Somewhere in the long effort to build machines that understand the world through many senses at once, a small and quiet discovery has surfaced: a model can paint a fact it cannot yet speak. Researchers at Tsinghua and Zhipu, working with an internal 9-billion-parameter model called GLM-Omni, tested this gap inside an invented world where no prior text could leak the answers. Their findings suggest that learning to create images and learning to explain in words are not automatically the same kind of learning, and that letting a model look at what it has drawn may help it recall what it knows. W
Zhipu Paper Finds Image Generation Can Teach Models Knowledge They Can't Yet Express in Text
A model can draw what it cannot say
Mark
So what is the headline finding here?
Mimi
That a model can show knowledge in a picture that it can't put into words. GLM-Omni got 72% on numerical image tasks and 2% in text.
Mark
Why invent a fictional world for the test?
Mimi
So the model couldn't lean on anything it had read. MagicWorld's teams and badges don't exist, so any knowledge had to come from the training images.
Luke
Worth noting, though, that every number comes from that fictional dataset. It tells us about the mechanism, not about real-world performance.
Mark
And the fix was to let the model draw before answering?
Mimi
Roughly. Interleaved training lifted text accuracy from 12.57% to 26.86%, and visual chain-of-thought took it to 41.14%.
Luke
The control matters too. Removing the image loss dropped accuracy to 13.71%, which points to the image signal as the driver. But this is one 9-billion-parameter model.
Mark
Does this mean the next GLM will work this way?
Luke
No. The reporting says the paper gives no direct evidence of that. GLM-6.0 plans mention multimodal unification, but nothing confirms these methods.
Mimi
And the cost question is unresolved. Specialized models are cheaper to tune, which is why OpenAI and Google split their image models off.
The Pulse
- A model drew fictional jersey numbers correctly 54% of the time yet could state them in text only 14.57% of the time, exposing a stubborn gap between what it shows and what it says.
- On numerical questions the split became stark, with 72% accuracy in images against just 2% in text, and rival models from ByteDance and SenseTime showed the same disconnect.
- Researchers tried to close the gap by training generation and answering together, lifting text accuracy from 12.57% to 26.86%, and a control run without the image signal fell back to 13.71%.
- A visual chain-of-thought, drawing an image and then reasoning over it before answering, pushed average text accuracy to 41.14% and numerical accuracy from 0% to 36%.
- The gains rest on a fictional dataset, and the paper has not shown they justify unified models over specialized ones, so the field waits to see whether GLM-6.0 adopts these methods.
Somewhere in the long effort to build machines that understand the world through many senses at once, a small and quiet discovery has surfaced: a model can paint a fact it cannot yet speak. Researchers at Tsinghua and Zhipu, working with an internal 9-billion-parameter model called GLM-Omni, tested this gap inside an invented world where no prior text could leak the answers. Their findings suggest that learning to create images and learning to explain in words are not automatically the same kind of learning, and that letting a model look at what it has drawn may help it recall what it knows. Whether this promise is worth the cost of bundling these abilities into one architecture remains an open question that the field has not yet answered.
A model asked to draw a fictional team's jersey number usually gets it right. Asked to simply say that number in words, it fails. That puzzling gap is the heart of a new arXiv paper from researchers at Tsinghua University and Zhipu, with Tang Jie, Zhipu's chief scientist and a Tsinghua professor, as corresponding author. The paper discloses GLM-Omni, an internal model built on Zhipu's open-source GLM-4.1V-Thinking, and introduces a method called Knowledge Transfer Probing to separate what a model learns while generating images from what it can express in language.
To keep the model from simply repeating text it had already absorbed, the team built MagicWorld, a world with no real teams, characters or badges. Across seven knowledge categories, including color, quantity, shape and spatial position, the results showed what the authors call "can draw but can't say." GLM-Omni averaged 54% accuracy in images but only 14.57% in text, and on numerical tasks the gap was 72% against 2%. Rival models from ByteDance and SenseTime showed similar patterns.
The team then tested two training approaches. Simple text alignment proved unstable. Interleaved image-text training, which placed drawing and answering in one sequence, raised text accuracy to 26.86%, while a control without the image signal fell back to 13.71%. A further step, Visual Chain-of-Thought, had the model generate an image, observe it, and then answer, lifting average text accuracy to 41.14% and numerical accuracy from 0% to 36%.
