ProEval

Slash GenAI evaluation costs by up to 100x while actively discovering model failure patterns to guide better AI development.

1. 💰 Cut GenAI eval costs up to 100× — achieve ±1% accuracy with a fraction of the samples
2. 🔍 Discover failure cases — proactively surface diverse bugs under strict evaluation budgets
3. 🧠 Transfer learning over benchmarks — pre-trained GP surrogates generalize to new models instantly
4. 🧩 Easy Integration - Easily to integrate into the GenAI evaluation systems with different modalities
5. ✅ Validated on reasoning, safety & classification — GSM8K, MMLU, StrategyQA, Jigsaw, and more

Installation

Published release

Install ProEval from PyPI. The current 0.1.0 release requires Python 3.10 or newer:

python -m pip install --upgrade proeval
python -m pip install --upgrade "proeval[topic]"  # Optional topic generation

The 0.1.0 package predates recent API and sampling improvements on main, installs PyTorch as a core dependency, and does not include the repository's research data.

Current main

The current source supports Python 3.9 or newer and provides the APIs used by the examples below. Clone the repository for the latest library code or the research artifacts under data/:

git clone https://github.com/google-deepmind/proeval.git
cd proeval
python -m pip install --upgrade pip
python -m pip install -e .

Optional extras:

python -m pip install -e ".[encoder]"   # PyTorch — for BQEncoderSampler and encoder training
python -m pip install -e ".[topics]"    # BERTopic + HDBSCAN — for TopicAwareGenerator
python -m pip install -e ".[datasets]"  # HuggingFace datasets — for evaluator.load_dataset_data
python -m pip install -e ".[all]"       # everything above
python -m pip install -e ".[dev]"       # pytest, ruff, build tooling

Quick Start

This example targets the current main API. It uses an in-memory prediction table, so it works without an API key or the repository's research data. Prediction columns must be named label_<model> and use 1 = failure, 0 = correct.

import pandas as pd

from proeval import BQPriorSampler

predictions = pd.DataFrame(
    {
        "label_reference_a": [0, 0, 1, 0, 1, 0],
        "label_reference_b": [0, 1, 1, 0, 0, 0],
        "label_candidate": [0, 0, 1, 0, 1, 1],
    }
)

result = BQPriorSampler(noise_variance=0.3).sample(
    predictions=predictions,
    target_model="candidate",
    budget=3,
    pretrain_mode="all",  # Required for an unnamed in-memory DataFrame.
    seed=42,
)

print(f"Estimated error rate: {result.estimates[-1]:.4f}")
print(f"Selected rows: {result.selected_indices}")

See the Python API guide for custom datasets, model evaluation, and test-case generation.

Reproducing the Paper Experiments

The pre-computed prediction CSVs and embeddings under data/ are research artifacts stored only in the GitHub repository. They are approximately 381 MB and are intentionally not included in the PyPI wheel or source distribution. Clone the repository before using dataset-name shortcuts such as predictions="svamp", or pass your own DataFrame or explicit data_dir.

From a source checkout, run the bundled example with:

python experiment/sample_usage.py

Integrating an external evaluation harness

Plan evaluations using only historical source-model scores, then send the selected IDs to any agent or inference benchmark. Source and target scores should use the same per-item metric convention (for example, 1=error and 0=correct).

from proeval import BQPriorSampler

# Rows are benchmark items; columns are previously evaluated source models.
# A DataFrame's index is used as the stable item ID.
plan = BQPriorSampler(noise_variance=0.3).plan(
    source_scores=historical_scores,
    budget=50,
)

# This loop can be replaced by any third-party agent or inference harness.
target_scores = {
    item_id: run_target_model(item_id)
    for item_id in plan.item_ids
}
result = plan.estimate(target_scores)
print(f"Estimated target error rate: {result.estimates[-1]:.4f}")

estimate also accepts a list or NumPy array ordered like plan.item_ids.

Experiments

Here is an example of how to run the experiments:

# BQ performance estimation (runs BQ-SF, BQ-RPF, etc.)
python -m experiment.exp_performance_estimation --dataset svamp --n-runs 5

You can find the comprehensive experiment details and dataset settings here.

Citation

If the work did some helps on your research/project, please cite our ICML 2026 paper. Thank you!

@inproceedings{huang2026proeval,
  title={{{ProEval}: Proactive Failure Discovery and Efficient Performance Estimation for Generative AI Evaluation}},
  author={Huang, Yizheng and Zeng, Wenjun and Kumaresan, Aditi and Wang, Zi},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2026},
  url={https://arxiv.org/abs/2604.23099}
}

License

Copyright 2026 DeepMind Technologies Limited

All software is licensed under the Apache License, Version 2.0 (Apache 2.0);
you may not use this file except in compliance with the Apache 2.0 license. You
may obtain a copy of the Apache 2.0 license at:
https://www.apache.org/licenses/LICENSE-2.0

All other materials are licensed under the Creative Commons Attribution 4.0
International License (CC-BY). You may obtain a copy of the CC-BY license at:
https://creativecommons.org/licenses/by/4.0/legalcode

Unless required by applicable law or agreed to in writing, all software and
materials distributed here under the Apache 2.0 or CC-BY licenses are
distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND,
either express or implied. See the licenses for the specific language governing
permissions and limitations under those licenses.

This is not an official Google product.