ELBench - Safety and Trustworthiness: leaderboard
The safety module of ELBench (Jiang et al., 2026): 1,000 items covering requests an education assistant should refuse, harmful requests it should redirect, benign questions it should not over-refuse, and curated EduGuardBench teaching-safety and jailbreak items. Mean per-item score in percent, higher is better; the refusal items include region-specific normative content.
Metric: Module score (%; mean per-item score on 1,000 items: synthesized refusal, safe-guidance and benign-answering requests and curated EduGuardBench teaching-safety and jailbreak items; zero-shot, temperature 0; open-ended items scored by a Qwen3.6 rubric judge, majority of 9 calls). Source: arxiv.org. Saturation forecast: Not forecast. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Flash | 89.7 | #258 |
| 2 | GLM-5.1 | 89.5 | #179 |
| 3 | DeepSeek V4 Pro | 86.1 | #188 |
| 4 | Seed 2.0 Pro | 83 | #211 |
| 5 | Claude Opus 4.8 | 78.5 | #56 |
| 6 | GPT-5.4 | 76.6 | #123 |
| 7 | Gemini 3.5 Flash | 70.2 | #101 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=elbench-safety-and-trustworthiness · How It Works · Data refreshed daily, snapshot 2026-11-13.