RHELM (RAG Top-20) - Misleading: leaderboard
RHELM category (Zhang et al., Microsoft and Renmin University of China, 2026) that makes requests that implicitly conflict with how the persona's life has changed (an injury, a move, a new job), which the assistant must notice, decline and answer with a compatible alternative. The model answers from the top 20 retrieved chunks of dialogue history and external documents; accuracy in percent on 65 questions judged 0 or 1 by an LLM, higher is better.
Metric: Accuracy (%; 65 misleading questions, LLM judge, dialogue history plus external sources, dense top-20 retrieval). Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 41.5 | #152 |
| 2 | Gemini 2.5 Pro | 26.2 | #262 |
| 3 | GPT-4.1 | 6.2 | #492 |
| 4 | GPT-4.1 Mini | 3.1 | #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-misleading · How It Works · Data refreshed daily, snapshot 2026-11-12.