RAI-Bench - RAG Robustness (HY Factuality): 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. This board is the HyDE setting, where retrieval follows a hypothetical answer. Replies that do not abstain are graded by an LLM judge on a three-tier factuality rubric, and the number is the share judged factual, so higher is better.
Metric: Rate (%). Source: huggingface.co. Saturation forecast: Not forecast. 69 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 49 | #280 |
| 2 | Grok 4 | 49 | #181 |
| 3 | O3 | 49 | #124 |
| 4 | Claude Opus 5 | 48 | #12 |
| 5 | Grok 4.3 | 48 | #188 |
| 6 | O4 Mini | 47 | #180 |
| 7 | Claude Opus 4.7 | 45 | #49 |
| 8 | Gemini 2.5 Flash Lite | 45 | #422 |
| 9 | Grok 4.1 Fast (Reasoning) | 44 | #217 (Grok 4.1 Fast) |
| 10 | DeepSeek V4 Pro | 43 | #102 |
| 11 | Claude Sonnet 5 | 43 | #57 |
| 12 | GLM-5 | 42 | #141 |
| 13 | Kimi K3 | 42 | #30 |
| 14 | GPT-5.6 Luna | 42 | #68 |
| 15 | Claude Sonnet 4.5 | 41 | #142 |
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-factuality · How It Works · Data refreshed daily, snapshot 2026-10-20.