Zhipu AI and Tsinghua researchers found that training a small multimodal model to generate and inspect images could improve its answers to text questions. The result offers a concrete research lead for GLM’s next generation, not proof that larger models or real-world tasks will see the same gains.
Zhipu AI / Z.ai GLM Watch analysis
What happened
The researchers tested GLM-Omni, a 9-billion-parameter model based on GLM-4.1V-Thinking, in a fictional world designed to test whether knowledge learned through image generation could transfer to language. In one example, the model learned to draw a made-up team’s jersey number without being given that number in text.
The paper, covered by 36Kr, reports that GLM-Omni’s average text-answer accuracy rose from 12.57% to 26.86% after the researchers trained image generation and text answers in an interleaved sequence. When the model could generate an image, inspect it and then answer, accuracy reached 41.14%. Removing the image-generation training signal in a control test brought the score down to 13.71%.
Key findings
- Drawing knowledge did not automatically become language knowledge
GLM-Omni reached 54% average accuracy on image-generation tasks, but 14.57% on text questions about the same fictional knowledge. - Interleaved training improved text answers
Combining image generation and question-answering in one training sequence raised average accuracy to 26.86%. - Images also helped at answer time
Generating and inspecting an image before responding lifted average accuracy to 41.14% in this test.
Why it matters
The findings suggest image generation may be more than an output feature: in this experiment, it helped the model learn or retrieve information for language answers. That makes training design, not just a model’s architecture or list of supported inputs, part of the multimodality story.
The evidence has clear limits. GLM-Omni is a small experimental model, and the researchers used a purpose-built fictional dataset. These results do not establish that the method improves general reasoning or will carry over to deployed models.
Our read
This is a useful research clue, rather than a preview of a confirmed GLM product feature. The control result gives the paper’s central claim some experimental bite, while the narrow test means bigger promises should wait their turn. If you follow multimodal models, watch for replications on broader tasks, not just a better score in a carefully designed toy world.
What to watch
- Whether other researchers reproduce the results with different models and datasets.
- Whether the approach improves performance on practical, less constrained tasks.
- Whether Zhipu incorporates the method into a future GLM model.
Discussion spark: Should multimodal models be encouraged to generate images as part of reasoning, even when the user only wants a text answer?
Sources and evidence
- Tang Jie Signed Paper Unveils Zhipu AI’s In-House Multimodal Model: Is Next-Gen GLM Nearing Launch? – 36Kr (26 October 2026, 00:44 UTC)
not affiliated with or endorsed by Zhipu AI, Z.ai or GLM