RHELM (RAG Top-20): leaderboard
Long-term memory benchmark for personal assistants (RHELM, Zhang et al., Microsoft and Renmin University of China, 2026): ten synthetic personas, each with a simulated year of dialogue plus emails, journals and reports, and 1,305 questions in seven categories that an LLM judge scores 0 or 1 against a reference answer. The model answers from the top 20 chunks of dialogue and external documents retrieved by a bge-large dense index. The score is accuracy in percent over all 1,305 questions; higher is better.
Metric: Accuracy (%; question-weighted over 1,305 questions in seven categories, LLM judge, dialogue history plus external sources, dense top-20 retrieval). Source: arxiv.org. Saturation forecast: Around July 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 38.1 | #152 |
| 2 | Gemini 2.5 Pro | 35.5 | #262 |
| 3 | GPT-4.1 | 29.5 | #492 |
| 4 | GPT-4.1 Mini | 25.1 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-12.