EditReward-Compass - Instruction Awareness: leaderboard

Accuracy (%) of multimodal LLMs used as reward models on the preference pairs that differ in how well the edit follows the instruction. 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 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 February 2027. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)83.24#99
2Qwen 3.5 35B A3B80.74#492
3Gemini 3 Flash80.42#175
4Qwen 3.6 27B79.61#351
5Qwen 3.6 35B A3B79.21#481
6Qwen 3.5 27B76.74#375
7Qwen 3.5 9B76.15#828
8Gemma 4 31B75.27#382
9GPT-4.174.71#505
10Gemma 4 26B A4B69.47#590
11Qwen 3 VL 32B67.63#622
12Gemma 3 27B59.09#1663
13Qwen 3 VL 8B56.46#917
14Qwen 3 VL 30B A3B56.33#784
15Qwen 3 VL 4B52.09#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-instruction-awareness · How It Works · Data refreshed daily, snapshot 2026-11-21.