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

Motivation

• Generative AI evaluation is increasingly unsustainable because slow inference, costly raters or judges, and large benchmarks make repeated comprehensive testing expensive and time-consuming.Downsampling can reduce estimate accuracy and miss rare but critical failures during iterative model development.Source: §1 · pp. 1–2

Framework

• ProEval models performance as an unknown input-to-metric function and uses transfer learning to construct an informed GP prior instead of learning the performance function from scratch.Its priors derive covariance from historical within-benchmark evaluations or features from pretrained text and image embeddings for new domains and modalities.Source: §1 · p. 2

Active evaluation

• Bayesian quadrature estimates aggregate performance by selecting inputs that minimize posterior variance, while superlevel set sampling targets likely failure regions through exploration and exploitation.The shared probabilistic framework supports both performance estimation and failure discovery as complementary objectives.Source: §1 · p. 2

Failure discovery

• An LLM synthesizes harder tests from identified failures using bi-level topic-aware sampling, reducing collapse into one failure mode and promoting diverse failure patterns.High-level semantic topics are sampled before specific cases, with hard examples serving as in-context anchors.Source: §1 · pp. 2–3

Empirical results

• Within 1% estimation error, ProEval often requires only 1–27 evaluated inputs and detects 2–5x more failures with greater semantic diversity than competing methods.The estimator is theoretically unbiased and bounded under mild assumptions, while ablations identify transfer learning, active sampling, and LLM-enhanced discovery as critical components.Source: §1 · pp. 2–3

Contributions

• ProEval unifies performance estimation and failure discovery in a proactive, sample-efficient Bayesian framework that integrates informed priors, objective-specific acquisition functions, and hierarchical synthesis.This unified design differentiates it from benchmark pruning, surrogate testing, and red-teaming methods that are often model-agnostic, high-variance, manual, or sample-inefficient.Source: §1 · p. 3

2.1. Bayesian Formulation of Evaluation

• ProEval formalizes evaluation as estimating expected score S with minimal evaluations while identifying the failure superlevel set X_λ where scores exceed threshold λ.A GP surrogate models the noisy score function and yields posterior distributions for both objectives.Source: §2.1 · pp. 3–4

2.2. Transfer Learning for GP Priors

• Transfer learning constructs GP priors from historical evaluation data, exploiting structural correlations across inputs so performance on correlated examples informs unobserved inputs.Figure 2 supports the shared-covariance assumption across models and benchmark questions.Source: §2.2 · pp. 4–5, Figure 2 · p. 5

2.2.1. Score features via empirical statistics

• On shared benchmarks, empirical score features produce a GP prior whose posterior performance estimator is unbiased and has a worst-case error bound that improves with more historical models.The bound is non-vacuous when the number of historical models N grows faster than the number of test samples M.Source: §2.2.1 · pp. 5–6

2.2.2. Prompt features via learned embeddings

• When historical scores are unavailable, learned input embeddings define a GP prior that enables zero-shot predictions for unseen inputs through semantic similarity.The embedding-based GP can use linear or Matérn kernels and is trained by maximizing likelihood across historical datasets.Source: §2.2.2 · p. 6

2.2.3. Selecting source data

• To prevent negative transfer, ProEval clusters historical models by behavioral profiles and builds priors from the target model’s cluster, abstaining when fewer than three models are available.Out-of-distribution targets can violate the shared-distribution assumption underlying transfer learning.Source: §2.2.3 · p. 7

2.3. Active Performance Estimation

• Active performance estimation greedily selects inputs that maximize posterior variance reduction, and its observation-independent acquisition function permits precomputed batches for parallel evaluation.For linear kernels, Sherman–Morrison updates avoid repeatedly inverting a (t+1)-dimensional Gram matrix.Source: §2.3 · p. 7

2.4. Proactive Failure Case Discovery

• ProEval discovers failures through three increasingly flexible strategies: superlevel-set retrieval, LLM-based synthesis from likely-failing anchors, and topic-aware synthesis for broader diversity.Superlevel-set sampling combines probable failure membership with uncertainty, while topic-aware exploration selects topics via UCB1 independently of anchors.Source: §2.4 · p. 8

3.1.1. Datasets and models

• The evaluation covers reasoning, world knowledge, and safety-alignment datasets across text and image modalities, using 16 LLMs and VLMs with binary failure scores.Scenarios include Default, New Model, and New Bench, with score, raw-prompt, and tuned-prompt feature configurations.Source: §3.1.1 · pp. 9–10, §3.1 · p. 9

3.1.2. Evaluation Metrics

• Performance estimation uses MAE and samples required to reach ≤1% MAE, while failure discovery measures cumulative failures, failure rate, first-failure samples, and semantic-topic diversity.Overall diversity combines normalized topic entropy with embedding diversity and penalizes near-duplicates or mode collapse.Source: §3.1.2 · p. 10

