall evaluation of the cafe is satisfactory.

(1.2) QevalQ_{\mathrm{eval}}: formal review. The service and coffee taste at this cafe are both satisfactory. The wooden decor and quiet setting define its spatial character. For customers who wish to relax or talk quietly, these environmental factors may contribute importantly to the overall experience.

In this example, QstyleQ_{\mathrm{style}} changes the original everyday language into a formal review while retaining the original evaluations of service, coffee taste, decor, and the setting. QevalQ_{\mathrm{eval}} also uses a formal-review style, while expanding the content from separate descriptions of the experience to how the setting meets needs such as relaxing or having a quiet conversation, adding suitability of the setting as an evaluative relationship. The two texts have the same expression style, and the evaluation text further expands the evaluative content and relationships between attributes. This expansion is grounded in source cues, and its allowed scope is specified by the generation instructions for the task.

(2) Which outcome-related information becomes easier for the predictor to use? For an externally annotated overall experience rating, predictive cues include individual evaluations such as good service and good coffee, relationships between the setting and its uses, the basis for quality judgments, and how several attributes relate to the overall rating. The LLM forms these features from the source and its existing semantic knowledge; they may provide incremental predictive information relative to a single encoding and finite-dimensional working representation of the original text. Real labels determine their relationship with the outcome and their importance. MaSI maps evaluation texts into common content coordinates for prediction, density-ratio calibration, and within-source aggregation.

(3) Do synthetic texts contain predictive information related to their real source? In MHS, we use Ridge and Extra Trees to examine linear and nonlinear predictive relationships. We keep real reviews, real ratings, and the synthetic text pool unchanged, and compare adding each review's own evaluation-group features with adding another review's evaluation-group features. Both configurations predict real ratings on real samples excluded from training. Appendix F.4F.4, Table F.11F.11, Panel B shows that both models perform better with correct source matching at every examined sample size. For example, at n=800n=800, the held-out R2R^2 of Extra Trees is 0.66280.6628 with correct matching and 0.60210.6021 after shuffling, a difference of 0.06070.0607; a larger R2R^2 indicates better prediction. Under the same configuration, R2R^2 is 0.64190.6419 using real-review features alone and rises to 0.66280.6628 after adding the review's own evaluation-group features. These results show that the predictive value of evaluation texts is source-specific and supplements real-review features in this configuration.

AuthorsOpenReview ↗

Response to Reviewer kBDB [2/11]

2. After using the same density-ratio calibration and downstream inference procedure, what incremental benefit does the MaSI content representation provide over PCA? Under which data conditions is this representation learning worthwhile?

After matching evaluation texts, outcome learning, density-ratio calibration, and source-group variance estimation, representation benefits mainly arise when high-variance style factors cause PCA to omit outcome directions or interaction information. MaSI retains more outcome signal, reduces residuals, and shortens intervals; this difference remains with true density ratios. The matched MHS comparison also shows incremental benefits. If fixed PCA already retains the main predictive information and calibration is stable, a simpler calibrated baseline may suffice. Following your suggestion, we added information-retention diagnostics, PCA comparisons with matched downstream inference, true-density-ratio comparisons, and MHS experiments in Appendix G.3G.3 of the revised paper.

The methods share synthetic data, fit outcome models and density ratios under the same rules, and use the same variance calculation, changing only the representation supplied to these steps. We examine information retention, finite-sample prediction, and final interval precision in turn.

