MTRAG-UN (RAG) - LLM Judge: leaderboard
Metric: reference-based LLM-judge score RB_llm (GPT-OSS-120B judge in place of MTRAG's GPT-4o-mini; GPT-OSS-120B is also a subject) (%) with the top 5 passages retrieved by Elser after query rewriting, on MTRAG-UN's 666 human-written conversation tasks (unanswerable, underspecified, non-standalone and clarification turns over six corpora); metrics follow MTRAG and are conditioned on answerability with an IDK judge; printed as fractions and shown times 100; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B | 65 | #330 |
| 2 | GPT-OSS-20B | 63 | #499 |
| 3 | Mistral Large 3 | 60 | #388 |
| 4 | Qwen 3 30B A3B | 60 | #488 |
| 5 | Qwen 3 8B | 58 | #667 |
| 6 | DeepSeek V3 | 58 | #312 |
| 7 | Phi-4 | 57 | #701 |
| 8 | Llama 4 Maverick Instruct | 57 | #439 |
| 9 | Llama 3.3 70B Instruct | 55 | #520 |
| 10 | Granite 4.0 H Small | 53 | #657 |
| 11 | DeepSeek R1 | 52 | #245 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=mtrag-un-rag-llm-judge · How It Works · Data refreshed daily, snapshot 2026-10-11.