acceptodds
Under review as a conference paper at ICLR 2027

Accuracy-Optimal Stopping Can Reverse Treatment Effect Estimates from LLM Labels

Abstract

Repeated language model calls can improve the accuracy of outcome labels in randomized studies. We show that stopping optimized for pooled accuracy can nevertheless reverse the reported treatment effect while accuracy rises and the population and response kernel remain fixed. When stopping may depend on treatment, this occurs even with conditionally independent calls, answers better than chance on every record, and oracle posterior decoding. In a construction with at most three calls, we characterize the attainable region of accuracy and effect, identify an open budget interval where every accuracy maximizer reports the wrong sign, and derive the exact accuracy cost of preserving the effect. For any prescribed number of sign reversals, we construct one population and a finite call cap under which price optimal policies produce those reversals as call prices decrease and accuracy strictly increases. A first revision representation characterizes all attainable effect displacements within a call horizon, without requiring independent calls or Bayesian decoding. In the four type model, a known response catalog permits unbiased correction using at most two additional calls. Semisynthetic experiments using measured response catalogs and recorded model answers show that stopping optimized separately by arm produces substantial contrast displacement and severe confidence interval undercoverage on held out records. Ambiguous polarity pools can exhibit reversals, and policies with nearly identical accuracy can yield opposite effect signs. In the held out comparisons, fitted stopping rules shared by both arms produce much smaller mean absolute displacement with similar mean accuracy.

Primary area:
causal reasoning
Keywords:
optimal stopping, LLM annotation, measurement error, treatment effects, proxy outcomes, partial identification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_5iykOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates how optimizing repeated LLM queries for classification accuracy can distort estimated treatment effects, potentially reversing their signs. It develops theoretical counterexamples, characterizes the effects of adaptive stopping, and proposes correction procedures under known response distributions. Semi-synthetic experiments using measured LLM responses illustrate the phenomenon.

Critical strengths

The paper raises a relevant concern about using LLM-generated labels for downstream causal inference. The exact stopping-rule constructions and characterization of attainable effect distortions are technically interesting.

Critical weaknesses
  1. My first main concern is limited conceptual novelty. The central observation that higher classification accuracy does not guarantee less bias in prevalence or treatment effect estimates is already well established. The elaborate stopping-rule constructions do not clearly establish a broadly new methodological contribution.

  2. Also, the main results rely on carefully constructed response distributions, oracle posterior decoding, and treatment-dependent stopping. The repeated sign reversals require very small treatment effects and narrow query-price intervals, limiting their practical relevance.

  3. Theory and experiments seem somewhat disconnected. The theoretical framework assumes oracle posterior decoding, whereas the empirical implementation uses a pooled decoder with arm-specific stopping. This weakens the connection between the theoretical findings and empirical evidence.

  4. My second main concern lies in limited relevance to observational causal inference. Although motivated by downstream causal inference, the analysis focuses primarily on randomized treatment contrasts or synthetically constructed treatment arms. It remains unclear how the findings translate into observational studies requiring confounding adjustment, or how adaptive LLM labeling affects commonly used causal estimators. A theoretical analysis or realistic observational-data application would help establish the broader implications for causal inference.

  5. My third concern is questionable practical contribution. A common stopping rule largely eliminates the observed reversals with negligible accuracy loss. The proposed correction requires known response distributions and thus can substantially increase variance or become unreliable under misspecification. The practical advantage over simpler alternatives remains unclear. Further, sign reversals occur mainly under specially calibrated treatment distributions and selected ambiguous records, but not under natural designs or among records satisfying the main theoretical better-than-chance assumption.

Questions for authors

see my comments in the weakness section

Reviewer_nH8pOpenReview ↗
Rating
3
Confidence
2(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how adaptive stopping of repeated LLM annotations can distort treatment effect estimation. The authors show theoretically that accuracy-optimal stopping can reverse the sign of the estimated treatment effect even as classification accuracy improves. They further provide theoretical characterizations, correction methods, and semi-synthetic experiments illustrating this phenomenon.

Critical strengths
  • Interesting motivation. The paper highlights an important distinction between classification accuracy and the validity of downstream causal inference.

  • Substantial theoretical analysis. The authors provide detailed theoretical results characterizing treatment effect reversals and their relationship with stopping policies.

Critical weaknesses
  • Limited practical relevance. The sign-reversal results rely heavily on treatment-aware stopping and carefully constructed settings. In the empirical experiments, reversals mainly occur in specially constructed ambiguous subsets, rather than under more natural stopping procedures.

  • Limited applicability of the proposed correction. The correction requires a known response catalog and can substantially increase estimation variance. Its robustness and feasibility in real applications remain unclear.

Questions for authors
  • Could the authors further demonstrate whether sign reversals occur under realistic stopping policies that do not explicitly depend on treatment assignment?

  • How practical is the proposed correction when the response catalog must be estimated rather than assumed known? Could the authors provide additional empirical evidence?

Reviewer_AhkWOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies randomized experiments on binary outcomes that are measured through repeated LLM calls. If the number of LLM calls is optimized using a stopping rule for pooled classification accuracy, resulting treatment effects can be biased or even reversed. Specifically, the paper characterizes an accuracy– treatment effect frontier, establishes repeated sign reversals as computation becomes cheaper, and presents calibration results. Semisynthetic experiments underline the finding that stopping optimized separately by arm produces contrast displacement and too narrow confidence intervals.

Critical strengths
  • The paper outlines an important observation in that accurate individual LLM labels do not necessarily yield valid aggregate inference.
  • The paper showcases the problem on a three-call construction showing how accuracy improvements can disrupt cancellation of measurement errors across treatment arms.
Critical weaknesses
  • Motivation and practical setting are insufficiently developed. The paper does not clearly establish who chooses the stopping rule, why treatment assignment would be available to that procedure, or how accuracy optimization is feasible when the true outcome is unobserved. This is particularly important because common stopping rules achieve similar accuracy with much smaller displacement in the reported comparisons (Table 1).
  • The abstract is difficult to follow. It introduces numerous technical results and concepts before clearly explaining the practical problem, the mechanism behind the distortion, and the main takeaway.
  • The notation is insufficiently introduced. For example, Eq. (1) uses RtR_t, rr, and Qr(Z)Q_r(Z) before explicitly defining them.
  • The overall presentation is extremely hard to follow and lacks a clear line of thought.
  • The connection to treatment-effect estimation needs sharper explanation. Eq. (2) defines a causal target, but the mechanism is distortion from adaptively measuring the outcome, but not a failure of identification under randomization. The paper should clarify what is specific to causal estimation beyond estimating a contrast between two populations.
  • The stated sign-reversal result not sufficiently supported through empirical studies. In the experiments, sign reversals occur only under anchoring and on pools containing records that violate the stated recordwise better-than-chance assumption.
  • The practical impact and robustness of the findings is unclear. Full-corpus displacement is relatively small, and near-optimal policies on ambiguous polarity pools can produce either sign. For real-world applications, the overall message/finding of the paper thus remains unclear.
Questions for authors
  • What annotation workflow motivates treatment-aware accuracy optimization? Who observes treatment assignment, and what labeled or calibration data are available to fit the stopping rule? How would one obtain the response catalog in practice?
  • Which findings are specific to (causal) treatment-effect estimation, rather than to estimating a difference between two populations?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.