RHELM (RAG Top-20) - Fact: leaderboard
RHELM category (Zhang et al., Microsoft and Renmin University of China, 2026) that asks for specific facts about the persona scattered through a simulated year of conversation, including multi-hop look-ups, entity disambiguation, attributes at a past state and exclusion constraints. The model answers from the top 20 retrieved chunks of dialogue history and external documents; accuracy in percent on 207 questions judged 0 or 1 by an LLM, higher is better.
Metric: Accuracy (%; 207 fact questions, LLM judge, dialogue history plus external sources, dense top-20 retrieval). Source: arxiv.org. Saturation forecast: Around September 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 | 50.2 | #492 |
| 2 | Claude Opus 4.5 | 50.2 | #152 |
| 3 | Gemini 2.5 Pro | 43 | #262 |
| 4 | GPT-4.1 Mini | 42.5 | #775 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=rhelm-rag-top-20-fact · How It Works · Data refreshed daily, snapshot 2026-11-12.