DPrivBench - Permute-and-Flip: leaderboard
Metric: Accuracy (0-1, times 100) on the report-noisy-max with exponential noise (permute-and-flip) instances of the DPrivBench Category 1 mechanism-level instances (a mechanism template instantiated with a concrete query function from a 49-function bank; positive instances calibrate the noise to the true sensitivity, paired negatives under-calibrate it); binary yes/no verdict on whether the stated differential privacy guarantee holds, five seeds, no tools; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 (High) | 99.6 |
| 2 | Gemini 3 Pro | 93.9 |
| 3 | Gemini 2.5 Flash | 66.5 |
| 4 | Claude Opus 4.5 | 65.7 |
| 5 | DeepSeek V3.1 (Non-reasoning) | 65.1 |
| 6 | Claude Sonnet 4.5 | 60 |
| 7 | GPT-5 (Minimal) | 53.3 |
| 8 | DeepSeek R1 | 52.2 |
| 9 | Qwen 3 30B A3B 2507 Instruct | 52 |
| 10 | Qwen 3 30B A3B 2507 (Thinking) | 46.5 |
Interactive version: theaggregate.ai/benchmark?slug=dprivbench-permute-and-flip · How It Works · Data refreshed daily, snapshot 2026-10-07.