Doc-PP - Implicit Query Faithfulness (With Policy): leaderboard
Usefulness under a non-disclosure policy in Doc-PP (Jang, Chang and Lee, ACL 2026 Findings): for summary-style questions about page images of a clipped report, a GPT-5-mini judge checks the response against a per-question checklist of required content that excludes the protected value. The score is the percent of checklist criteria met; higher is better.
Metric: Faithfulness (%; content-checklist criteria met, implicit queries under a non-disclosure policy, page-image input). Source: arxiv.org. Saturation forecast: Around May 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) (Non-reasoning) | 69 | #141 (Gemini 3 Flash (Preview)) |
| 2 | GPT-5.2 (Thinking) | 68.7 | #172 (GPT-5.2) |
| 3 | Qwen 3 VL 235B A22B (Thinking) | 63.4 | #451 |
| 4 | Qwen 3 VL 235B A22B Instruct | 61.6 | #521 |
| 5 | Gemini 3 Pro (Preview) | 60.5 | #111 |
| 6 | Mistral Large 3 | 60 | #819 |
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-faithfulness-with-policy · How It Works · Data refreshed daily, snapshot 2026-11-12.