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

TL;DR

ProEval is a sample-efficient evaluation framework for generative AI that frames benchmark performance estimation as Bayesian quadrature (BQ) and failure discovery as superlevel-set sampling over a pre-trained Gaussian Process surrogate of the model’s score function. Pre-training the GP on historical evaluation statistics (within-benchmark) or on semantic embeddings (cross-benchmark) gives an informed prior; an uncertainty-aware active-sampling loop then picks the next input to evaluate so as to either minimize posterior variance of the integrated score (estimation mode) or to localize regions where the score exceeds a failure threshold (discovery mode). Empirically the method reaches estimates within 1% of ground-truth using 8-65× fewer model evaluations than competitive baselines on reasoning, safety alignment, and classification benchmarks, while surfacing 2-5× more diverse failure cases under matched evaluation budgets. The work matters to anyone running large-scale model evaluation: sample efficiency directly translates to slower-iteration / lower-cost benchmarking, especially when raters or inference is expensive.

Key claims

Performance estimation can be cast as Bayesian quadrature over a GP posterior on the score function, so picking inputs to minimize posterior variance of the integral $S = \mathbb{E}_{x \sim p}[f(x)]$ yields an unbiased, theoretically bounded estimator [§2.1, §2.3].
A pre-trained GP prior — built either from historical evaluations on the same benchmark (within-benchmark transfer) or from semantic embeddings of related benchmarks (cross-benchmark transfer) — is the key mechanism delivering the 8-65× sample reduction; without transfer, naive BQ does not beat baselines [§2.2].
Failure discovery is framed as superlevel-set sampling: identify the input region $\{x : f(x) > \tau\}$ where the score function exceeds a failure threshold, sharing the same GP surrogate used for estimation [§2.4].
The estimator is proved unbiased and bounded for the pre-trained GP-based BQ setting, giving a theoretical handle that benchmark-pruning baselines lack [§2.3, Theorem statements].
On reasoning, safety-alignment, and classification benchmarks, ProEval reaches within 1% estimation error using as few as 1-27 evaluated inputs, vs the 8-65× larger sample counts required by competitive baselines such as Polo 2024-style benchmark pruning and Kossen 2021/2022-style active testing [§3, Abstract].
Failure-discovery mode achieves 2-5× higher failure detection rates with better semantic diversity than LLM-based generation methods (Perez 2022, Rainbow Teaming / Samvelyan 2024) at matched budgets [§3, Abstract].
The ablation pins responsibility on three components specifically — the transfer-learning prior, the active-sampling acquisition function, and the LLM-augmented failure-discovery loop — none of which alone explains the gains [§3, Ablations].

Method

The paper formalizes model evaluation as a dual-objective problem over input space $\mathcal{X}$ with score function $f : \mathcal{X} \to \mathbb{R}$ (e.g. error severity, safety-violation magnitude) under test distribution $p(x)$ . The two objectives are (1) estimate $S = \int f(x) p(x) dx$ to within 1% using as few evaluations of $f$ as possible, and (2) recover the superlevel set $\{x : f(x) > \tau\}$ for a fixed failure threshold.

A Gaussian Process is fit to $f$ , with the prior constructed by transfer learning. Within-benchmark transfer uses historical statistics on the same benchmark across past models; cross-benchmark transfer uses a semantic embedding of inputs to align with priors fit on related benchmarks. Bayesian quadrature reduces $S$ to a closed-form Gaussian whose variance can be computed analytically from the GP posterior; the next input is picked to minimize this variance. For discovery, the same GP is used to estimate $P(f(x) > \tau)$ , and inputs are picked to maximize information about the superlevel-set boundary. An LLM-assisted module synthesizes new candidate inputs when the benchmark’s static input pool is exhausted or biased.

Results

Performance estimation: ProEval reaches within 1% of ground-truth integrated score on reasoning, safety alignment, and classification benchmarks with 8-65× fewer evaluations than baselines — sometimes with as few as 1-27 inputs total (HTML body §3). Baselines named include subset-selection methods (Polo, Kipnis, Vivek) and surrogate-based active testing (Kossen, Berrada).
Failure discovery: 2-5× higher failure detection rate with better semantic-space diversity vs LLM-based red-teaming baselines (Perez et al. 2022, Chao et al. 2024, Mehrotra et al. 2024, Samvelyan et al. 2024 / Rainbow Teaming) under matched evaluation budgets [Abstract, §3].
Theoretical: pre-trained GP-based BQ estimator proved unbiased and bounded [§2.3].
Ablations attribute the gains to transfer-learning priors, active acquisition, and LLM-augmented input synthesis — removing any one degrades performance materially [§3].

Why it’s interesting

This is the first filed paper that attacks the cost axis of VLM/LLM benchmarking directly, rather than the fidelity axis that VLM-as-Evaluator is organized around. The seven existing papers on that page (UnifiedReward-Flex, ERNIE-Image-Aes, IGenBench, RISE-Video, UniG2U-Bench, VBVR, DiffThinker × MLLM-verifier) all argue about who the evaluator should be — a context-adaptive VLM, a static rubric, a rule-based scorer — but treat the per-sample evaluation as a one-shot pass over the full benchmark. ProEval is orthogonal: take whatever score function you have (rule-based scorer, VLM judge, human rater) and pick which inputs to feed it. Compose-ability with Unified Personalized Reward Model for Vision Generation‘s context-adaptive VLM judge or A Very Big Video Reasoning Suite‘s rule-based scorers is the natural next experiment — the GP surrogate is agnostic to what $f$ is. It also sits adjacent to Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling (compute-optimal sampling for training data) — both argue that the right axis to optimize is sample efficiency rather than raw quality, but for evaluation rather than synthetic SFT. Sample efficiency on the order of 8-65× would meaningfully shift the cost equation for any team running per-checkpoint benchmark sweeps.