Facet-Probe - Document Rank Order: leaderboard
One ordering facet of Facet-Probe (Paruchuri, Koyejo and Adeli, 2026): the order of ranked retrieved documents in a RAG prompt is permuted on MultiHop-RAG and HotpotQA, and a flip is a change in the normalized short answer. The score is the share of items whose answer changes across six sampled orderings, in percent; lower is better.
Metric: Flip rate (% of items whose normalized answer changes across six sampled orderings of ranked retrieved documents (MultiHop-RAG, HotpotQA replay); lower is better). Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 8 | #91 |
| 2 | Claude Opus 4.7 (Thinking) | 11 | #75 (Claude Opus 4.7) |
| 3 | Gemini 3 Flash (Preview) | 13 | #141 |
| 4 | Qwen 3.5 4B | 23 | #1180 |
| 5 | GPT-5.5 (Medium) | 26 | #48 (GPT-5.5) |
| 6 | GPT-5.4 Mini (Medium) | 27 | #389 (GPT-5.4 Mini) |
| 7 | Qwen 3.5 9B | 28 | #803 |
| 8 | Claude Sonnet 4.6 (Thinking) | 33 | #158 (Claude Sonnet 4.6) |
| 9 | MedGemma-4B-IT | 34 | #2732 |
| 10 | Qwen 3.5 2B | 41 | #4849 |
| 11 | Qwen 3.5 0.8B | 41 | #8139 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=facet-probe-document-rank-order · How It Works · Data refreshed daily, snapshot 2026-11-12.