DPrivBench - Mechanism Level: leaderboard

Differential privacy verification on the mechanism-level instances of DPrivBench (2026), where the model decides whether a stated differential privacy guarantee holds for a textbook mechanism instantiated with a concrete query, no tools, five seeds: mean accuracy over the six mechanism tasks (Laplace, Gaussian under GDP and zCDP, and three report-noisy-max variants). Percent; higher is better.

Metric: Accuracy (0-1, times 100) on the six mechanism tasks (mean of the per-task accuracies) 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. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5High99.5#178
2Gemini 3 Pro92.3#145
3DeepSeek V3.1Non-reasoning84.1#511
4Gemini 2.5 Flash77.8#463
5DeepSeek R174.1#482
6Claude Sonnet 4.573.8#261
7Claude Opus 4.572.8#160
8Qwen 3 30B A3B 2507 Instruct67.1#1100
9Goedel-Prover-V2-32B62.8
10Qwen 3 30B A3B 2507 (Thinking)56.5#764

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=dprivbench-mechanism-level · How It Works · Data refreshed daily, snapshot 2026-11-21.