ProEval: Proactive Failure Discovery and Efficient Performance Estimation for Generative AI Evaluation

Yizheng Huang; Wenjun Zeng; Aditi Kumaresan; Zi Wang

Source-linked AI summary

ProEval: Proactive Failure Discovery and Efficient Performance Estimation for Generative AI Evaluation

Yizheng Huang, Wenjun Zeng, Aditi Kumaresan, Zi Wang

arXiv:2604.23099v2Apr 25, 2026cs.LGcs.AIstat.ML

TL;DR

Generative AI evaluation is costly and can miss rare failures when test data are downsampled. ProEval uses transfer-learned Gaussian-process surrogates with Bayesian quadrature and active sampling, reducing evaluation samples 8-65x and discovering 2-5x more failure cases than competitive baselines.

Problem
Generative AI evaluation is increasingly costly, while downsampling can reduce estimate accuracy and miss rare but critical failures.
Method
ProEval combines transfer learning, Gaussian-process surrogates, Bayesian quadrature, and active input selection to estimate performance and discover failures.
Results
8-65x reduction in evaluation sample sizes and 2-5x more discovered failure cases than competitive baselines were achieved across performance estimation and failure discovery.
Takeaways & Limitations
ProEval improves sample efficiency and failure-case discovery for expensive-to-query and expensive-to-rate generative AI models.
Takeaways & Limitations
ProEval depends on strong Gaussian-process priors and may abstain from predictions when reliable prior data are unavailable.

Abstract

from arXiv · show

Evaluating generative AI models is increasingly resource-intensive due to slow inference, expensive raters, and a rapidly growing landscape of models and benchmarks. We propose ProEval, a proactive evaluation framework that leverages transfer learning to efficiently estimate performance and identify failure cases. ProEval employs pre-trained Gaussian Processes (GPs) as surrogates for the performance score function, mapping model inputs to metrics such as the severity of errors or safety violations. By framing performance estimation as Bayesian quadrature (BQ) and failure discovery as superlevel set sampling, we develop uncertainty-aware decision strategies that actively select or synthesize highly informative inputs for testing. Theoretically, we prove that our pre-trained GP-based BQ estimator is unbiased and bounded. Empirically, extensive experiments on reasoning, safety alignment, and classification benchmarks demonstrate that ProEval is significantly more efficient than competitive baselines. It requires 8-65x fewer samples to achieve estimates within 1% of the ground truth, while simultaneously revealing more diverse failure cases under a stricter evaluation budget.

1. Introduction

ProEval addresses the resource-intensive nature of generative AI evaluation with a transfer-learning and Bayesian framework that jointly estimates performance and discovers failures. It combines Bayesian quadrature, superlevel set sampling, active synthesis, and uncertainty-aware acquisition to improve evaluation efficiency and failure diversity.

Motivation: Generative AI evaluation is increasingly unsustainable because inference, human or LLM judging, and large benchmarks make comprehensive assessment costly.Repeated evaluation during model development also encourages downsampling, which can reduce estimate accuracy and miss rare but critical failures.
Framework: ProEval uses pre-trained Gaussian Processes and transfer learning to model performance scores as an unknown function mapping inputs to metrics such as error severity or unsafe-response severity.Its priors come from historical within-benchmark evaluation statistics or pre-trained semantic embeddings for new domains and modalities.
Method: ProEval formulates performance estimation as Bayesian quadrature and failure discovery as superlevel set sampling within one active probabilistic framework.Variance-minimizing acquisition supports estimation, while exploitation–exploration sampling and bi-level LLM-based synthesis target severe and diverse failures.
Results: 1 to 27 evaluated inputs can bring ProEval within 1% estimation error, while failure detection rates are about 2-5x higher than competitive baselines.Theoretical analysis proves the pre-trained-GP Bayesian quadrature estimator unbiased and bounded under mild assumptions.

2. Our Framework: Proactive Evaluation

ProEval frames performance estimation and failure discovery as dual Bayesian objectives, using transfer-learned Gaussian Process priors to actively sample informative inputs. The framework estimates global expected scores and identifies diverse high-severity failures, with theoretical guarantees for its performance estimator.

