Put a heart next to a coding problem, and Claude and GPT write recursive code

A study of vision-language models examines whether simple images can shape how they solve problems and communicate. Adding a heart beside a coding prompt led Claude and GPT to produce recursive code in every tested case, compared with none without the symbol. Simple icons such as hearts and triangles steered behavior about as effectively as direct verbal instructions, while captions describing the images, visual noise, and ordinary photographs had little effect. The influence extended beyond coding. Yet models’ reasoning traces often ignored the image or called it irrelevant, and their explanations of what symbols meant did not reliably match the behavior they triggered: Claude, for example, produced recursion in response to a triangle but described the triangle as indicating iteration.

The research also tested models’ ability to choose images to influence another model. They initially favored familiar symbols such as spirals and arrows, but effective symbols emerged through repeated exchanges, sometimes becoming less intuitive to people; human readers identified the intended meaning in 67% of first attempts but only 44% of refined messages. When models could draw freely using code, they conveyed concepts and short stories successfully on 91% of first tries; with icons alone, success fell to 48%. The findings suggest models can acquire visual associations they cannot explain, complicating oversight based on verbal reasoning. Although the results do not point to a fully developed private visual language, model-to-model communication may become effective while growing harder for humans to interpret.