(1) Theoretically, which errors are affected by content coordinates? A content representation may affect two parts of prediction error: how much outcome information it omits, and whether the predictor can accurately learn outcome relationships from the retained information. Decomposing the squared error of outcome prediction distinguishes these two parts. Following the paper's notation, U=ψ(X)U=\psi(X) is the text vector, Z=V⊤UZ=V^\top U is the content coordinate vector, and h(Z)h(Z) is the scalar outcome prediction. Conditional on the training fold, we treat the fitted mapping and predictor as fixed objects. When the relevant second moments are finite, we have $\mathbb E\!\left

Y-\mathbb E(Y\mid U)\mid U\right

=0,$ and since ZZ is a function of UU, we have $\mathbb E\!\left

\mathbb E(Y\mid U)-\mathbb E(Y\mid Z)\mid Z\right

=0.$ Therefore,

�2� = \mathbb E �3� + \mathbb E �4� .$$ Further, $$\begin{aligned} \mathbb E �5� ={}& \mathbb E �6� \\\\ &+ \mathbb E �7� . \end{aligned}$$ Thus, $$\begin{aligned} \mathbb E �8� ={}&\mathbb E �9� \\\\ &+\mathbb E �10� \\\\ &+\mathbb E �11� . \end{aligned}$$ The final three terms correspond, respectively, to outcome variation that the text vector cannot yet explain, outcome-predictive information omitted by the content coordinates, and the approximation and estimation errors of the predictor in these coordinates. Density-ratio calibration adjusts differences between real and synthetic distributions in the selected coordinates; it cannot recover omitted outcome directions. PCA selects coordinates by population variance and may discard directions with smaller variance that are closely related to the outcome. MaSI combines expression stability with real-label supervision to retain this information, with Theorems $1$–$2$ controlling content-projection learning error under the corresponding conditions. The simulations separately examine the information retained by the coordinates, the predictor's use of that information, and the inference benefits after applying a matched calibration procedure. **(2) New mechanism simulations: from information retention to practical inference benefits.** The true content factors usually cannot be directly observed in real texts. We use numerical simulations in which these factors and their relationships with the outcome are known to directly examine the information retained by different representations. The $40$-dimensional observed vector is generated jointly by five outcome-related content factors, thirty-five outcome-irrelevant style factors, and independent measurement error. The outcome depends on both individual content factors and an interaction between the first and third content factors: the effect of one factor varies with the value of the other, so both directions need to be retained. We keep the content distribution, outcome relationship, and generation rules fixed and vary only the vector variation due to style factors across weak, moderate, and strong levels; these levels describe style interference. Appendix $G$, Table $G.3$ (p. $77$) gives the full population PCA ranks: under weak, moderate, and strong interference, $\text{PCA-5}$ omits one, three, and five content directions, respectively; under strong interference, increasing to $10$ or $20$ dimensions still omits three content directions.
AuthorsOpenReview ↗

Response to Reviewer kBDB [3/11]

PCA-5\text{PCA-5} has the same coordinate dimension as MaSI; PCA-10\text{PCA-10}, PCA-20\text{PCA-20}, and the full 4040-dimensional coordinates examine whether increasing dimension can compensate for omitted information. The known-content comparator uses the true five-dimensional content projection, while still fitting its outcome predictor from real training labels. Each condition examines n=200,400,800n=200,400,800, with 200200 repetitions per cell. The representations share real samples, evaluation texts, and fold splits, and use the same quadratic ridge regression and training-side fitting rules in their respective coordinates.

(2.1) Can these coordinates predict real outcomes? Each sample receives a prediction from a model that did not use it for training. In each repetition, we calculate R2R^2 using all held-out predictions and then average across repetitions; within the same dataset, a larger R2R^2 means a smaller prediction error. Appendix GG, Table G.4G.4 (p. 7777) reports all style conditions and sample sizes. Under weak interference, PCA-5\text{PCA-5} and MaSI perform similarly because the former omits only one relatively weak content factor. As interference increases, PCA-5\text{PCA-5} omits more content factors and its predictive performance declines substantially. Under strong interference, from n=200n=200 to 800800, MaSI's R2R^2 increases from 0.9350.935 to 0.9550.955, whereas that of PCA-20\text{PCA-20} increases from 0.2040.204 to 0.7750.775, leaving a clear gap. Increasing the input dimension also increases the individual and interaction relationships that the quadratic predictor must learn from limited labels; PCA-20\text{PCA-20} does not recover the three content directions omitted under strong interference and instead predicts less accurately than PCA-10\text{PCA-10} in small samples. Using all 4040 coordinates also does not match the content representation. This is consistent with the distinction between "whether information is retained" and "whether the predictor can be learned well in finite samples."

