acceptodds
Under review as a conference paper at ICLR 2027

Which Model Should Be Chosen? Joint Performance Estimation and Ranking of Black-Box Models on Unlabeled Data

Abstract

Evaluating a pool of models on an unseen and unlabeled dataset requires both accurate performance estimates and reliable rankings. Many existing evaluation methods estimate performance for one model at a time, often requiring model-specific training or adaptation, access to training data, or internal representations to characterize distribution shifts. These requirements limit evaluation of black-box models. Moreover, methods that evaluate models independently miss cross-model evidence that could improve ranking within a model pool. We propose PoolEvaluator, a framework for jointly estimating and ranking model performance without target labels or access to model internals. It estimates prior accuracies from relevant calibration datasets, then combines these priors with model agreement through expectation-maximization to infer latent correctness and refine each model's accuracy estimate using a closed-form update. An external budgeted judge is used later to refine ambiguous latent answers. Experiments across Text2SQL, image classification, and node classification show improved accuracy estimation and ranking with lower end-to-end evaluation latency than existing methods. This opens a new direction for jointly evaluating and ranking black-box model pools on unlabeled data through shared prediction evidence. Our code is available at https://anonymous.4open.science/r/PoolEvaluator-664E.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
model evaluation, pool ranking, unlabeled data

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kGXqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PoolEvaluator, a framework for estimating and ranking the accuracy of multiple black-box models on an unlabeled target dataset. It uses cross-model agreement, optionally supplemented by labeled calibration datasets and external judge verdicts, to reduce reliance on target ground-truth labels. An EM procedure estimates model accuracies and agreement rates, while selective judging refines the estimates.

Critical strengths
  1. Practical problem. Estimating model performance and selecting among candidate models when target labels are scarce or expensive is an important evaluation problem.
  2. Potentially useful evaluation signal. Cross-model agreement provides information beyond evaluating each model independently. Combining this signal with calibration data and selective judging is a promising approach.
  3. Broad empirical coverage. The experiments span Text2SQL, image classification, and node classification across multiple datasets, measuring both accuracy estimation and ranking quality.
Critical weaknesses
  1. Identifiability of accuracy from agreement.
    My primary concern is whether the stated model can identify accuracy from unlabeled agreement observations. Marginalizing Equation 1 gives
P(Cij=1)=αjβj+(1−αj)γj.P(C_i^j=1)=\alpha_j\beta_j+(1-\alpha_j)\gamma_j.

This constrains one probability while three parameters are unknown. For example, both (α,β,γ)=(0.8,0.8,0.3)(\alpha,\beta,\gamma)=(0.8,0.8,0.3) and (0.5,0.9,0.5)(0.5,0.9,0.5) imply an agreement probability of 0.7 despite different accuracies. Both also satisfy β>γ\beta>\gamma.

Under the factorized likelihood in Equation 2, I do not see an additional constraint that resolves this ambiguity. Calibration-based initialization may influence which solution EM reaches, but initialization alone does not establish identifiability. This is particularly relevant when a strong model disagrees with a weaker majority or several models share systematic errors.

  1. The contribution of unlabeled agreement needs clearer separation from labeled information.
    The default configuration uses labeled calibration subsets and judge-provided correctness assignments. The paper also permits random initialization and claims competitive performance with both budgets set to zero (K=0,V=0K=0,V=0, Section 6).

    Figure 7 varies the budgets separately, but explicit results for the joint zero-budget setting would help assess this claim. A comparison of calibration-only estimates, calibration plus EM, and calibration plus EM plus judging would also clarify how much the unlabeled agreement stage contributes. The distinction between evaluation on an unlabeled target and evaluation without any labeled information should be explicit.

  2. Judge verdicts are treated as exact correctness information.
    Equations 3 and 9 assign hard correctness labels from the external judge, while Appendix A assumes exact validation. In practice, judges may make mistakes or share errors with the evaluated models. Using a different model family does not by itself establish independent errors.

    Reporting judge accuracy on the selected examples and assessing sensitivity to erroneous verdicts would strengthen the evaluation. Practical guidance on validating judge quality and handling disagreement between judges would also improve the deployment case.

Questions for authors
  1. What assumptions or constraints identify αj\alpha_j separately from βj\beta_j and γj\gamma_j? How does the method distinguish parameter settings with identical agreement probabilities but different accuracies?
  2. Can you provide explicit results for K=0,V=0K=0,V=0 and isolate the improvement from EM over calibration-only estimates?
  3. How accurate are the judges on the queried examples, and how sensitive are accuracy estimates and rankings to judge errors? How should practitioners evaluate this risk on a new dataset?
Reviewer_aRhPOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper deals with performance estimation and ranking of black-box models on unlabeled data. Previous works estimated performance for a model at a time, or required specific model-training or access to training data. The paper proposes PoolEvaluator, that is, a framework for estimating and ranking models performance without target labels. Experiments show improved accuracy estimation and ranking while simultaneously lowering end-to-end evaluation latency.

