RAI-Bench - RAG Robustness (HY Abstention): leaderboard
GovTech Singapore's RAI-Bench asks 331 questions on four Singapore government policy documents, each with its own answer left out of the supplied context, so the right reply is to abstain. This board is the HyDE setting, where retrieval follows a hypothetical answer. An LLM judge detects abstention, and the number is the share of questions on which the model abstained, so higher is better.
Metric: Rate (%). Source: huggingface.co. Saturation forecast: Not forecast. 69 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash Lite | 93 | #422 |
| 2 | DeepSeek V4 Pro | 91 | #102 |
| 3 | DeepSeek V4 Flash | 91 | #137 |
| 4 | Grok 4.3 | 91 | #188 |
| 5 | Claude Sonnet 4.5 | 90 | #142 |
| 6 | GPT-5.4 Mini | 90 | #212 |
| 7 | Gemini 3.5 Flash | 90 | #60 |
| 8 | Gemini 3.6 Flash | 90 | #51 |
| 9 | Claude Sonnet 4 | 89 | #199 |
| 10 | GLM-5 | 89 | #141 |
| 11 | Gemini 3.5 Flash Lite | 89 | #156 |
| 12 | Grok 4 | 88 | #181 |
| 13 | Grok 4.1 Fast (Reasoning) | 88 | #217 (Grok 4.1 Fast) |
| 14 | Claude Opus 5 | 88 | #12 |
| 15 | GLM-4.6 | 88 | #252 |
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-rag-robustness-hy-abstention · How It Works · Data refreshed daily, snapshot 2026-10-20.