(2.2) How much outcome-related signal do the learned coordinates retain? Predictive performance is affected by both representation and fitting, so we further use simulation ground truth to examine outcome-signal retention. The key content direction is the 33rd direction; the two-factor interaction is the joint effect of the 11st and 33rd content factors; and the overall outcome signal is the scalar signal MM formed by the five content factors through the outcome-generating function, excluding independent label noise. Let the unit vector of the key direction be v⋆v^\star and the learned coordinate basis be V^\widehat V; its projection retention is Rdir=∣V^⊤v⋆∣22R_{\mathrm{dir}}=\\|\widehat V^\top v^\star\\|_2^2. A value of 11 means that the learned space contains the true direction completely, and 00 means complete exclusion. Let Z=V^⊤UZ=\widehat V^\top U and let TT denote the product of the two centered content factors. The other two signal-retention rates are

Rint=Var⁡E(T∣Z)Var⁡(T),Rall=Var⁡E(M∣Z)Var⁡(M).R_{\mathrm{int}} = \frac{\operatorname{Var}\\{\mathbb E(T\mid Z)\\}} {\operatorname{Var}(T)}, \qquad R_{\mathrm{all}} = \frac{\operatorname{Var}\\{\mathbb E(M\mid Z)\\}} {\operatorname{Var}(M)}.

In these formulas, the denominator is the total variance of the original signal, and the numerator is the signal variance explained by knowing only the learned coordinates. The former examines the specified two-factor interaction, whereas the latter combines the linear, nonlinear, and interaction components of the outcome function. The calculation uses the known generative mechanism, so it measures the information-retaining ability of the coordinates themselves, separately from the R2R^2 actually achieved by a finite-sample predictor.

Appendix GG, Table G.5G.5 (p. 7878) reports all diagnostics under the strong-style condition. At all three sample sizes, MaSI retains nearly all of the key direction, interaction, and overall outcome signal, with retention improving as the sample size increases and corresponding smaller held-out residuals. For example, at n=800n=800, the interaction-retention rates of MaSI and PCA-20\text{PCA-20} are 98.898.8\\% and 14.514.5\\%, their overall outcome-signal retention rates are 99.499.4\\% and 89.189.1\\%, and their held-out residual variances are 0.2010.201 and 0.9970.997, respectively. PCA-20\text{PCA-20} still retains other important outcome components, so an interaction-retention rate of 14.514.5\\% cannot be interpreted as retaining only 14.514.5\\% of all outcome information, nor can it be directly subtracted from the overall retention rate.

Reviewer_T7LMOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces MaSI, a method for improving confidence intervals when only a small number of texts with known outcomes are available. The approach uses a language model to generate alternative versions of observed texts and extract useful information from them. It separates information related to the outcome from changes in writing style and adjusts for differences between real and generated data. The method also accounts for the dependence between texts generated from the same original source. The authors provide theoretical results showing that the confidence intervals are valid under specific assumptions. Simulations and text experiments show improved precision in several settings while generally maintaining the intended coverage.

Critical strengths

I appreciate that the paper clearly shows that generated texts are not new population draws, and that the independent text control, which treats them as independent, covers only 0.782 instead of 0.90. I think this is an important warning for the many people who want to use LLM augmentation for inference. The second strength is the honest experimental reporting, including settings where MaSI gives no advantage, such as the controlled rendering study, and many negative controls where uncalibrated PPI is biased. This honesty makes the paper more trustworthy.

