CCBench-Health - Avoid Rate: leaderboard
CCBench-Health (2026) tests whether LLMs adapt health advice to cultural norms that 60 personas from six cultures signal only implicitly in earlier conversation, scoring answers against persona-specific checklists of norms to follow or avoid. Share of recommendations satisfied for norms the persona avoids (stereotype resistance), in percent; higher is better.
Metric: Avoid rate (%; share of checklist recommendations satisfied for norms the persona avoids, a measure of stereotype resistance; 60 personas with graded norm-adherence states across six cultures (Afghan, Burmese, Chinese, Maori, Nepali, Vietnamese), 52 real health questions each (3,120 interactions); GPT-5.2 checklist evaluator; temperature 0.7; the persona conversation history is given before the question and the norms must be inferred from it)). Source: arxiv.org. Saturation forecast: Around October 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 51.7 | #176 |
| 2 | Gemini 2.5 Pro | 45.3 | #266 |
| 3 | Qwen 3.5 397B A17B | 43.7 | #246 |
| 4 | Llama 3.2 90B | 37.9 | #1733 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ccbench-health-avoid-rate · How It Works · Data refreshed daily, snapshot 2026-11-13.