LIT-RAGBench (Japanese) - Integration: leaderboard
Metric: Accuracy (%) on the 12 Integration questions (integrating information from multiple documents) among the 54 hand-written Japanese questions, judged correct or incorrect against the reference answer by GPT-4.1 (2025-04-14); the generator receives the question with relevant and irrelevant documents of about 512 tokens each, temperature 0 where supported and the maximum reasoning length for reasoning models; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 91.7 | #237 |
| 2 | GPT-4.1 Mini | 91.7 | #346 |
| 3 | GPT-5 Mini | 91.7 | #176 |
| 4 | Qwen 3 235B A22B 2507 Instruct | 91.7 | #291 |
| 5 | Qwen 3 235B A22B 2507 (Thinking) | 91.7 | #253 (Qwen 3 235B A22B 2507) |
| 6 | O4 Mini | 91.7 | #172 |
| 7 | GPT-5 | 83.3 | #91 |
| 8 | GPT-4.1 | 83.3 | #240 |
| 9 | O3 | 83.3 | #121 |
| 10 | Llama 3.3 70B Instruct | 75 | #520 |
| 11 | GPT-5 Nano | 75 | #415 |
| 12 | Claude Sonnet 4 | 75 | #194 |
| 13 | Gemini 2.5 Pro | 66.7 | #145 |
| 14 | Gemma 3 27B (IT) | 66.7 | #509 |
| 15 | Llama 3.1 8B Instruct | 16.7 | #1018 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=lit-ragbench-japanese-integration · How It Works · Data refreshed daily, snapshot 2026-10-11.