Active Sampling: ProEval actively samples for Bayesian quadrature and probabilistic superlevel-set discovery, progressing from static-dataset retrieval toward diversity-driven failure synthesis.These strategies target diverse, high-severity inputs within the failure region.
Dual Evaluation Objectives: ProEval jointly estimates the global expected score and identifies the superlevel set of inputs whose failure scores exceed threshold λ.It minimizes evaluations of the performance function while targeting failure regions containing safety violations or reasoning errors.
Transfer-Learned GP Priors: Transfer learning constructs GP priors from historical evaluation data, capturing correlations across inputs so observed examples inform performance on unobserved inputs.The shared-prior assumption models score functions across datasets with unknown mean, kernel, and noise variance.
Prior Construction: For shared benchmarks, ProEval extracts score features from historical model results; otherwise, it uses semantic embeddings and learned kernels to transfer across benchmarks or missing inputs.Embedding-based priors enable zero-shot predictions for previously unevaluated inputs based on semantic similarity, while experiments use a Matérn kernel over embeddings.
Theoretical Guarantees: Theorem 3 establishes that the learned-prior performance estimator is unbiased and provides a worst-case deviation bound from the ground-truth estimate.The bound improves with more historical models and is non-vacuous when their number grows faster than the number of test samples.

3. Empirical Analyses and Results

ProEval is evaluated across default, new-model, and new-benchmark scenarios using benchmarks, models, metrics, and diversity measures. Experiments show strong performance estimation and failure discovery, while exposing limitations from transfer quality, generator capability, and modeling choices.

Experimental setup: ProEval is assessed in Default, New Model, and New Bench scenarios using historical model–benchmark data, novel models, and novel benchmarks.The evaluation introduces benchmarks, models, and specific metrics, with additional details provided in §C.
Performance estimation: At a 1% benchmark-size budget, ProEval variants consistently outperform all baselines in MAE under the reported settings.The comparison includes transfer-learning BQ variants, BQ Rounded, random selection, and active selection; BQ Rounded can be less stable because rounding discards posterior uncertainty.
Performance estimation: BQ-RPF outperforms Random Sampling on 6 out of 9 datasets in the New Bench scenario, despite lower performance than in the Default setting.It falls behind only on DICES, StrategyQA, and SVAMP, indicating generalization in zero-shot prompt-feature transfer.
Transfer analysis: Cross-modality transfer reduces MAE from 0.111 to 0.055 when transferring from DICES text data to DIVE image data with 15 samples.The experiment compares knowledge transfer against no knowledge transfer.
Failure discovery: TSS-RPF achieves overall diversity scores of 0.74 on StrategyQA and 0.73 on GSM8K, matching or nearing Rand-Anchor-Gen while outperforming other structured alternatives.The result indicates that generated failures remain diverse rather than collapsing onto a single mode.
Failure discovery: TSS-RPF discovers 41.5% failures on StrategyQA with GPT 5, compared to 28.7% with Qwen3 (32b) and 26.3% with Gemma3 (27b).The results show that stronger LLM generators consistently improve failure discovery rates, while generator quality is necessary for SS-Gen and TSS effectiveness.

4. Discussion and Conclusion

ProEval combines Bayesian ideas and transfer learning to improve sample efficiency for performance estimation and failure discovery in expensive generative AI evaluation. Its results support strong GP priors and learned embeddings while identifying embedding quality as an important factor and outlining future improvements.

Contributions and results: ProEval uses Bayesian ideas and transfer learning to improve the sample efficiency and effectiveness of performance estimation and failure case discovery.The framework targets modern generative AI models that are expensive to query and rate.
Discussion: Strong GP priors are a core advantage because they avoid learning surrogate models from scratch through expensive direct observations.ProEval also evaluates prior quality by verifying the sufficiency of source data.
Discussion: Stronger embedding models tend to have more accurate performance estimation, making prompt-feature embeddings an important factor in ProEval’s quality.This conclusion comes from an ablation study on BQ-RPF using different embedding models.
Discussion and future work: Learned embeddings can transfer underlying performance patterns to entirely new datasets, while incorporating CoT reasoning into embeddings can reduce estimation error.Future work includes reducing reliance on high-quality embeddings and developing better acquisition functions, surrogate models, and variance reduction techniques.

Impact Statement · A. Literature Review

ProEval is presented as an efficient framework for evaluating resource-intensive generative AI models and systematically discovering failures. The literature review distinguishes it from benchmark pruning, active learning, agentic evaluation, red teaming, white-box methods, and prior Bayesian quadrature by its black-box, global failure-mapping and metric-estimation focus.

