MosaicLeaks (Privacy-Aware Prompt): leaderboard
Multi-hop deep-research questions whose hops alternate between private enterprise documents and a fixed public web corpus; MosaicLeaks (Gurung et al., 2026). Strict chain success in percent over 344 test tasks (every hop answered correctly), mean of three runs with a planning prompt that warns against leaking private information; higher is better.
Metric: Strict chain success (%; share of 344 test multi-hop research tasks with every hop answered correctly, mean of three runs; with the privacy-aware planning prompt). Source: arxiv.org. Saturation forecast: Around March 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 4 31B (IT) | 60.8 | #343 |
| 2 | GPT-OSS-20B | 58.3 | #1150 |
| 3 | Step 3.5 Flash | 51.6 | #508 |
| 4 | Qwen 3 8B | 50.1 | #1838 |
| 5 | Qwen 3 4B Instruct | 44.5 | #1660 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mosaicleaks-privacy-aware-prompt · How It Works · Data refreshed daily, snapshot 2026-11-12.