Safety-Flag - AURC: leaderboard

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

#ModelScore
1Gemma 2 9B (IT)0.08
2GPT-4.1 Mini0.09
3Qwen 2.5 7B Instruct0.1
4DeepSeek R1 Distill Llama 8B0.12
5Mistral 7B Instruct (v0.3)0.14
6OLMo-2-1124-7B-Instruct0.25
7Llama 3.1 8B Instruct0.3

Interactive version: theaggregate.ai/benchmark?slug=safety-flag-aurc · How It Works · Data refreshed daily, snapshot 2026-09-26.