RealGuardrails - Distractors: leaderboard
Same system prompts, but 504 user turns try to pull the model off task with in-context demonstrations and other distractions; the number is the % that still follow the system prompt. It tests holding the guardrails under distraction rather than under a direct conflicting request.
Metric: Distractor test pass rate (%, higher is better). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 Mini | 81.5 | #533 |
| 2 | DeepSeek R1 | 69 | #473 |
| 3 | Gemini 1.5 Flash-8B (001) | 64.9 | #2331 |
| 4 | Gemini 1.5 Flash (002) | 64.5 | #1464 |
| 5 | Llama 3.1 8B Instruct | 61.9 | #4246 |
| 6 | GPT-4o (2024-08-06) | 54.6 | #740 |
| 7 | Llama 3 8B Instruct | 53.6 | #4612 |
| 8 | GPT-4o Mini (2024-07-18) | 48.2 | #1510 |
| 9 | DeepSeek V3 | 48 | #716 |
| 10 | Llama 3.2 3B Instruct | 46.6 | #7147 |
Interactive version: theaggregate.ai/benchmark?slug=realguardrails-distractors · How It Works · Data refreshed daily, snapshot 2026-11-13.