3.2. Results on Performance Estimation

• At a 1% labeling budget, ProEval variants consistently outperform all baselines, while active BQ selection is almost always better than random BQ selection.BQ-SF benefits from transferred priors, and BQ-TPF generally converges faster as samples are acquired.Source: §3.2 · p. 11, Table 1 · p. 12
• Transfer improves estimation when source data are aligned: cross-modality transfer reduces MAE from 0.111 to 0.055, whereas unfiltered negative transfer can increase MAE by up to 100×.In New Bench, BQ-RPF outperforms Random Sampling on 6 out of 9 datasets; reasoning traces generally reduce error, although the best integration varies by task.Source: §3.2 · pp. 11–12, Figure 7 · p. 14

3.3. Results on Failure Case Discovery

• ProEval’s active generative strategies discover failures faster and in greater numbers than random baselines, requiring fewer than 8 samples for math failures versus 11–27 for random methods.Cumulative failures grow near linearly for SS-Gen and TSS, with the largest gap in the math domain.Source: §3.3 · p. 14, Figure 8 · p. 15
• Across five target models, guided generation outperforms Rand-Gen; on GPT 5, TSS-TPF finds 18.9% failures versus 5.1%, a 3.7× improvement, on StrategyQA.TSS-TPF also achieves 0.74 overall diversity on StrategyQA and 0.73 on GSM8K, while relative gains persist across target-model difficulty levels.Source: §3.3 · p. 15, Table 4 · p. 16

• ProEval combines Bayesian ideas and transfer learning to improve performance estimation and failure case discovery.The framework targets evaluations where querying and rating modern generative AI models are expensive.Source: §4 · p. 17
• ProEval reduces evaluation sample sizes by 8-65x and discovers 2-5x more failure cases than competitive baselines.These results support the framework’s promise for expensive-to-query and expensive-to-rate generative AI models.Source: §4 · p. 17
• Strong GP priors improve efficiency, while prior-quality checks let ProEval abstain when source data are insufficient for reliable prediction.Without informative priors, the surrogate must learn from scratch through costly direct observations.Source: §4 · p. 17
• Stronger embedding models produce more accurate performance estimation, making embedding quality an important factor in ProEval’s effectiveness.This conclusion follows from an ablation study on BQ-RPF using different embedding models.Source: §4 · p. 17
• Learned embeddings capture transferable latent performance patterns across datasets, and incorporating CoT reasoning into thinking-trace embeddings can reduce estimation error.Future work includes reducing dependence on high-quality embeddings and improving acquisition functions, surrogate models, variance reduction, and cost-aware strategies.Source: §4 · p. 17

Impact Statement

• ProEval’s reduced evaluation overhead is intended to accelerate GenAI research and quality iteration while improving transparency, energy efficiency, and model safety and equity.The framework prioritizes diverse and challenging test cases and systematically uncovers and forecasts model failures.Source: Impact Statement · p. 17

A. Literature Review

• Unlike static benchmark pruning and adaptive testing methods, ProEval can deviate from source-model priors, synthesize new questions, and explore target-specific weaknesses.Static subsets may miss idiosyncratic failures, while Fluid Benchmarking is restricted to static item pools and a single scalar ability score.Source: A. Literature Review · p. 25
• ProEval frames evaluation as Bayesian quadrature to estimate aggregate metrics directly, rather than reconstructing individual scores, while also supporting failure discovery and query synthesis.This contrasts with active-testing approaches based on importance sampling, imputation, label weighting, or cold-start evaluation without cross-model structure.Source: A. Literature Review · pp. 26–27
• ProEval differs from standard active learning by keeping the model fixed and selecting samples to estimate test metrics or discover failure regions, not to reduce model generalization error.Its acquisition functions prioritize variance reduction for integral estimation or superlevel-set locations rather than information gain about model parameters.Source: A. Literature Review · p. 26
• For costly simulated-user and agentic evaluations, ProEval treats the simulation as a costly black-box oracle and samples only the most critical test cases.This is intended to make high-fidelity interactive evaluation more practical for iterative development.Source: A. Literature Review · p. 26
• ProEval addresses red-teaming limitations by mapping the entire superlevel set of model failures through black-box Bayesian level-set estimation instead of optimizing for a single failure mode.Prior methods trade efficiency and diversity through attacker fine-tuning, prompt optimization, evolutionary search, concept edits, or heuristic diversity penalties.Source: A. Literature Review · p. 26
• ProEval adapts Bayesian quadrature to high-dimensional generative-AI metric estimation and addresses its cold-start problem with pre-trained Gaussian processes and score-space transfer learning from historical data.The cited BQ literature primarily targets scientific-computing quantities such as marginal likelihoods, whereas ProEval targets metrics such as accuracy and safety over semantic input spaces.Source: A. Literature Review · pp. 26–27

