RAI-Bench - RAG Robustness (LC 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 long in-context setting, where the rest of the knowledge base sits in the prompt. 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 | Claude Haiku 4.5 | 97 | #280 |
| 2 | Claude Sonnet 4 | 97 | #199 |
| 3 | Claude Sonnet 4.5 | 96 | #142 |
| 4 | DeepSeek V4 Pro | 95 | #102 |
| 5 | Grok 4.1 Fast | 95 | #217 |
| 6 | DeepSeek V4 Flash | 93 | #137 |
| 7 | Grok 4.3 | 93 | #188 |
| 8 | GPT-5.4 Nano | 92 | #328 |
| 9 | GPT-5.2 | 90 | #110 |
| 10 | DeepSeek V3.2 | 89 | #205 |
| 11 | Claude Opus 4.5 | 89 | #83 |
| 12 | Grok 4.1 Fast (Reasoning) | 89 | #217 (Grok 4.1 Fast) |
| 13 | Claude Opus 5 | 89 | #12 |
| 14 | GPT-5 Chat | 89 | #241 |
| 15 | Gemini 3.5 Flash | 88 | #60 |
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-lc-abstention · How It Works · Data refreshed daily, snapshot 2026-10-20.