RHELM (RAG Top-20) - Aggregation: leaderboard
RHELM category (Zhang et al., Microsoft and Renmin University of China, 2026) that asks for conditional counts, trends, extreme values or events that never happened across many sessions. The model answers from the top 20 retrieved chunks of dialogue history and external documents; accuracy in percent on 192 questions judged 0 or 1 by an LLM, higher is better.
Metric: Accuracy (%; 192 aggregation questions, LLM judge, dialogue history plus external sources, dense top-20 retrieval). Source: arxiv.org. Saturation forecast: Around 2029. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 | 32.3 | #492 |
| 2 | Claude Opus 4.5 | 31.8 | #152 |
| 3 | GPT-4.1 Mini | 30.7 | #775 |
| 4 | Gemini 2.5 Pro | 26 | #262 |
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-aggregation · How It Works · Data refreshed daily, snapshot 2026-11-12.