• The proof establishes that the sample mean estimator ˆ𝜇_t(x) is unbiased and bounded by the posterior variance.Source: B. Proof for Theorem 3 · p. 27
• The integral estimator constructed from the estimated GP mean and kernel is unbiased for the ground-truth expectation of the integral.Source: B. Proof for Theorem 3 · p. 27
• With probability 1 − 𝛿, the integral estimate satisfies the stated bound, where 𝜅 is a constant bounding the kernel value.Source: B. Proof for Theorem 3 · p. 27

• The experiments include additional results alongside detailed experimental descriptions.Source: C. Details on Experiments and… · p. 28
• Source data are selected with a GMM-based approach using all benchmarks except the target benchmark as reference data.Source: C. Details on Experiments and… · p. 28

C.1. Datasets

• ProEval’s evaluation spans mathematical reasoning, general world knowledge, and safety alignment, covering challenges from objective correctness to nuanced human judgment.Most datasets are subsampled to accelerate experimentation and reduce inference costs.Source: §C.1 · p. 28
• The reasoning suite includes GSM8K and SVAMP for multi-step mathematical reasoning, StrategyQA for implicit reasoning, and GQA for multimodal reasoning.GSM8K emphasizes numeric solutions, SVAMP tests phrasing robustness, and StrategyQA requires latent decomposition of Yes/No questions.Source: §C.1 · p. 28
• MMLU evaluates general-purpose knowledge across 57 subjects, using abstract algebra and professional law subsets with differing difficulty levels.The benchmark uses a multiple-choice format spanning STEM, humanities, and social sciences.Source: §C.1 · p. 28, Table 6 · p. 28
• ToxicChat, Google Civil Comments, Google DICES-350, and text-to-image DIVE assess safety and human alignment using toxicity data and expert ratings.DIVE contains expert safety ratings across 123 unique raters.Source: §C.1 · p. 28
• For safety datasets, models are evaluated as reward models, with f(x) measuring agreement between predicted labels and ground-truth human annotations.This evaluation differs from the reasoning tasks by directly measuring alignment with human judgments.Source: §C.1 · p. 28

C.1.1. Baselines for performance estimation

• Random Sampling estimates global accuracy with uniformly selected test cases and the sample mean, while Active Testing uses surrogate-guided acquisition with selection-bias correction Kossen et al. (2021).The active surrogates are trained on evaluation results from 11 models to define acquisition probabilities q(i).Source: §C.1.1 · p. 29
• The active baseline variants are LR+IS, LR+LURE, RF+IS, and RF+LURE, combining logistic-regression or random-forest surrogates with two unbiased estimation schemes.Importance Sampling uses inverse acquisition-probability weights, whereas LURE accounts for sequential sampling without replacement.Source: §C.1.1 · p. 29

C.2. Ablation Studies with Pre-training Data Selection

• Uniformly positive performance correlations across model pairs on StrategyQA and SVAMP show that models share failure modes.These correlations support using historical model data to pre-train ProEval’s Gaussian-process surrogate.Source: §C.2 · p. 29, Figure 9 · p. 29
• ProEval selects historical pre-training datasets by identifying models outside the target-relevant performance cluster and removing them.For Gemini 2.5 Flash on StrategyQA or GSM8K, Gemma-3-12B and GPT-3.5 Turbo are excluded using SVAMP as the hold-out benchmark.Source: §C.2 · p. 29, Figure 10 · p. 30
• 100× higher MAE can result from poor pre-training model pairs, and the optimal pair varies by target model on GSM8K.Figure 11 reports pairwise pre-training ablations for two target models, with lower MAE indicating better prior quality.Source: §C.2 · p. 30, Figure 11 · p. 31

• Active Selection + BQ achieves the lowest MAE on most benchmarks across the evaluated target models, outperforming random sampling and LURE.The comprehensive evaluation uses the same experimental setting as Table 1 and reports results in Tables 7–21.Source: §C.3.1 · p. 31, Table 7 · p. 32, Table 8 · p. 32
• Tables 11–16 report MAE at a 1% budget for GPT-5.2, Claude 3.5 Haiku, Claude 3.7 Sonnet, Claude 4.5 Sonnet, Claude 4.5 Opus, and Gemini 2.5 Pro.Active BQ is deterministic, while other methods report mean ± standard deviation over five runs.Source: Table 11 · p. 33, Table 12 · p. 33, Table 13 · p. 34
• Tables 17–21 extend the 1% MAE evaluation to Gemini 3 Flash, Gemini 3 Pro, Gemma 3 12B, Gemma 3 27B, and Qwen 3 32B.DIVE and GQA results are unavailable for GPT-3.5 Turbo and Qwen 3 32B because those models do not support image inputs.Source: Table 7 · p. 32, Table 17 · p. 35, Table 18 · p. 35

