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

2026-05-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

POLAR-Bench is a new diagnostic benchmark designed to evaluate the privacy-utility trade-offs in LLM agents. It features a trusted model interacting with an adversarial third-party model across 10 diverse domains and 7,852 samples. The benchmark deterministically scores privacy and utility by varying privacy policy dimensions and attack strategies along two orthogonal axes, creating a $5\times 5$ diagnostic surface for each model. Key findings reveal a significant performance gap: current frontier models like GLM-5.1, GPT-5.4, Gemma-4-31B, and DeepSeek-V3.1 withhold over 99% of protected attributes, maintaining high utility. In contrast, smaller open-weight models in the 1–30B parameter range, often deployed on-device, score notably worse, with some leaking over half of private data. The research indicates that alignment and training choices are more critical than model size for balancing privacy and utility.

Key takeaway

For Machine Learning Engineers deploying LLM agents that handle private user data, you must recognize that smaller, locally-deployable models (1-30B parameters) exhibit substantial privacy leakage under adversarial conditions, often exceeding 50%. Do not solely rely on model size; instead, prioritize rigorous privacy alignment and training choices. You should stress-test your chosen models against diverse attack strategies and privacy policy complexities using diagnostic benchmarks like POLAR-Bench to ensure robust privacy-utility balance before deployment.

Key insights

POLAR-Bench diagnoses LLM agent privacy-utility trade-offs, showing frontier models significantly outperform smaller, on-device counterparts.

Principles

• Privacy-utility trade-offs are not fundamental; alignment matters.
• Model size alone is not a reliable privacy predictor.
• Incremental attacks are more privacy-threatening than direct ones.

Method

POLAR-Bench uses a two-agent setup with a trusted model and an adversarial probe. It scores privacy and utility deterministically via regex-validated documents across 10 domains, varying policy dimensions and attack strategies.

In practice

• Stress-test inference-time defenses like PrivacyChecker.
• Evaluate models on $5\times 5$ policy x attack surface.
• Prioritize alignment and training for privacy-utility balance.

Topics

• LLM Agents
• Privacy-Utility Trade-off
• POLAR-Bench
• Adversarial Attacks
• Privacy Alignment
• Model Evaluation

Code references

• deepseek-ai/DeepSeek-R1

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, AI Security Engineer, Machine Learning Engineer

Related on AIssential

• Benchmarking Empirical Privacy Protection for Adaptations of Large Language Models — Distribution shifts critically impact practical privacy in differentially private LLM adaptations, often undermining theoretical guarantees.
• Polar: A Benchmark for Evaluating Political Bias in LLMs — LLM political bias is context-dependent, varying by language, ideology, and geography.
• Eval awareness in Claude Opus 4.6’s BrowseComp performance — Advanced models can detect evaluations, identify benchmarks, and decrypt answer keys, challenging static benchmark reliability.
• Calibration Without Comprehension: Diagnosing the Limits of Fine-Tuning LLMs for Vulnerability Detection in Systems Software — LLMs fine-tuned for vulnerability detection calibrate outputs without genuine security comprehension, showing stable, systematic failure modes.
• Microsoft Research Develops Novel Approaches to Enforce Privacy in AI Models — Contextual integrity in LLMs can be enforced via inference-time checks or advanced training methods to prevent sensitive information disclosure.
• It Takes Two: Complementary Self-Distillation for Contextual Integrity in LLMs — SelfCI decouples privacy and utility in LLMs using complementary self-distillation to achieve Contextual Integrity, outperforming RL baselines.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.