FrameBench - Japanese: leaderboard
FrameBench (arXiv 2609.03370v1, Table 3) Japanese set: multiple-choice questions testing whether a model distinguishes the semantic frames evoked by the same verb across contexts, built from Japanese FrameNet-style resources with native-speaker validation. Accuracy averaged over five prompt templates, in percent; higher is better. Reasoning and non-reasoning runs are separate rows.
Metric: Accuracy (%; mean of five prompt templates). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 4 31B (IT) (Thinking) | 99.1 | #178 (Gemma 4 31B (IT)) |
| 2 | Gemma 4 31B (IT) (Non-reasoning) | 97.6 | #178 (Gemma 4 31B (IT)) |
| 3 | Gemini 3.1 Pro (Preview) | 97.5 | #64 |
| 4 | Qwen 3.5 27B (Thinking) | 97.3 | #192 (Qwen 3.5 27B) |
| 5 | GPT-5 | 96.3 | #105 |
| 6 | Qwen 3 32B (Thinking) | 93.9 | #449 (Qwen 3 32B) |
| 7 | Gemma 4 E4B (IT) (Thinking) | 92.2 | |
| 8 | Qwen 3.5 9B (Thinking) | 91.8 | #391 (Qwen 3.5 9B) |
| 9 | Qwen 3.5 4B (Thinking) | 89.9 | #514 (Qwen 3.5 4B) |
| 10 | Qwen 3 8B (Thinking) | 85 | #678 (Qwen 3 8B) |
| 11 | Gemma 4 E4B (IT) (Non-reasoning) | 84.6 | |
| 12 | llm-jp-4-8B-Thinking | 84.5 | #529 |
| 13 | GPT-5 Nano | 83.9 | #439 |
| 14 | Qwen 3.5 27B (Non-reasoning) | 83.7 | #192 (Qwen 3.5 27B) |
| 15 | Gemma 4 E2B (IT) (Thinking) | 82.3 | #779 (Gemma 4 E2B (IT)) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=framebench-japanese · How It Works · Data refreshed daily, snapshot 2026-11-05.