MedPIC-Bench - Counterfactual: leaderboard
Accuracy on MedPIC-Bench counterfactual questions, where a controlled change in patient information changes whether a medication-safety rule applies (MedPIC-Bench, arXiv 2608.03028v1, Table 2). Percent; higher is better; every evaluated model scores lower here than on guideline-following questions.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around May 2027. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 69.9 | #64 |
| 2 | GPT-5 | 63.4 | #105 |
| 3 | Qwen 3.5 Plus | 57.9 | #133 |
| 4 | GPT-OSS-120B | 56.8 | #346 |
| 5 | DeepSeek V4 Pro | 56.8 | #112 |
| 6 | Qwen 3.5 27B | 55.7 | #192 |
| 7 | Qwen 3.5 35B A3B | 54.6 | #268 |
| 8 | GPT-5.2 | 52.5 | #115 |
| 9 | Claude Sonnet 4.6 | 48.1 | #94 |
| 10 | Qwen 3.5 9B | 48.1 | #391 |
| 11 | GPT-OSS-20B | 47.5 | #523 |
| 12 | Llama 3.1 70B | 47 | #588 |
| 13 | Gemini 2.5 Pro | 44.8 | #159 |
| 14 | Llama 3.1 8B | 33.3 | #1130 |
| 15 | Gemma 3 27B | 33.3 | #606 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=medpic-bench-counterfactual · How It Works · Data refreshed daily, snapshot 2026-11-05.