EvoHarmBench (Direct Audit): leaderboard
EvoHarmBench (Findings of EMNLP 2026) submits 5,002 real-world harmful posts from content platforms to LLM content moderators without any rewriting. The miss rate is the share of posts the moderator fails to flag, in percent; lower is better.
Metric: Miss rate (%; share of the 5,002 real-world harmful posts in five violation categories that the moderator fails to flag when they are submitted unchanged, with the same category-specific moderation prompts as the iterative protocol). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 31.7 | #48 |
| 2 | Claude Sonnet 4.6 | 36.2 | #158 |
| 3 | Gemini 3.1 Pro (Preview) | 37.1 | #91 |
| 4 | Qwen 3.6 Plus | 39.5 | #217 |
| 5 | Kimi K2.6 | 39.6 | #173 |
| 6 | DeepSeek V4 Pro | 41.3 | #181 |
| 7 | GLM-5.1 | 43.3 | #175 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=evoharmbench-direct-audit · How It Works · Data refreshed daily, snapshot 2026-11-12.