EditReward-Compass - Visual Quality: leaderboard

Accuracy (%) of multimodal LLMs used as reward models on the preference pairs that differ in the visual quality of the edit. 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: Preference accuracy (%) on the pairs that differ in visual quality 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 2028. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.5 27B58.78#380
2Qwen 3.6 27B57.43#358
3Qwen 3.5 35B A3B56.08#494
4Qwen 3.6 35B A3B56.08#485
5GPT-4.153.38#506
6Gemma 4 31B49.32#390
7Qwen 3.5 9B48.98#838
8Gemini 3 Flash48.65#180
9Qwen 3.5 2B44.66#5380
10Gemini 3.1 Pro (Preview)44.59#102
11Gemma 4 26B A4B43.92#602
12Qwen 3 VL 32B36.49#624
13Qwen 3 VL 8B34.46#927
14Qwen 3 VL 4B33.78#1163
15Qwen 3 VL 30B A3B33.78#791

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=editreward-compass-visual-quality · How It Works · Data refreshed daily, snapshot 2026-11-22.