Critical weaknesses

As I understand it, the theoretical guarantees depend on several non-intuitive assumptions concerning representation quality, distribution adjustment, and dependence between observations. I think checking these conditions may be difficult with real data and different language models. Moreover, the improvement over simpler methods is inconsistent across experimental settings. In the controlled text experiment, some simpler methods perform similarly or slightly better. This raises questions about whether MaSI's added complexity is justified in every application. As I mentioned above, the variance issue should be addressed.

Questions for authors

I am curious: could the authors explain how practitioners can determine whether the assumptions required by MaSI are reasonably satisfied for a new dataset? Moreover, how sensitive are the results to the choice of language model and the instructions used to generate texts? I would also like to understand under which conditions the authors recommend using MaSI instead of a simpler method that requires less computation. Please explain/estimate the computational cost of generation and fitting compared with simple PPI baselines, such as PPI on PCA coordinates.

5 comments
AuthorsOpenReview ↗

Response to Reviewer T7LM [1/5]

1. How can practitioners determine whether MaSI's assumptions about representation quality, distribution calibration, and dependence between observations are reasonably satisfied on a new dataset?

These conditions can be assessed through four checks: representation quality, distribution calibration, source-group independence, and cross-covariance. Paired differences and held-out prediction assess content–expression separation and retention of scoring information; prediction-mean differences and relative ESS assess calibration; sampling and generation design determine independence across sources; relative cross terms and two incremental prediction comparisons assess the sources and effects of covariance.

(1) Representation stability and retention of scoring information. Following Appendix B, calculate the second-moment matrix of differences between original texts and style rewrites, and order its eigenvalues from smallest to largest. The first rr eigenvalues correspond to directions that change less after rewriting, while subsequent directions change more. Plot the eigenvalues and examine the gap between eigenvalues rr and r+1r+1:

γ^k=λr+1(Σ^Δ,k)−λr(Σ^Δ,k).\widehat\gamma_k = \lambda_{r+1}(\widehat\Sigma_{\Delta,k}) - \lambda_r(\widehat\Sigma_{\Delta,k}).

Resample complete source groups and re-estimate the matrix and content space. A gap that is large relative to matrix variation across resamples, together with similar estimated content spaces, supports stable separation. This corresponds to the requirement in Assumption A that the spectral gap exceed estimation error. Compare projection matrices to avoid mistaking coordinate rotations within the same space for representation changes. If the gap is unclear, increase variation in sentence structure and expression while retaining scoring cues, and check whether separation improves.

In the content coordinates, use paired differences between originals and their own rewrites to assess stability, and classification AUC to assess distinguishability of real/generated identity. In the remaining expression coordinates, use style classification to assess retention of expression differences. MHS Table F.11F.11, Panel A shows that real/rewrite classification AUC falls from 0.6430.643 in the original coordinates to about 0.4950.495–0.4980.498 in the content coordinates; style classification accuracy in the expression coordinates is 0.6030.603, above the chance level of 0.2000.200.

Using held-out real labels, compare predictors using content coordinates alone and additionally using the remaining expression coordinates. Consistent improvement by the latter suggests that scoring information remains outside the content coordinates. For example, if irony affects ratings, style rewrites should preserve its meaning. Figure 22(a) shows projection and risk changes in simulations, while Figure 22(b) and Appendix Table F.5F.5 show constraint residuals; the same optimization and constraint diagnostics can be recorded on new data.

(2) Generation coverage and prediction-mean calibration. Examine the ranges and proximity of the two sets of content coordinates, especially whether synthetic texts support extreme prediction scores among real texts, to assess the coverage relation in Assumption C. For example, if evaluation generation consistently softens strongly hostile comments, synthetic support may be missing in high-score regions, calling for instructions that preserve the original scoring cues.

A diagnostic directly linked to the final estimator is the difference between the same predictor's mean predictions on real and weighted synthetic texts. Let H^k,i\widehat H_{k,i} be the prediction for a held-out original in fold kk, and A^k,i\widehat A_{k,i} the weighted mean prediction for its synthetic group. Calculate

Δ^w,k,n=1∣Ik∣∑i∈Ik(A^k,i−H^k,i).\widehat\Delta_{w,k,n} = \frac{1}{|I_k|} \sum_{i\in I_k} \left( \widehat A_{k,i}-\widehat H_{k,i} \right).

