DPrivBench - Algorithm Level: leaderboard

Algorithm-level differential privacy verification on DPrivBench (2026): for 125 algorithms from the research literature and perturbed versions of them, the model decides whether the stated privacy guarantee holds, no tools, five seeds. F1 of the yes verdict in percent; higher is better.

Metric: F1 score (0-1, times 100) of the yes verdict on the DPrivBench Category 2 algorithm-level instances (125 algorithms from the differential privacy literature with perturbed negative versions); binary yes/no verdict on whether the stated guarantee holds, five seeds, no tools; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.5High82.9#54
2Gemini 3.1 Pro (Preview)81#99
3GPT-5.4High79.3#135
4Gemini 3 Pro74.8#145
5GPT-5High74.2#178
6DeepSeek R161.4#482
7DeepSeek V3.1Non-reasoning60.3#511
8Gemini 2.5 Flash60.2#463
9Claude Sonnet 4.551.8#261
10Claude Opus 4.548.7#160
11Qwen 3 30B A3B 2507 Instruct46#1100
12Qwen 3 30B A3B 2507 (Thinking)40.7#764
13Goedel-Prover-V2-32B36.2

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

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