GLM-Omni Research Tests Knowledge Transfer From Images to Language
A joint Tsinghua University and Zhipu AI paper disclosed the previously unreleased internal model GLM-Omni, with Zhipu chief scientist Tang Jie as corresponding author. In the MagicWorld experiments, GLM-Omni achieved 54% average visual-generation accuracy but only 14.57% on text answers; interleaved image-text alignment raised text accuracy to 26.86%, while Visual Chain-of-Thought further raised it to 41.14%. The paper did not establish which product version, if any, corresponds to GLM-Omni.
The paper provides concrete evidence that visual generation training does not automatically make knowledge available to language responses, while joint image-text training and generating images before answering improved performance in controlled experiments.
导语:图像生成正确率 54%,文字问答仅 14.57%。

图像生成正确率 54%,文字问答仅 14.57%。
作者丨郑佳美
编辑丨岑 峰
智谱内部的 GLM-Omni,先在一篇论文里露了面。
#Tsinghua University#Zhipu AI#GLM-Omni#GLM-4.1V-Thinking#X-Omni#MagicWorld#BAGEL#ByteDance#SenseNova-U1#SenseTime