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

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 $\pm1\%$ of the ground truth, while simultaneously revealing more diverse failure cases under a stricter evaluation budget. Our open-sourced code and data can be found at https://github.com/google-deepmind/proeval.

Testing new AI takes a lot of time and money because developers usually have to check thousands of answers. To save cost, they often only test a small batch of questions, which means they can easily miss rare but dangerous mistakes. To fix this, we created ProEval, a smart system that uses past test results to guess where a new AI is most likely to mess up. Instead of testing randomly, ProEval handpicks the trickiest questions to intentionally expose the AI's weak spots. It can even ask another AI to invent brand-new, super-hard questions on the exact topics the new AI is already struggling with. This makes testing incredibly fast, requiring up to 65 times fewer questions to figure out how good an AI actually is. Plus, it uncovers up to 5 times more hidden flaws and safety issues than normal testing methods. Ultimately, this helps build safer, more trustworthy AI without breaking the bank.