RAI-Bench - Refusal Rate (Physics): leaderboard
GovTech Singapore's RAI-Bench runs adversarial prompts from RabakBench in four Singapore deployment scenarios, fifty per scenario. This board is the physics tutoring chatbot scenario, where users ask for physics knowledge that could be misused, for example to improvise weapons. An LLM judge labels each reply Reject, Reframe or Engage, and the number is the share labelled Reject or Reframe, so higher is better.
Metric: Rate (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 77 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 100 | #152 |
| 2 | Gemini 3.1 Pro (Preview) | 100 | #58 |
| 3 | GPT-5 | 100 | #98 |
| 4 | Gemini 2.5 Flash | 100 | #245 |
| 5 | Claude Sonnet 4.6 | 100 | #90 |
| 6 | Claude Sonnet 4.5 | 100 | #142 |
| 7 | GPT-5.4 | 100 | #79 |
| 8 | GPT-5.2 | 100 | #110 |
| 9 | GPT-5.5 | 100 | #31 |
| 10 | Claude Opus 4.6 | 100 | #63 |
| 11 | GPT-OSS-120B | 100 | #341 |
| 12 | Claude Haiku 4.5 | 100 | #280 |
| 13 | Kimi K2.5 | 100 | #144 |
| 14 | DeepSeek V3.2 | 100 | #205 |
| 15 | GPT-5 Nano | 100 | #425 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rai-bench-refusal-rate-physics · How It Works · Data refreshed daily, snapshot 2026-10-20.