Compare unit and estimated weights to assess whether calibration improves alignment. Following Table F.8F.8, aggregate absolute mean differences across folds and divide by MaSI's estimated standard error to obtain R^β,n\widehat R_{\beta,n}; a value of 11 means that the observed difference equals one standard error. Account for sampling uncertainty in the difference rather than treating a single held-out discrepancy as population bias.

Relative ESS near 11 indicates dispersed weighting; a substantial decrease suggests dependence on a few large weights. In MHS, standardized mean differences are about 0.740.74–0.800.80 and relative ESS is 0.9930.993–0.9990.999 (Tables F.8F.8–F.9F.9). If weighted mean differences repeatedly have the same direction and are non-negligible relative to the standard error, examine generation coverage and the weight model: improve instructions when coverage is inadequate, and improve calibration when coverage is present but alignment remains inadequate.

AuthorsOpenReview ↗

Response to Reviewer T7LM [2/5]

(3) Independent sampling and generation across sources. Write all synthetic texts from source ii as X~i=G(Xi,ξi),\widetilde{\boldsymbol X}_i=G(X_i,\xi_i), where GG is a fixed generation rule and ξi\xi_i is generation randomness. Complete source groups are independent when real sources are sampled independently and generation randomness is independent across sources and of the sample; rewrites within a source may remain dependent.

Check whether original comments share authors or threads, whether requests use only the current source and specified context, and whether random draws are shared across sources or generation rules are dynamically updated. Then assign each original and all its synthetic texts to the same fold and calculate variance by source group. Grouped splitting preserves the independence structure and cannot replace checking sampling and generation.

(4) Conditional-mean sources and relative contribution of cross-covariance. Proposition E.1E.1 in Appendix E relates covariance to

rY†(X)=E(Y∣X)−E(Y∣Z†),r_Y^\dagger(X) = \mathbb E(Y\mid X)-\mathbb E(Y\mid Z^\dagger),

rQ†(X)=E(AQ†∣X)−E(AQ†∣Z†).r_Q^\dagger(X) = \mathbb E(A_Q^\dagger\mid X) - \mathbb E(A_Q^\dagger\mid Z^\dagger).

Here, Z†Z^\dagger denotes the theoretical content coordinates and AQ†A_Q^\dagger the centered weighted synthetic-group mean prediction. The two terms measure outcome-mean information omitted by the content coordinates and changes in the synthetic-group mean associated with this additional information. When generation does not use real labels,

�0� .$$ On held-out sources, assess whether adding remaining expression information improves prediction of the real outcome $Y$ and of the synthetic-group mean, respectively. Under squared loss, the reductions in population optimal prediction error equal \$\\mathbb E

