A family of 7B–72B reward models based on EditReward-Bench is presented as support for RL training and, according to the post, outperforms GPT-5 on that benchmark.
Published on October 25, 2025, the post introduces EditScore, a family of reward models ranging from 7B to 72B parameters and based on EditReward-Bench. Its stated aim is to evaluate and guide complex image edits and enable reinforcement learning (RL) training. The post says EditScore outperforms GPT-5 on the cited benchmark.
The result may interest engineers training image-editing systems, but the performance claim concerns that benchmark and may not generalize to other uses. Consult the original post and EditReward-Bench to verify the methodology, metrics, and comparison conditions; the supplied material gives no further details. If using AI to study or apply the method, avoid submitting images or documents containing personal data without authorization.