FrameBench - English: leaderboard
FrameBench (arXiv 2609.03370v1, Table 3) English set: multiple-choice questions testing whether a model distinguishes the semantic frames evoked by the same verb across contexts, built from 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. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 99.3 | #64 |
| 2 | GPT-5 | 99.1 | #105 |
| 3 | Qwen 3.5 27B (Thinking) | 98.6 | #192 (Qwen 3.5 27B) |
| 4 | Gemma 4 31B (IT) (Thinking) | 98.6 | #178 (Gemma 4 31B (IT)) |
| 5 | Gemma 4 31B (IT) (Non-reasoning) | 97.5 | #178 (Gemma 4 31B (IT)) |
| 6 | Qwen 3.5 9B (Thinking) | 97.3 | #391 (Qwen 3.5 9B) |
| 7 | Qwen 3.5 27B (Non-reasoning) | 96.1 | #192 (Qwen 3.5 27B) |
| 8 | Qwen 3.5 4B (Thinking) | 95.5 | #514 (Qwen 3.5 4B) |
| 9 | Gemma 4 E4B (IT) (Thinking) | 94.5 | |
| 10 | GPT-5 Nano | 93.5 | #439 |
| 11 | Gemma 4 E2B (IT) (Thinking) | 86.3 | #779 (Gemma 4 E2B (IT)) |
| 12 | Qwen 3.5 9B (Non-reasoning) | 85.8 | #391 (Qwen 3.5 9B) |
| 13 | Gemma 4 E4B (IT) (Non-reasoning) | 82.5 | |
| 14 | Qwen 3.5 4B (Non-reasoning) | 81.1 | #514 (Qwen 3.5 4B) |
| 15 | Qwen 3.5 2B (Thinking) | 80.3 | #1114 (Qwen 3.5 2B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=framebench-english · How It Works · Data refreshed daily, snapshot 2026-11-05.