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

2026-04-24

AI Summary

This work addresses the challenges in evaluating generative AI models—namely high inference costs, expensive human annotation, and rapidly evolving benchmarks that strain evaluation resources. The authors propose ProEval, a novel framework that, for the first time, integrates pre-trained Gaussian processes with Bayesian quadrature and super-level set sampling. By leveraging transfer learning to construct a surrogate model, ProEval formulates performance estimation as Bayesian quadrature and failure discovery as super-level set sampling, actively selecting or synthesizing high-information inputs based on predictive uncertainty. Empirical results demonstrate that ProEval achieves unbiased estimates within 1% error using only 1/8 to 1/65 of the samples required by baseline methods across multiple benchmarks, while uncovering more diverse failure cases under stringent budget constraints, all backed by theoretical guarantees and computational efficiency.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Search and Optimization: Sampling/Simulation-based Search
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• User Modeling, Personalization and Recommendation: Metrics for user behavior and evaluating success

Abstract

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.

Problem

• generative AI evaluation
• performance estimation
• failure discovery
• resource efficiency
• model benchmarking

Innovation

• Proactive Evaluation
• Gaussian Processes
• Bayesian Quadrature
• Superlevel Set Sampling
• Transfer Learning