EditReward-Compass: leaderboard

Accuracy (%) of multimodal LLMs used as image-editing reward models at picking the preferred edit of each pair. EditReward-Compass (2026) holds 2,251 human-verified preference pairs of edited images produced by the same editing models under one instruction, as a reward model meets them during RL; higher is better.

Metric: Average preference accuracy (%) over all pairs 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 February 2027. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)74.33#99
2Gemini 3 Flash72.68#175
3Qwen 3.6 27B71.83#351
4Qwen 3.5 35B A3B70.89#492
5Qwen 3.6 35B A3B70.51#481
6Qwen 3.5 27B69.98#375
7Gemma 4 31B67.09#382
8Qwen 3.5 9B66.81#828
9GPT-4.166.11#505
10Gemma 4 26B A4B59.6#590
11Qwen 3 VL 32B57.9#622
12Gemma 3 27B49.96#1663
13Qwen 3 VL 30B A3B47.87#784
14Qwen 3 VL 8B46.5#917
15Qwen 3 VL 4B43.22#1151

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=editreward-compass · How It Works · Data refreshed daily, snapshot 2026-11-21.