EditReward-Compass - Instruction Awareness: leaderboard
Metric: Preference accuracy (%) on the pairs that differ in instruction awareness on EditReward-Compass (2,251 human-verified preference pairs of edited images sampled from the same editing models under the same instruction), general MLLMs prompted with the Edit-Compass rubric; thinking disabled unless the row is a thinking run; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 24 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 83.24 |
| 2 | Qwen 3.5 35B A3B | 80.74 |
| 3 | Gemini 3 Flash | 80.42 |
| 4 | Qwen 3.6 27B | 79.61 |
| 5 | Qwen 3.6 35B A3B | 79.21 |
| 6 | Qwen 3.5 27B | 76.74 |
| 7 | Qwen 3.5 9B | 76.15 |
| 8 | Gemma 4 31B | 75.27 |
| 9 | GPT-4.1 | 74.71 |
| 10 | Qwen 3.5 27B (Non-reasoning) | 73.22 |
| 11 | Qwen 3.5 35B A3B (Non-reasoning) | 72.79 |
| 12 | Qwen 3.6 27B (Non-reasoning) | 71.47 |
| 13 | Gemma 4 26B A4B | 69.47 |
| 14 | Qwen 3.6 35B A3B (Non-reasoning) | 68.24 |
| 15 | Qwen 3.5 9B (Non-reasoning) | 66.82 |
Interactive version: theaggregate.ai/benchmark?slug=editreward-compass-instruction-awareness · How It Works · Data refreshed daily, snapshot 2026-10-07.