(r_Y^\dagger)^2

$and$mathbbE \$ and \$\\mathbb E

(r_Q^\dagger)^2

\$; changes in held-out error estimate these gains in practice. A small gain in either comparison supports the corresponding covariance-control path. This also corresponds to the setting in which style no longer affects the conditional outcome mean given content. For example, if irony improves both predictions, it may be a shared cue producing cross-covariance; check whether the content representation omits irony and whether style rewriting changes that scoring cue. Also estimate covariance directly from held-out real residuals and synthetic-group means, and report its relative contribution: $$\widehat\rho_{\perp,n} = \frac{2\widehat C_\perp} {\widehat\sigma_P^2+\widehat\sigma_Q^2}.$$ Conditional-mean prediction comparisons explain the source of the cross term, while $\widehat\rho_{\perp,n}$ measures its effect on the current variance. In MHS, this proportion falls from $0.113$ at $n=100$ to $0.055$ at $n=800$ (Table $F.8$). After changing the generation model or instructions, repeat these calculations, plots, and prediction comparisons on the same pilot sources to assess the sources and effects of finite-sample departures; asymptotic rates remain governed by the population conditions stated in the paper. > ### 2. How sensitive are the results to the language model and generation instructions? In paired MHS experiments, two prompts produce similar precision gains, and all three generation models retain interval shortening, with different magnitudes. The key is to preserve cues determining real ratings and make evaluation changes outcome-related and amenable to calibration. Candidate models can first rewrite the same small set of sources and be assessed against these task criteria. **(1) Semantic requirements of the two channels and perturbation direction.** The style channel changes sentence structure and expression while preserving scoring cues; for example, when tone affects friendliness, politeness and sarcasm should be preserved. The evaluation channel develops outcome cues supported by the original text while preserving the target, stance, and facts, making existing sarcasm or hostility easier to identify. The linear simulation in Appendix Figure $F.2$ fixes perturbation strength and changes only its direction. For perturbations orthogonal to the outcome direction, MaSI's width ratio and point-estimation RMSE ratio relative to Classical are $1.004$ and $0.998$; along outcome-related directions, these fall to $0.769$ and $0.761$, with empirical coverage of $92.3\\%$, close to $91.7\\%$ under orthogonal perturbations. Thus, gains from changes of equal strength depend on their outcome relevance. MHS Table $F.9$ further shows that scores based only on style rewrites align more easily but yield wider intervals than the dedicated evaluation channel at similar observed coverage.
AuthorsOpenReview ↗

Response to Reviewer T7LM [3/5]

(2) MHS sensitivity to instructions. All MHS sensitivity experiments use 200200 paired repetitions at n=100n=100, sharing real samples and fold splits. Widths are relative to Classical and nominal coverage is 9090\\%. The instruction comparison fixes DeepSeek-V3\text{DeepSeek-V3} and the Qwen3-Embedding-8B\text{Qwen3-Embedding-8B} encoder. Both evaluation prompts preserve the target, stance, facts, and quotation context. The original prompt generates nearby texts around perceived intensity and allows small intensity changes, such as slightly softening a clearly hostile comment while preserving its target and hostile meaning; for originals likely to produce annotation disagreement, it can generate a supported version near the decision boundary. The revised prompt specifies edits to the original sentence: when removing repetition, retain words determining tone and hostility; when the original supports two interpretations, retain the wording and context causing the ambiguity.

Both prompts have coverage of 90.590.5\\%, with width ratios of 0.9310.931 and 0.9220.922, respectively. One possible explanation for their similar gains is that, despite handling tone and ambiguity differently, both preserve scoring cues and restrict rewrites to what the original supports. Together with Figure F.2F.2, this suggests specifying which cues must remain and which can be made clearer. For example, when offensive language is quoted to refute it, preserve quotation attribution and the refuting stance; when the original establishes a pronoun's referent, make it explicit; when two readings remain plausible, preserve context supporting both.

(3) MHS sensitivity to the generation model. Keeping the prompt and other conditions fixed, we compared DeepSeek-V3\text{DeepSeek-V3}, Qwen3.6-35B-A3B\text{Qwen3.6-35B-A3B}, and Moonshot AI Kimi K2.6\text{Moonshot AI Kimi K2.6}. Their total parameter counts are 671B671B, 35B35B, and 1T1T; coverages are 90.590.5\\%, 90.590.5\\%, and 89.089.0\\%; and width ratios are 0.9310.931, 0.9470.947, and 0.9360.936, respectively. Width differences are mainly reflected in synthetic-data variance, showing that the generator affects the retained precision gain. Before fitting MaSI, use the same instructions on the same small set of originals for paired review of target, stance, and quotation-context retention, and whether existing scoring cues become clearer.

3. Computational complexity of MaSI and PCA-PPI.

With coordinate dimensions, generated texts per source, folds, and solver budgets fixed, the main row-wise matrix operations of both methods grow linearly with the number of real sources. MaSI's additional cost mainly comes from repeated representation-constraint evaluation and density-ratio fitting; matrix construction, eigendecomposition, and linear-system solution increase costs in high dimensions.

