IRB1K (RAG) - English: leaderboard
Metric: Correctness (%, 0-100): share of answers judged correct (mean of two judges, GPT-4.1-mini and Qwen3-Next-80B, each labelling correct, incorrect or not attempted) on the 548 valid-premise questions whose supporting evidence is in English, of IRB1K's 1,000 short-answer factuality questions generated from citing sentences of 2024-2025 Wikipedia articles; RAG: the top-5 documents retrieved by text-embedding-3-small from the IRB1K corpus (512-token chunks, documents ranked by their best chunk) are given with their source and publication date; the prompt sets the current date to 29 September 2025, invites 'I don't know' and false-premise answers, and runs once; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 | 84.9 | #240 |
| 2 | GPT-5 (Medium) | 84.9 | #91 (GPT-5) |
| 3 | GPT-OSS-120B | 84.4 | #330 |
| 4 | GPT-5 Mini (Medium) | 83.1 | #176 (GPT-5 Mini) |
| 5 | GPT-4.1 Mini | 82.8 | #346 |
| 6 | DeepSeek R1 | 79.9 | #245 |
| 7 | Llama 3.3 70B Instruct | 78.6 | #520 |
| 8 | Llama 4 Scout | 76 | #646 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=irb1k-rag-english · How It Works · Data refreshed daily, snapshot 2026-10-11.