S-RuLES (System-Prompt RuLES): leaderboard
This paper's variant of the RuLES rule-following scenarios: the rules sit in the system prompt with no instruction to prioritise it, and users try to make the model break them; the number is the mean pass rate (%) of the basic and redteam suites. It is not the original RuLES score: this paper moves the rules into the system prompt and removes the explicit instruction to prioritise system messages.
Metric: Basic and redteam pass rate, rules in the system prompt (%, higher is better). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 Mini | 93.5 | #533 |
| 2 | GPT-4o (2024-08-06) | 91.7 | #740 |
| 3 | GPT-4o Mini (2024-07-18) | 80.2 | #1510 |
| 4 | Gemini 1.5 Flash (002) | 67 | #1464 |
| 5 | Gemini 1.5 Flash-8B (001) | 59.7 | #2331 |
| 6 | DeepSeek R1 | 56.6 | #473 |
| 7 | Llama 3.1 8B Instruct | 54.9 | #4246 |
| 8 | Llama 3 8B Instruct | 52.5 | #4612 |
| 9 | Llama 3.2 3B Instruct | 52.2 | #7147 |
| 10 | DeepSeek V3 | 47.1 | #716 |
Interactive version: theaggregate.ai/benchmark?slug=s-rules-system-prompt-rules · How It Works · Data refreshed daily, snapshot 2026-11-13.