A September 16 preprint reports that components helping AI models read text in pictures also reveal object concepts. The finding offers researchers a route for inspecting how visual information becomes language. [1]
Existing tools already make hidden image representations readable. LatentLens, first released in January, compares those representations with a stored collection of words represented in their surrounding sentences. The nearest matches provide descriptions of image regions. Its implementation lets researchers extract internal model states and search that collection without training another model. This supplies useful labels, but a label alone leaves a further question: which parts of the model actually contribute to producing its answer? [2] [3]
The new study starts with optical character recognition, or OCR: reading words from images. Researchers overlaid English words on photographs and ranked attention heads, components that route information, by their contribution to transcription. In a 100-image test, replacing the highest-ranked heads' outputs with averages damaged reading more than replacing equally many randomly selected late-layer heads. [1]
The selected heads then supplied the mathematical transformation for a verbalization lens, which turns internal image representations into vocabulary words. Directing the heads toward a bird's wing, for example, produced the word feathers. The study covered four models spanning Qwen, Molmo and LLaVA. [1]
These readable outputs still have limits. Object localization remained weak, and some interpretation and editing assessments used AI judges. Internal edits changed captioned objects while leaving the input photograph unchanged. Those experiments show a way to influence descriptions; they do not establish a general improvement in visual accuracy. [1]
For model researchers, our reading is that transcription provides a useful starting task for locating components to investigate. The removal test asks whether a component matters to behavior; decoding asks what information can be recovered from it. Keeping those questions separate prevents a plausible word label from becoming a complete explanation of recognition. A useful next test would combine independent human assessment with repeated interventions on unfamiliar images, checking both intended changes and damage to unrelated descriptions.