Impact Statement: ProEval reduces evaluation overhead, accelerates GenAI research and quality iteration, and supports deeper understanding of model failures for more trustworthy AI.The impact statement anticipates enhanced transparency and other positive societal outcomes.
A. Literature Review: Efficient-evaluation research centers on static benchmark pruning and standard active testing, whereas ProEval differs in its assumptions and statistical implications.Static methods seek fixed representative subsets, while active testing dynamically selects queries during evaluation.
A. Literature Review: TinyBenchmarks, MetaBench, HELM allocation strategies, Anchor Points, and DISCO target representative or disagreement-based subsets but remain target-agnostic about failure modes.These methods support benchmark standardization and leaderboard construction rather than explicitly mapping model-specific failures.
A. Literature Review: Standard Active Learning selects samples to train or fine-tune models, while ProEval selects inputs to estimate test metrics and discover failure regions for a fixed model.Its acquisition functions prioritize integral-estimator variance reduction and/or superlevel-set locations rather than information gain for model parameters.
A. Literature Review: Simulated-user and interactive-agent evaluations capture capabilities beyond static QA but impose computational costs that make them unsuitable for iterative development.Related approaches also use diverse personas to model real-world usage variance.
A. Literature Review: Adversarial attacks and automated red teaming identify blind spots, but prompt optimization can lack diversity, while Rainbow Teaming is sample-inefficient and Bayesian Red Teaming remains heuristic optimization.The reviewed methods include attacker fine-tuning, PAIR, TAP, evolutionary Quality-Diversity, concept edits, and Gaussian-process-based optimization.
A. Literature Review: White-box methods detect or correct failures through residual streams, intermediate features, or activation shifts, unlike ProEval’s input-space black-box evaluation.Examples include residual-stream error prediction, confidence networks, and truth-direction activation changes.
A. Literature Review: ProEval treats evaluation as black-box Bayesian Level Set Estimation and Bayesian Quadrature, mapping global failure superlevel sets while estimating accuracy or safety over high-dimensional semantic inputs.Prior Bayesian quadrature work focused mainly on scientific-computing tasks, whereas ProEval adapts it to generative AI evaluation.

B. Proof for Theorem 3

The proof of Theorem 3 uses the sample mean’s unbiasedness and posterior-variance bound to establish guarantees for the integral estimator. It shows that the estimator is unbiased for the ground-truth expectation and bounded with probability 1 − δ.

Proof strategy: The proof relies on the sample mean ˆ𝜇𝑡(𝑥) being unbiased and bounded by the posterior variance.This is the stated basis of the proof of Theorem 3.
Estimator construction: The integral estimator is constructed from the estimated GP mean and kernel.The supplied passage introduces this estimator as the central object of the proof.
Estimator guarantee: The estimator ˆ𝑆𝑡 is unbiased for the ground-truth expectation of the integral.The proof explicitly identifies ˆ𝑆𝑡 as an unbiased estimator.
Estimator guarantee: With probability 1 − δ, the estimate is bounded.The supplied proof states a high-probability bound, with κ introduced as a constant bounding the kernel value.

C. Details on Experiments and Additional Results

This section describes the experimental setup and presents additional results, including a GMM-based source-data selection strategy. The strategy uses all benchmarks except the target benchmark as the reference benchmark.

The section describes experiment details and presents additional results.
The experiments use a GMM-based approach for source-data selection.
The selection strategy uses all benchmarks except the target benchmark as the reference benchmark.

C.1. Datasets · C.1.1. Baselines for performance estimation

ProEval is evaluated for performance estimation and failure-case discovery across mathematical reasoning, world knowledge, and safety-alignment benchmarks. Its baselines combine random sampling with active-testing surrogates and estimators, using subsampled datasets to reduce costs.

