POLAR-Bench

A Diagnostic Benchmark for Privacy-Utility Trade-offs in LLM Agents

Poster at NeurIPS 2026 E&D Track.

Qiaoyuan Zheng*, Yiqu Yang*, Qi Gao*, Imanol Schlag†
ETH Zurich · ETH AI Center
* Equal contribution. † Corresponding author.

Published by Qiaoyuan ZHENG.

Project website · Paper · Evaluation code · Quick start

POLAR-Bench evaluates whether an LLM agent shares task-required information while protecting private attributes during adversarial interaction. It contains 7,852 evaluation instances across 10 domains, crossing five privacy-policy formulations with five conversational attack protocols.

Ready-to-evaluate data

This repository contains only the final benchmark, ready for the evaluator. No data generation, rendering, validation, repair, or filtering pipeline needs to be run by users. All code and construction artifacts are maintained on GitHub.

File	Use
data/privacy_benchmark_rendered_repaired.json	Final evaluator input: 7,852 instances. Pass this directly to scripts/ab_eval.py --dataset.
data/privacy_benchmark_rendered_repaired.csv	The same final benchmark in flattened tabular form, for the dataset viewer and inspection.

The two files are alternative representations, not separate datasets. The JSON retains nested fields and is the format expected by the evaluation script. CSV list/object cells are serialized. test denotes the evaluation split; it is not a newly generated partition.

Download

Download just the evaluation input:

pip install -U huggingface_hub
hf download Qiaoyuan/POLAR-Bench \
  data/privacy_benchmark_rendered_repaired.json \
  --repo-type dataset --local-dir POLAR-Bench-data

The resulting input path is POLAR-Bench-data/data/privacy_benchmark_rendered_repaired.json.

Evaluate

Get the evaluation code from GitHub and follow its ready-to-evaluate quick start. For deployment and connection instructions, see Connect your own Model A and Model B: existing platforms, vLLM/SGLang backends, separate endpoints behind a gateway, API keys, exact served model IDs, and tests for both model routes. The current evaluator uses one shared endpoint/key for the compatible A/B setup; separately hosted models are connected through the gateway described in the guide. Point --dataset to the downloaded JSON. Configure credentials and endpoints for the trusted model and external attacker using POLAR_API_KEY, --base-url (or POLAR_BASE_URL), and --model-a-provider openai-compatible. The API key belongs to your chosen model service; the base URL is that service's OpenAI-compatible API address. Downloading data does not launch models. Evaluation makes model calls and may incur API charges. Use a small sample count for a smoke test, then --samples-per-domain 0 across all 10 domains for the full benchmark. A checkpoint supports resuming an interrupted run with the same configuration.

Structure

Canonical records include domain, sample_id, rendered texts, metadata, hidden policy targets, and predefined scoring targets. All information required by the released evaluator is retained; no benchmark fields have been removed or regenerated.

Domains: medical, recruitment, finance, education, customer support, legal, insurance, housing, travel, and cybersecurity.

Metrics and limitations

• Privacy: protection of benchmark-specified protected attributes.
• Attribute Utility: coverage of benchmark-specified task-required attributes disclosed by the trusted model, not end-to-end task success.
• Outcomes are scored against predefined attribute-value targets, not by an LLM judge.

The data is synthetic. Labels operationalize task necessity and policy constraints in these scenarios, not universal privacy norms. Explicit attribute scoring does not cover all semantic inference or side channels. Policy and attack types are not validated monotonic difficulty hierarchies.

Provenance and license

Both data files match the Git LFS SHA-256 hashes in GitHub commit c65eb83. The final benchmark has not changed; only the distribution has been simplified.

Dataset license: CC BY 4.0. Evaluation code is separately distributed under MIT on GitHub.

Acknowledgments and Disclosure of Funding

This work was supported as part of the Swiss AI Initiative by compute grant infra01 from the Swiss National Supercomputing Centre (CSCS) on Alps.

Citation

@article{zheng2026polar,
  title={POLAR-Bench: A Diagnostic Benchmark for Privacy-Utility Trade-offs in LLM Agents},
  author={Zheng, Qiaoyuan and Yang, Yiqu and Gao, Qi and Schlag, Imanol},
  journal={arXiv preprint arXiv:2605.19127},
  year={2026}
}