AdversaRiskQA: leaderboard
Adversarial factuality benchmark (AdversaRiskQA; Szelestey et al., Eindhoven University of Technology, 2026): health, finance and law questions at basic and advanced difficulty open with a confidently stated false premise (“As we know, ...”), and a GPT-5-mini judge marks a response correct when it corrects or contradicts the premise. The score is the paper's filtered accuracy in percent: the mean of the six domain-by-difficulty accuracies over the 225 of 400 prompts that every evaluated model answered with a valid response, at temperature 0; higher is better.
Metric: Accuracy (%; mean of six domain-by-difficulty accuracies on the 225 prompts every model answered validly, GPT-5-mini judge). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 Next 80B A3B Instruct | 94.7 | #763 |
| 2 | GPT-5 | 91.4 | #163 |
| 3 | Qwen 3 30B A3B 2507 Instruct | 87.5 | #1078 |
| 4 | Qwen 3 4B 2507 Instruct | 85.5 | #2174 |
| 5 | GPT-OSS-120B | 76.8 | #699 |
| 6 | GPT-OSS-20B | 74.9 | #1150 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=adversariskqa · How It Works · Data refreshed daily, snapshot 2026-11-12.