C.1. Datasets: ProEval evaluates performance estimation and failure-case discovery across mathematical reasoning, general world knowledge, and safety alignment.These domains span objective numeric correctness to nuanced human judgment.
C.1. Datasets: Most datasets use subsets of the original data to accelerate experimentation and minimize inference costs; the MMLU subset is Professional Law.Table 6 compares dataset sizes and modalities.
C.1. Datasets: GSM8K and SVAMP assess multi-step chain-of-thought reasoning, with SVAMP additionally testing robustness to linguistic variation.GSM8K contains grade-school math problems requiring numeric solutions.
C.1. Datasets: StrategyQA evaluates implicit reasoning by requiring decomposition of Yes/No questions into latent reasoning steps, while GQA tests multimodal reasoning ability.The passage gives “Did Aristotle use a laptop?” as an example StrategyQA question.
C.1. Datasets: MMLU covers 57 subjects across STEM, humanities, and social sciences, with abstract algebra and professional law used as multiple-choice subsets.These subsets evaluate general-purpose capabilities across varying difficulty levels.
C.1. Datasets: ToxicChat, Google Civil Comments, Google DICES-350, and text-to-image DIVE provide toxicity annotations, real-world interactions, online comments, or expert safety ratings.DIVE contains expert safety ratings across 123 unique raters.
C.1.1. Baselines for performance estimation: The baselines include Random Sampling and Active Testing with Logistic Regression or Random Forest surrogates trained on evaluation results from 11 models.Active Testing guides acquisition while correcting for selection bias.
C.1.1. Baselines for performance estimation: Combining the two surrogates with importance sampling or LURE produces LR+IS, LR+LURE, RF+IS, and RF+LURE.These four combinations are the active variants used for comparison.

C.2. Ablation Studies with Pre-training Data Selection · C.3. Additional Results on Performance Estimation

ProEval’s transfer-learning premise is supported by shared failure patterns across models, while selective pre-training data substantially improves Bayesian quadrature performance. Ablations show that poor pre-training pairs can severely degrade estimation quality and that the best pair depends on the target model.

C.2. Ablation Studies with Pre-training Data Selection: Shared failure structure validates transferring information from historical model evaluations to pre-train Gaussian-process performance surrogates.The framework relies on related model behaviors rather than idiosyncratic failure modes.
C.2. Ablation Studies with Pre-training Data Selection: Positive correlations across all model pairs on StrategyQA and SVAMP indicate that difficult questions tend to challenge models similarly.This supports ProEval’s assumption that model failure modes are rarely orthogonal.
C.2. Ablation Studies with Pre-training Data Selection: Historical datasets are selected by identifying models whose performance profiles form suitable clusters for the target evaluation.For Gemini 2.5 Flash on StrategyQA or GSM8K, SVAMP serves as the hold-out benchmark for selecting pre-training data.
C.2. Ablation Studies with Pre-training Data Selection: Gemma-3-12B and GPT-3.5 Turbo fall outside the relevant cluster and are consequently removed from the candidate pre-training set.This illustrates the model-selection procedure using SVAMP to guide evaluation of Gemini 2.5 Flash.
C.2. Ablation Studies with Pre-training Data Selection: 100× higher MAE can result from poor pre-training model pairs on GSM8K, demonstrating that blindly selecting data can substantially degrade BQ performance.Figure 11 compares all pairwise pre-training combinations for two target models, with lower MAE indicating better prior quality.
C.2. Ablation Studies with Pre-training Data Selection: The optimal pre-training pair varies by target model, so effective Gaussian-process initialization requires target-specific data selection.The ablation evaluates pairwise combinations on GSM8K and shows that pre-training selection matters significantly.

C.3.1. Results on Different Target Models

Across frontier and legacy target models, ProEval’s Active Selection + BQ methods generally achieve the lowest MAE and outperform random sampling and LURE. This indicates that active BQ’s variance-reduction benefits extend across model architectures.

Cross-model results: Active Selection + BQ consistently achieves the lowest MAE on most benchmarks across the evaluated target models.The evaluation uses the same experimental setting as Table 1 and reports results in Tables 7–21.
Baseline comparison: Active Selection + BQ outperforms both random sampling and the strong importance-sampling baseline LURE.The comparison is reported across the comprehensive set of frontier and legacy generative AI models.
Architectural generalization: The variance-reduction property of active BQ acquisition generalizes well across different model architectures.Results cover GPT, Claude, Gemini, Gemma, and Qwen target models, with image-input metrics unavailable for GPT-3.5 Turbo and Qwen 3 32B.

C.3.2. Ablation Study on Different Embedding Models

A preliminary ablation evaluated four existing embedding models in ProEval on Gemini 2.5 Flash with StrategyQA using BQ-RPF and a sampling budget of 50. Models with larger embedding dimensions tended to produce lower MAEs.

