MedPriv-Bench (Privacy Prompt) - Instance-Level Leakage: leaderboard
Metric: Instance-level leakage rate (%): share of answers that reveal at least one injected sensitive fact, judged by a fine-tuned RoBERTa NLI model; MedPriv-Bench's 738 test questions: the model answers a clinical question from a retrieved patient record (PMC-Patients summaries) into which identifying or stigmatizing health facts were injected; temperature 0.4; an explicit privacy constraint is prepended to the prompt; lower is better. Source: arxiv.org. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Nano | 54.6 | #415 |
| 2 | Claude Haiku 4.5 | 61.4 | #271 |
| 3 | DeepSeek V3.2 | 80.2 | #198 |
| 4 | Llama3-Med42-70B | 83.6 | #594 |
| 5 | Gemini 3 Flash | 93.2 | #93 |
Interactive version: theaggregate.ai/benchmark?slug=medpriv-bench-privacy-prompt-instance-level-leakage · How It Works · Data refreshed daily, snapshot 2026-10-11.