QuarkMedBench - Pitfall: leaderboard
Metric: Pitfall penalty rate in negative percentage points (0 means no penalty): the share of Pitfall rubric penalties (factual errors and medical safety risks) the answer incurs, on a stratified sample of 2,000 real Chinese online medical and health queries (single- and multi-turn), answers scored against query-specific rubrics (Essential, Important, Highlight and Pitfall items) by a Qwen3-235B judge, without an output length limit; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | -5.8 | #105 |
| 2 | Qwen 3 30B A3B 2507 (Thinking) | -6.23 | #366 (Qwen 3 30B A3B 2507) |
| 3 | Gemini 2.5 Pro | -6.39 | #145 |
| 4 | DeepSeek R1 0528 | -7.07 | #217 |
| 5 | GPT-5.1 | -7.78 | #131 |
| 6 | DeepSeek V3.2 | -8.66 | #198 |
| 7 | Gemini 3 Pro | -9.32 | #77 |
| 8 | Seed-1.6 | -11.2 | #257 |
| 9 | Qwen 3 235B A22B (Non-reasoning) | -12.1 | #304 (Qwen 3 235B A22B) |
| 10 | Qwen 3 32B (Non-reasoning) | -14.42 | #424 (Qwen 3 32B) |
| 11 | Qwen 3 30B A3B (Non-reasoning) | -14.75 | #488 (Qwen 3 30B A3B) |
| 12 | Qwen 3 8B (Non-reasoning) | -18.04 | #667 (Qwen 3 8B) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=quarkmedbench-pitfall · How It Works · Data refreshed daily, snapshot 2026-10-11.