Researchers from Google DeepMind and Stanford University published a paper on October 27, 2026, titled A Heart Means Recursion: Exploring Visual Meaning and Communication in VLMs. The study examines how unrelated images attached to prompts influence the coding style of vision-language models.
During tests involving binary tree traversal tasks, researchers attached various colored shapes to prompts. For models Claude Opus 4.7 and GPT-5.4, the inclusion of a red heart image increased the use of recursive code solutions from 0% to 100%. Gemini 3.1 Pro showed a 90% recursion rate when prompted with a yellow checkmark.
The team evaluated 288 different shape-color combinations, finding that each model possessed a unique steering profile. The study noted that text captions describing the symbols or images of random Gaussian noise had minimal impact on model behavior, suggesting the effect is specific to the visual input itself.
The research highlights a transparency issue, as models often failed to mention the visual cues in their reasoning traces. This suggests that current methods for auditing AI logic through self-explanation may have significant blind spots regarding how visual data influences output.
