MedPIC-Bench - Counterfactual Pair: leaderboard
Share of the 89 linked MedPIC-Bench counterfactual pairs for which a model answers both linked questions correctly, so a pair counts only when the patient-information change is handled in both directions (MedPIC-Bench, arXiv 2608.03028v1, Table 2). Percent; higher is better.
Metric: Both-correct pair rate (%). Source: arxiv.org. Saturation forecast: Around April 2027. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 48.3 | #64 |
| 2 | GPT-5 | 38.2 | #105 |
| 3 | DeepSeek V4 Pro | 32.6 | #112 |
| 4 | Qwen 3.5 Plus | 32.6 | #133 |
| 5 | Qwen 3.5 27B | 31.5 | #192 |
| 6 | GPT-OSS-120B | 30.3 | #346 |
| 7 | Qwen 3.5 35B A3B | 30.3 | #268 |
| 8 | GPT-5.2 | 29.2 | #115 |
| 9 | Qwen 3.5 9B | 24.7 | #391 |
| 10 | GPT-OSS-20B | 24.7 | #523 |
| 11 | Claude Sonnet 4.6 | 21.3 | #94 |
| 12 | Gemini 2.5 Pro | 20.2 | #159 |
| 13 | Llama 3.1 70B | 16.9 | #588 |
| 14 | Llama 3.1 8B | 5.6 | #1130 |
| 15 | Gemma 3 27B | 5.6 | #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-pair · How It Works · Data refreshed daily, snapshot 2026-11-05.