Tsinghua Finds AI Image Generation Secretly Teaches Models Textual Facts
A new probing protocol shows that teaching unified multimodal models to draw facts partially teaches them to say those facts in words, with an alignment trick closing the gap.
- New paper introduces Knowledge Transfer Probing (KTP) to test if image-generation training teaches text-answerable facts. Paper
- Tests three unified models: BAGEL, SenseNova-U1, and Z.AI's new GLM-Omni built on GLM-4.1V-Thinking.
- Direct text-to-image training does transfer some knowledge to text, but unevenly across seven categories.
- Models reproduce facts in images far more reliably than they can state them in words.
- Interleaved Image-Text Alignment on a small entity subset generalizes mapping skill to unseen entities.
- Project page with results and ablations at kevinlllee.github.io/Knowledge-Transfer-Probing.
Can Models Say What They Learn to Draw?
Researchers from Tsinghua University and Z.AI tested whether a multimodal model can answer questions about facts it encountered only while learning to generate images. Their paper finds that some visual knowledge transfers into text responses, but access varies sharply by model and knowledge category.
Unified multimodal models, or UMMs, use one system to understand and generate text and images. They train on large image collections whose captions often omit details such as object counts, spatial relationships, shapes, and embedded text. If those details remain accessible only through image generation, developers cannot reliably use them in text-based applications.
A Controlled Test for Cross-Modal Recall
The researchers created Knowledge Transfer Probing (KTP), a protocol built around fictional entities and attributes. Fictional associations reduce the chance that a model already knows an answer from pretraining.
KTP covers seven categories:
- Color
- Count
- Number
- Shape
- Size
- Spatial relationships
- Text displayed inside images
The protocol separates exposure from retrieval so researchers can identify the source of each answer:
| Stage | Model input | Training target | What it measures |
|---|---|---|---|
| Exposure | A text prompt naming an entity and scene | An image containing the attribute | Whether generation training encodes the association |
| Evaluation | A text-only question about the entity | The attribute expressed in text | Whether the association transfers into language |
Attribute values appear only in target images during the initial training stage. A correct text answer therefore indicates that the model extracted a fact from the image-generation objective and made it available to its language pathway.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
