autotrust 发布 JEV-27B-VL：零样本看图决策的视觉模型
原标题：autotrust/JEV-27B-VL: a decision model that learned to see without a single image of training
AI 摘要
autotrust 发布 JEV-27B-VL，在 JEV-27B 基础上加入视觉能力，可同时接收图像与文本并在一次前向传播中输出每个选项的校准概率。其决策头训练时从未见过图像，却在 MicroLens 短视频封面推荐任务上达到 AUC 0.727，与基于 59,045 名用户行为训练的协同过滤（0.728）相当，且 top-5 命中率更高（59% vs 49%）。模型以 Apache-2.0 开放权重，支持 vLLM 单卡部署，并保留原文本决策能力。
正文节选
Today we are releasing autotrust/JEV-27B-VL, JEV-27B with vision. Show it images and text, ask a question, and it answers in one forward pass with a calibrated probability for every option. When a question needs more thought, the same model thinks it through step by step, and it can look at the images while it does. Its decision head never saw an image during training. It still makes image decisions well enough to compete with a recommender built from tens of thousands of users' behaviour. TL;DR