C.3.2. Ablation Study on Different Embedding Models

• Larger embedding dimensions tended to yield lower MAEs when evaluating Gemini 2.5 Flash on StrategyQA with BQ-RPF and a sampling budget of 50.The study compared Google gemini_embedding_001, OpenAI text_embedding_3_small, OpenAI text_embedding_3_large, and all_minilm_l6_v2.Source: Table 22 · p. 37, §C.3.2 · p. 37

C.4. Additional Results on Failure Case Discovery

• Score 1.0 active-generation examples consistently reveal complex failures, whereas random generation often produces simpler questions models solve with Score 0.0.The examples span StrategyQA-like multi-hop constraints and GSM8K-like problems requiring multi-step reasoning, systems of equations, or layered constraints.Source: §C.4 · p. 38, Table 23 · p. 38, Table 24 · p. 39

D. Generation Prompts for Failure Discovery

• Generation prompts vary topic constraints, hard-example anchors, and tuned prompt features while preserving a common base structure and verified JSON outputs.The prompt families combine no or selected topics with no anchors, random anchors, or reasoning-pattern anchors across GSM8K-like and StrategyQA-like generators.Source: D. Generation Prompts for Failure… · p. 40, §D.1.4 · p. 42, §D.1.5 · p. 42

E.1. Problem Setting

• The source-selection problem chooses historical models for estimating the GP prior of a new target under limited evaluation budgets, excluding the target to prevent leakage.Using all models can violate the shared-prior assumption for out-of-distribution targets, while irrelevant sources can corrupt Bayesian-quadrature predictions.Source: §E.1 · p. 46, §E.2.1 · p. 46

E.2.1. Leave-One-Out Prior (Baseline)

• Selection alternatives target different notions of similarity or typicality: Spearman correlation, Mahalanobis distance, leave-one-out likelihood, and distributional hypothesis testing.Leave-one-out uses all non-target models, correlation selects highly correlated prediction profiles, Mahalanobis selects geometrically closest models, and likelihood selects typical models.Source: §E.2.1 · p. 46, §E.2.3 · p. 47, §E.2.4 · p. 47

E.2.2. GMM Clustering

• Mean MAE drops from 0.0394 to 0.0274 when GMM selection abstains on clusters with fewer than 3 models, with 17 abstentions among 78 experiments.GMM clusters models by behavior features, with the number of components selected by BIC minimization; abstention is intended to improve reliability when too few sources are available.Source: §E.2.2 · pp. 46–47

E.2.7. Mardia’s Multivariate Normality Test

• Mardia’s test computes skewness and kurtosis on combined data X = [ϕ*, ϕ_i]^⊤ and selects the top-k models using combined p-values via Fisher’s method.Source: §E.2.7 · p. 48
• The combined normality assumption rarely holds in practice, making this method too strict for practical abstention.Source: §E.2.7 · p. 48

E.3. Experimental Results

• GMM + min≥3 achieves the best median MAE of 0.0109 across the experimental comparison.It abstains when clusters contain fewer than three models, covering 22% of cases.Source: §E.3 · p. 49
• Abstention improves GMM reliability by reducing mean MAE 30% and standard deviation 28%, identifying one- or two-model selections as failure modes.The evaluation uses MAE between the BQ-SF estimate and true accuracy at iteration 20 across six benchmarks and 13–15 LLMs per benchmark.Source: §E.3 · p. 49
• Spearman-based abstention for LOO Prior reaches the lowest standard deviation, 0.0175 at τ≥0.7, but abstains in 54% of cases.Spearman is described as more robust to outliers in per-question predictions than Pearson.Source: §E.3 · p. 49, Table 27 · p. 49
• Advanced selection methods using top-k=5 underperform because selecting too few models increases variance and can leave insufficient data for a stable prior.Using all models is more stable when the selective choice is wrong.Source: §E.3 · p. 49

E.4. Recommendations

• GMM + min≥3 abstention is recommended as the optimal pipeline choice, combining effective source selection with explicit failure detection.The recommendation is based on the verified experimental results.Source: §E.4 · p. 50
• When abstention is unacceptable, LOO Prior is recommended as a safe default; future work includes adaptive thresholds, ensembles, online adaptation, and regret bounds.These future directions target benchmark-specific abstention, combined selection strategies, evaluation-time adaptation, and theoretical guarantees.Source: §E.4 · p. 50