(1) Total complexity under fold-specific fitting. Following the paper's notation, nn is the number of real sources, KK the number of outer folds, IkI_k and IkcI_k^c the held-out and training sets in fold kk, mm the generated texts per source per channel, and dd and rr the original and content dimensions; PCA-5\text{PCA-5} PPI uses r=5r=5. Equal-sized folds have n−n/Kn-n/K training sources and n/Kn/K held-out sources. The dense-operation costs of one cross-fitting run are:

Fold-fitted PCA-5\text{PCA-5} PPI:

O((K−1)n(1+m)d2+Kd3).O\big((K-1)n(1+m)d^2+Kd^3\big).

MaSI:

�0� +G(K-1)n(1+m)r\big).$$ **(2) PCA fitting and inference after dimension reduction.** Combining real training data and evaluation data in each fold gives $(1+m)|I_k^c|$ rows and $d$ columns. Centering costs $O((1+m)|I_k^c|d)$. Each of the $d^2$ covariance entries is an inner product of two columns, so matrix construction costs $O((1+m)|I_k^c|d^2)$. Dense eigendecomposition costs $O(d^3)$, giving a total of $O((1+m)|I_k^c|d^2+d^3)$. Projecting real training rows and held-out real and evaluation rows costs $O((|I_k^c|+(1+m)|I_k|)dr)$. The outcome model uses only real training labels; constructing the ridge Gram matrix and right-hand side and solving the system costs $O(|I_k^c|r^2+r^3)$. Held-out scoring, averaging same-source synthetic scores, and calculating residuals and variance cost $O((1+m)|I_k|r)$. Each source enters $K-1$ training folds and one held-out fold. Summing across folds and using $r\leq d$ to absorb lower-dimensional terms gives the total above. Covariance construction dominates when there are many training rows; the $d^3$ eigendecomposition may dominate when the original dimension is very high. **(3) Frozen external PCA and predictor.** If PCA and a well-performing external predictor have been fitted on independent data and frozen, as in the given-predictor setting of the original PPI paper, inference requires only one projection and scoring pass over $n$ labeled and $N$ unlabeled samples, followed by prediction-mean, residual-mean, and variance calculations. Counting from the original coordinates, the total is $$O\big((n+N)dr\big).$$

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to Reviewers [1/2]

Thank you for the constructive feedback and for recognizing our treatment of dependence among texts generated from the same source. When labeled texts are scarce and no large external pool of real unlabeled texts is available, MaSI uses same-source generated texts to improve the precision of confidence intervals for the population mean. Style rewrites help learn a content representation that retains outcome cues. After calibration, separately generated evaluation texts are combined with residuals from real labels for source-group inference. Appendix G of the revision collects the additional analyses below, starting on page 73.

1. The added value of content representations beyond calibration

Content representations retain outcome information that variation in expression can obscure. PCA selects directions by overall variance. When variation in expression unrelated to the outcome dominates the data, PCA may discard directions related to the outcome and their interactions. MaSI uses same-source style rewrites to identify directions stable under changes in expression and uses real labels to guide the representation toward cues that predict the outcome. Density ratio calibration aims to align real and synthetic mean predictions in the given coordinates, so the information available for calibration depends on which cues the representation retains.

Retaining this information can reduce prediction residual variance and help shorten intervals. In the simulation with strong style variation, both methods use the same evaluation texts and follow the same rules for fitting the outcome model and density ratio and for estimating source-group variance. At n=800n=800, the average Main interval width ratios relative to Classical are 0.7500.750 for MaSI and 0.8370.837 for PCA-20. In the simulation with matching marginal distributions, the width difference persists with the known density ratio. Tables G.5–G.7 (pp. 78–79) report interaction signal retention and interval results. Table G.9 (p. 80) reports the matched MHS comparison.