Experimental setup: The study compared Google gemini_embedding_001, OpenAI text_embedding_3_small, OpenAI text_embedding_3_large, and all_minilm_l6_v2 using a sampling budget of 50.Evaluation used Gemini 2.5 Flash on StrategyQA with BQ-RPF.
Embedding-dimension effect: Larger embedding dimensions tended to yield lower MAEs in ProEval.This trend was observed in the preliminary experiment across the four embedding models.

C.4. Additional Results on Failure Case Discovery … E.2.7. Mardia’s Multivariate Normality Test

ProEval’s active generation discovers substantially harder failure cases than random generation, while its source-selection methods identify or reject historical models based on behavioral similarity and normality assumptions. Among these methods, GMM clustering improves MAE with abstention, whereas Hotelling’s T² and Mardia’s test face practical limitations.

C.4. Additional Results on Failure Case Discovery: Active Generation (TSS) consistently discovers complex failure cases, whereas Random Generation often produces simpler questions models easily solve.The reported examples contrast Score 1.0 for active discoveries with Score 0.0 for random questions.
C.4. Additional Results on Failure Case Discovery: Active sampling synthesizes multi-hop StrategyQA-like constraints and harder GSM8K-like problems requiring multi-step reasoning, systems of equations, or layered constraints.These examples are presented in Tables 23 and 24.
D. Generation Prompts for Failure Discovery: Generation prompts vary by topic constraints, anchor examples from previously identified hard problems, and tuned prompt features, with TSS combining topic and anchor information.The variable descriptions define anchor_context as incorrectly answered examples selected by Superlevel Set acquisition and selected_topic_label as a BERTopic-derived topic.
E. Source Data Selection Methods for Bayesian Quadrature: Source selection estimates the GP prior from historical model scores while choosing a subset that balances data quantity against similarity to the target model.Using all models can violate the shared-prior assumption for OOD targets, while irrelevant models can corrupt BQ predictions.
E.2.1. Leave-One-Out Prior (Baseline): Leave-One-Out Prior uses every historical model except the target, preventing leakage but potentially including outliers that corrupt the prior.It is simplest when models behave similarly and can be combined with Spearman-correlation abstention.
E.2.2. GMM Clustering: GMM Clustering groups models by behavior features and selects cluster members, with BIC minimizing the criterion used to choose the number of components.Models in one cluster are treated as having similar prediction patterns and potentially a common GP prior.
E.2.2. GMM Clustering: 22% abstentions occurred in 78 experiments when GMM clusters contained fewer than 3 models, reducing mean MAE from 0.0394 →0.0274.The abstention rule uses min_sources with default=3 and is reported to improve reliability.

E.3. Experimental Results · E.4. Recommendations

ProEval was evaluated across six benchmarks using 13–15 LLMs per benchmark, measuring Mean Absolute Error after 20 evaluation iterations. The recommended strategy is GMM + min≥3 abstention, with LOO Prior as a fallback when abstention is unacceptable.

E.3. Experimental Results: 6 benchmarks and 13–15 LLMs per benchmark were used to evaluate ProEval.The benchmarks were gsm8k, svamp, strategyqa, mmlu, jigsaw, and toxicchat.
E.3. Experimental Results: Features were PCA-reduced per-question predictions retaining 95% variance, with approximately 12 components.The metric was Mean Absolute Error between the BQ-SF estimate and true accuracy at 20 evaluation iterations.
E.3. Experimental Results: Overall performance was summarized across 78 experiments using MAE at iteration 20, sorted by median MAE with lower values preferred.The supplied passage identifies Table 26 as the overall performance summary.
E.3. Experimental Results: Pearson and Spearman correlation results were compared in a dedicated summary table.The supplied passage identifies Table 27 as the correlation comparison.
E.4. Recommendations: The method comparison summary supports GMM + min≥3 abstention as the optimal choice for combining good selection with explicit failure detection.This recommendation is based on the verified results.
E.4. Recommendations: LOO Prior is the safe default when abstention is not acceptable for the pipeline.The recommendation contrasts this fallback with the abstention-based choice.
E.4. Recommendations: The recommended implementation uses auto_select_with_abstention and can fall back to a default estimate or skip the prediction pass.The implementation also runs BQ with the selected source models.
E.4. Recommendations: Future improvements include adaptive abstention thresholds, ensembles of selection methods, online adaptation, and theoretical regret bounds.These directions are proposed for benchmark-specific adaptation and stronger guarantees.