Safety-Flag - AURC: leaderboard
Selective-prediction quality of LLM content moderators on Safety-Flag (Hu, 2026): items are ranked by the token-probability confidence of the chosen verdict and the area under the resulting risk-coverage curve is measured, pooled over the seven sources. Scale 0 to 1; lower is better.
Metric: Area under the risk-coverage curve (0-1 scale, lower is better; items ranked by the token-logprob confidence of the chosen option, pooled over 198-200 class-balanced items from each of BeaverTails, XSTest, Ethics, WildGuard, Aegis, ToxiChat and ToxiGen, recast as one flag / do-not-flag decision with harmonised source labels; one standardized two-option prompt with a JSON verdict and confidence; greedy decoding; not computable without option logprobs). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 2 9B (IT) | 0.08 | #2275 |
| 2 | GPT-4.1 Mini | 0.09 | #775 |
| 3 | Qwen 2.5 7B Instruct | 0.1 | #2583 |
| 4 | DeepSeek R1 Distill Llama 8B | 0.12 | #6773 |
| 5 | Mistral 7B Instruct (v0.3) | 0.14 | #6698 |
| 6 | OLMo-2-1124-7B-Instruct | 0.25 | #7377 |
| 7 | Llama 3.1 8B Instruct | 0.3 | #4277 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=safety-flag-aurc · How It Works · Data refreshed daily, snapshot 2026-11-12.