The authors conclude that knowledge sharing depends on whether training signals across tasks are connected, not merely on sharing an architecture, and that image generation may help both learning and recall. The cost question stays open. Major companies such as OpenAI and Google have split image work into specialized models, and Tang Jie has said separate development may still be most effective. The paper does not show its gains outweigh the cost of unification, and it offers no direct evidence of GLM-6.0's architecture. Zhipu has said GLM-6.0 will explore self-evaluation and native multimodal unified modeling, and whether it adopts these methods is the thing to watch.
A model asked to draw a fictional team's jersey number gets it right most of the time. Asked to simply state that number in words, it fails. That gap sits at the center of a new paper posted to arXiv by researchers from Tsinghua University and Zhipu, titled "Can Visual Generation Expand What Unified Multimodal Models Know?" Tang Jie, Zhipu's chief scientist and a Tsinghua professor, is the corresponding author.
The paper discloses a Zhipu internal model not previously made public, GLM-Omni. It is built on Zhipu's open-source GLM-4.1V-Thinking and has 9 billion parameters, which makes it a relatively small test model. The authors propose a method called Knowledge Transfer Probing, designed to separate two things: whether a model learned a piece of knowledge while generating images, and whether it can express that knowledge in language.
To make sure the model was not answering from text it had already absorbed, the team invented a world called MagicWorld, where teams, characters and badges do not exist. The dataset spans seven categories of knowledge, including color, quantity, number, shape and spatial position. It holds 350 entries, expanded into 8,552 image training samples.
The first results showed what the researchers describe as "can draw but can't say." GLM-Omni averaged 54% accuracy in image generation but only 14.57% when answering the same knowledge in text. On numerical tasks the split was starker: 72% in images, 2% in text. ByteDance's BAGEL and SenseTime's SenseNova-U1 were also tested and showed similar patterns, with SenseNova-U1 reaching 88% image accuracy and 0% text accuracy on numerical tasks. Housing generation and language in one model, the results suggest, does not make what each learns interchangeable.
The team then tried two training approaches. Text alignment, which added text question-answering for some fictional entities to see whether it would carry over to others, proved unstable; GLM-Omni's average text accuracy slipped from 14.57% to 12.57%. The second approach, Interleaved Image–Text Alignment, placed generation and answering in one training sequence: faced with a jersey-number question, the model first generates an image, looks at it, then answers in text. With drawing not allowed at test time, average accuracy rose from 12.57% to 26.86%. A control that kept the same samples and text supervision but removed the image-generation loss signal fell back to 13.71%.
A third round changed the reasoning at answer time, using a Visual Chain-of-Thought: generate an image, observe it, then compose the text answer. Average text accuracy climbed from 26.86% to 41.14%, a gain of 14.28 percentage points, and numerical tasks went from 0% to 36%. The authors draw two conclusions. Knowledge sharing in a unified model depends not only on a unified architecture but on whether the training signals of different tasks are connected. And image generation may help a model learn during training and also serve as an intermediate representation for recalling knowledge during inference.
The cost question remains open. In March 2025, OpenAI built native image generation into GPT-4o, stressing it should be a core language-model capability, but later products moved image work to specialized models such as GPT Image. Google similarly divides labor between Gemini and the Nano Banana image model. Specialized models are easier to optimize separately for quality, speed and cost. Zhipu itself uses GLM-5.3-Flash for multimodal understanding and GLM-Image for generation. In December 2025, Tang Jie said developing text, multimodal understanding and multimodal generation separately may still be most effective, while combining them is worth exploring but requires sufficient capital. The paper has not shown that its gains outweigh the cost of unification.
The study offers no direct evidence of the architecture of the next GLM. Zhipu's latest financing covers the next-generation GLM, fully self-trained models and computing infrastructure. At an August 31 earnings briefing, Tang Jie said GLM-6.0 will explore self-evaluation and continuous improvement, and Zhipu listed native context length and "native multimodal unified modeling capability" among its aims. The paper's co-first authors, Li Wenkai, Hong Wenyi and Yang Zhuoyi, did the work during internships at Zhipu. All figures come from the public paper and rest on the fictional MagicWorld dataset; whether GLM-6.0 adopts these methods is the thing to watch.
Notable Quotes
For text, multimodal understanding and multimodal generation, the most effective approach may still be to develop them separately.— Tang Jie, Zhipu chief scientist, December 2025
Want the full story?
Read the original at finance.biggo.com ↗
Science & Space