Doc-PP - Implicit Query Leakage: leaderboard
Policy-preserving document QA from Doc-PP (Jang, Chang and Lee, ACL 2026 Findings): given page images of a clipped report and a policy forbidding disclosure of one value, the model answers a summary-style question that cannot be answered fully without that value. A GPT-5-mini judge marks responses that reveal it; the score is the percent that leak, so lower is better.
Metric: Leakage Rate (%; responses revealing the protected value, implicit summary-style queries under a non-disclosure policy, page-image input). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (Thinking) | 25.7 | #172 (GPT-5.2) |
| 2 | Qwen 3 VL 235B A22B (Thinking) | 40.4 | #451 |
| 3 | Gemini 3 Pro (Preview) | 46.8 | #111 |
| 4 | Gemini 3 Flash (Preview) (Non-reasoning) | 64.6 | #141 (Gemini 3 Flash (Preview)) |
| 5 | Mistral Large 3 | 76.8 | #819 |
| 6 | Qwen 3 VL 235B A22B Instruct | 93.5 | #521 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=doc-pp-implicit-query-leakage · How It Works · Data refreshed daily, snapshot 2026-11-12.