FrameBench - Frame Identification - Japanese: leaderboard
Frame identification companion task from FrameBench (arXiv 2609.03370v1, Table 3), Japanese: given a sentence, a target word and candidate frame names with definitions from Japanese FrameNet, the model selects the frame the word evokes. Accuracy in percent; higher is better. Evaluated by the FrameBench authors; reasoning and non-reasoning runs are separate rows.
Metric: Accuracy (%; Japanese FrameNet candidate frames). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 78.3 | #105 |
| 2 | Gemma 4 E4B (IT) (Thinking) | 76.8 | |
| 3 | GPT-5 Nano | 76.3 | #439 |
| 4 | llm-jp-4-8B-Thinking | 75 | #529 |
| 5 | Gemma 4 31B (IT) (Thinking) | 73.8 | #178 (Gemma 4 31B (IT)) |
| 6 | Gemma 4 E2B (IT) (Thinking) | 72.3 | #779 (Gemma 4 E2B (IT)) |
| 7 | Qwen 3.5 9B (Thinking) | 72 | #391 (Qwen 3.5 9B) |
| 8 | Gemini 3.1 Pro (Preview) | 71.5 | #64 |
| 9 | Gemma 4 31B (IT) (Non-reasoning) | 71.5 | #178 (Gemma 4 31B (IT)) |
| 10 | Qwen 3 32B (Thinking) | 71.3 | #449 (Qwen 3 32B) |
| 11 | Qwen 3.5 4B (Thinking) | 71 | #514 (Qwen 3.5 4B) |
| 12 | Qwen 3.5 27B (Thinking) | 70.8 | #192 (Qwen 3.5 27B) |
| 13 | Qwen 3 8B (Thinking) | 69 | #678 (Qwen 3 8B) |
| 14 | Gemma 4 E4B (IT) (Non-reasoning) | 68.5 | |
| 15 | Gemma 4 E2B (IT) (Non-reasoning) | 67.8 | #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-frame-identification-japanese · How It Works · Data refreshed daily, snapshot 2026-11-05.