Stanford Finds a Red Heart Secretly Steers AI Coding Decisions
A new study shows that pasting a red heart next to a coding prompt flips frontier VLMs toward recursion, revealing a hidden visual channel their own reasoning cannot explain.
- Attaching a red heart image steers Claude, Gemini, and GPT toward recursive code solutions on binary tree traversal.
- Models' thinking traces deny any image influence, yet behavior shifts dramatically.
- Effect sweeps a 288-symbol library of 12 colors by 24 shapes with model-specific hotspots.
- Text captions describing the same symbols do not reproduce the steering effect.
- Pixel-shuffling and rotation disrupt steering, showing shape geometry drives the signal.
- Sender-receiver VLM games reveal a generation-interpretation gap that drifts from human visual conventions.
Attach a tiny red heart to an otherwise identical coding prompt, and a frontier vision-language model can become more likely to produce a recursive solution. Remove the image, and the model returns to loops. When asked to explain its choice, the model usually cites coding considerations while treating the heart as irrelevant.
A preliminary preprint, A Heart Means Recursion: Exploring Visual Meaning and Communication in VLMs, argues that latent visual associations can steer outputs on unrelated tasks without appearing in the model’s written rationale. Stanford, Google DeepMind, and Harvard researchers Alexa Tartaglini, Zhengxuan Wu, Thaddäus Wiedemer, Martin Wattenberg, Been Kim, and Yuxuan Li tested Claude Opus 4.7, Gemini 3.1 Pro, and GPT-5.4.
A Binary Tree Changes Course
The main experiment uses inorder traversal of a binary tree, which visits the left subtree, the current node, and then the right subtree. Developers can implement it recursively with the call stack or iteratively with an explicit stack. The tested models usually chose iteration when given a plain prompt.
Researchers then attached one of 288 symbols created by combining 12 colors with 24 shapes. They measured how each symbol changed the probability of receiving a recursive answer. A red heart produced the strongest reported shift, while pink crosses and light-blue spirals also influenced some models.
Each model responded to a different pattern of color-shape combinations. The resulting heatmaps contain distinct regions associated with recursion or iteration, indicating that the effect depends on both the model and the visual stimulus.
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
