Doc-PP - Explicit Query Faithfulness (No Policy): leaderboard
Baseline retrieval on the direct questions of Doc-PP (Jang, Chang and Lee, ACL 2026 Findings) with no non-disclosure policy: given page images of a clipped report, the model is asked for a specific value, and a GPT-5-mini judge checks whether the response states it. The score is the percent of responses containing the value; higher is better.
Metric: Faithfulness (%; responses stating the requested value, explicit queries without a policy, page-image input). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (Thinking) | 98 | #172 (GPT-5.2) |
| 2 | Gemini 3 Flash (Preview) (Non-reasoning) | 97.3 | #141 (Gemini 3 Flash (Preview)) |
| 3 | Qwen 3 VL 235B A22B (Thinking) | 96.8 | #451 |
| 4 | Gemini 3 Pro (Preview) | 96.6 | #111 |
| 5 | Qwen 3 VL 235B A22B Instruct | 93.5 | #521 |
| 6 | Mistral Large 3 | 89.5 | #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-explicit-query-faithfulness-no-policy · How It Works · Data refreshed daily, snapshot 2026-11-12.