Critical strengths
  • The paper is well motivated and deals with a real problem
  • The paper outperforms evaluated baselines in all settings, additionally, PoolEvaluator is more efficient compared to the baselines
Critical weaknesses
  • Unclear novelty and comparison with prior work: Model Selector predates this work and appears to address a similar setting, including noisy labels. The paper should explain what is new about its formulation and clarify its claim of being the first.

  • Unclear reliability of model selection: It is unclear how the method can reliably identify the best models without labels, especially when all models in the pool are noisy. The paper should also report the performance gaps between candidate models, particularly the top models. Without this information, it is difficult to judge whether the method can distinguish similarly performing models or whether the evaluated cases are relatively easy.

Questions for authors

“Formulation: To our knowledge, we are the first to formulate label-free model-pool evaluation where latent labels are estimated from the outputs of multiple candidate models.” — The Model Selector paper (https://arxiv.org/abs/2410.13609) predates this work and already addresses noisy labels and these aspects. Could the authors clarify whether this understanding is incorrect or reformulate their claim? Is the method limited to classification? Does it rely on accuracy or another metric? Can it work on general language generation tasks, such as code generation?

How do you know that you are selecting the best models without any labels or oracles? I assume labels are used to verify the selections against the ground truth. Additionally, what happens if all models in the pool are noisy? Why assume that no labels are available at all? Usually, the setting involves a limited labeling budget, which is reasonable. Can your label-free method outperform methods that use a small labeling budget? For example, how does it compare to Model Selector with a small budget (https://arxiv.org/abs/2410.13609)?

How does your method’s wall-clock time scale with the pool size?

I am interested in the distribution of model performance on the datasets you evaluate. For example, what are the performance differences among the top three models? What is the performance gap between the best and worst models? A histogram of model performance would be helpful.

1 comment

Thank you for addressing most of my concerns. However, I have a follow-up question regarding the following:

Q2b. Can PoolEvaluator outperform methods that use a small labeling budget, such as Model Selector?

You have compared PoolEvaluator against Model Selector on Text2SQL, whereas Model Selector is designed for classification tasks. How did you adapt Model Selector to these non-classification tasks? In my opinion, it would be more appropriate to evaluate on tasks for which the method was originally designed, such as image classification.

Reviewer_KAq9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PoolEvaluator, a framework for jointly estimating the accuracy of multiple black-box models and ranking them on an unlabeled target dataset. The method consists of three main components: (1) prior initialization using labeled calibration subsets selected according to their representation similarity to the target data, (2) joint estimation via an EM algorithm that models both model accuracy and prediction-consensus behavior, and (3) an optional judge-guided refinement procedure that selects informative samples for external LLM evaluation. Experiments verify the advantages of the proposed method.

Critical strengths
  1. The paper addresses model evaluation and selection when target labels are unavailable or expensive to obtain. This is relevant to changing data distributions and black-box model pools.
  2. The method performs well on Text2SQL, image classification, and node classification. It achieves lower accuracy-estimation MAE than MetaEvaluator.
Critical weaknesses
  1. The model estimates accuracy from unlabeled agreement, but the observable agreement probability depends jointly on (\alpha_j), (\beta_j), and (\gamma_j). Therefore, different parameter combinations may produce the same agreement rate, and EM convergence alone does not show that the recovered parameters correspond to true accuracies.
  2. The parameter (\gamma_j) captures incorrect agreement at the individual-model level but does not account for dependencies among models or shared systematic errors. Leave-one-out consensus does not remove correlations among the remaining models, so groups of models may reinforce the same wrong prediction.
  3. Selecting calibration data by representation similarity does not necessarily imply similar accuracy or error patterns, especially across datasets with different structures or difficulty levels. The paper should compare similarity-based selection with random and alternative selection strategies, and evaluate deliberately mismatched calibration sets. The effect of prior quality on estimation error and ranking stability should also be reported.
  4. The information-gain strategy depends on the quality of the current posterior estimates and may be unreliable when the model is misspecified or errors are correlated. In addition, an LLM judge may introduce bias related to the prompt, judge model, or task difficulty.
Questions for authors
  1. Under what assumptions are (\alpha_j), (\beta_j), and (\gamma_j) identifiable from unlabeled consensus data? Please characterize the non-identifiable cases.
  2. When (K=0,V=0), why should EM recover the correct accuracies rather than another parameter configuration with a similar likelihood? How sensitive are the estimates and rankings to initialization?
  3. How does the method perform when many models share systematic errors? Please include controlled experiments with different levels of model correlation.
  4. Under a fixed annotation budget, does information-gain selection outperform random selection when the model is misspecified or the judge is imperfect?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.