Individual constraint ablations and simulations further show which properties of the representation each of the four constraints protects and where their roles overlap (Tables G.10–G.11, pp. 82–83). See our responses to kBDB's Questions 2–3 for details.

AuthorsOpenReview ↗

General Response to Reviewers [2/2]

2. Sources of cross-covariance and the role of Main

Main's additive decomposition explains where precision gains come from. When the cross-covariance is negligible and the other inference conditions hold, MaSI's intervals are shorter than Classical intervals to first order if prediction reduces residual variance by more than the variance of the synthetic-group mean prediction. Main and Block use the same point estimate and compute uncertainty at the original source-group level. Within-source dependence enters the variance of the synthetic-group mean prediction. Block also includes its cross-covariance with the residual from the real label.

This covariance can arise when the content representation misses scoring cues. For example, two comments may have similar content coordinates, with one using sarcasm and the other a neutral tone. If this difference affects both the mean real-label residual and the mean prediction for the same-source synthetic group, the two quantities can vary together across sources. To test this mechanism, we hold the marginal distribution of generated data fixed and vary their dependence on scoring cues in the original text. Finite-sample predictor error can also affect the cross term (Appendix Sections G.5.2–G.5.5).

When the cross term is zero or asymptotically negligible relative to the additive variance, the two intervals have the same first-order width. In the simulation that retains sufficient scoring information and uses the population conditional mean as the predictor, the cross term is zero, the intervals have similar widths, and Main has lower variance-estimation RMSE (Table G.13, pp. 87–88).

In practice, held-out source groups can be used to estimate the cross term relative to the additive variance. Uncertainty in this estimate helps assess how much the cross term affects the standard error (Appendix G.5.7, p. 92). Across the five MHS sample sizes, Block intervals with estimated weights have an average empirical coverage of 90.490.4\\% and are 3.53.5\\% shorter on average than Classical intervals (Table F.20, p. 69).

In the learned content coordinates, if the evaluation-text distribution covers the real-text distribution and the true density ratio is bounded, weighting by the true ratio aligns the two population mean predictions. With estimated weights, the theorem further requires L2L_2 consistency of the weights and a population calibration error in the prediction mean of op(n−1/2)o_p(n^{-1/2}) (Appendix G.6, pp. 93–94). In the simulation with matching real and generated marginal distributions, the Block intervals have similar widths with estimated and known weights (Table G.8, p. 79). See our responses to kBDB's Question 4 and HcmK's Questions 1, 2, and 4 for details.

3. Generation design, diagnostics, and method choice

Generation should focus on outcome-relevant cues. The style channel changes expression while preserving scoring cues to help learn a stable representation. The evaluation channel expands on task-relevant semantic relationships within what the original text supports, providing useful cues for prediction (Appendix G.1, p. 73). In MHS, holding the synthetic text pool fixed and shuffling the links to original texts worsens held-out prediction (Table F.11, Panel B, p. 62). Both tested prompts and all three tested generators retain a width advantage over Classical (Tables G.1–G.2, p. 75).

On new data, practitioners can assess generation and calibration by checking whether scoring cues are retained, evaluating held-out prediction, and comparing real and weighted synthetic mean predictions (Appendix G.7, p. 94 onward). When a fixed representation already retains the main outcome signal and calibration is stable, a simple calibrated baseline can be used. When variation in expression may obscure the outcome signal, practitioners can compare held-out prediction and variance gains under a matched workflow before deciding whether to use MaSI (Appendix G.9, p. 98 onward). In paired MHS timings starting with text embeddings in memory, MaSI takes a median of about 11–33 seconds longer than PCA-20 PPI from fitting through interval construction (Table G.17, p. 98).

See our responses to T7LM's Questions 1–2 for diagnostics and sensitivity and Questions 3–5 for cost and method choice. Generation examples and their use in prediction are discussed in our responses to kBDB's Question 1 and HcmK's Question 3. We welcome further discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.