XHotpotQA (Gold Supporting Facts): leaderboard
XHotpotQA (2026) tests cross-lingual multi-hop question answering in which the question, bridge evidence and answer evidence may be in different languages. A one-shot reader receives the annotated supporting sentences of 7,405 validation items and is scored by answer F1 in percent under the legacy Hotpot normalization; higher is better.
Metric: Answer F1 (%; one-shot reader given the question and the annotated supporting sentences of the audited V1 validation set, 7,405 mixed-language HotpotQA-derived items, legacy Hotpot answer normalization). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 50.82 | #741 |
| 2 | Llama 3.1 405B Instruct | 48.32 | #1114 |
| 3 | GPT-4o Mini | 45.13 | #1627 |
| 4 | Gemma 2 27B (IT) | 39.54 | #1815 |
| 5 | Qwen 2 72B Instruct | 37.95 | #1489 |
| 6 | aya-23-35B | 36.92 | #2669 |
| 7 | Gemma 2 9B (IT) | 35.32 | #2275 |
| 8 | Llama 3.1 8B Instruct | 32.12 | #4277 |
| 9 | Qwen 2 7B Instruct | 28.12 | #3238 |
| 10 | aya-23-8B | 22.29 | #6778 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=xhotpotqa-gold-supporting-facts · How It Works · Data refreshed daily, snapshot 2026-11-12.