The Information Bottleneck of Vision Tokens: How Much Can One Token Hold?

In VLMs (Vision-Language Models), an image is first split into patches, and each patch is mapped to a vision token. This token is the model's "window" for seeing the image.

But here is a question: how much information can a single token actually hold? As image resolution increases and tasks become more complex, how many tokens do you need? Nobody had answered this quantitatively before.

In joint work with Shuxin Zhuang, we found that the information capacity of vision tokens follows a scaling law—a quantitative power-law relationship with image resolution, patch size, and task complexity.

Specifically:

• Higher resolution increases per-token information (with diminishing returns)
• Smaller patches (more tokens) increase total information
• More complex tasks (e.g., fine-grained recognition) demand higher per-token information

This scaling law directly informs VLM design. For instance, if you want the model to recognize small objects in images, our formula tells you exactly what resolution you need—given a specific patch size—for each token to carry sufficient discriminative information. No more trial and error.

• Title: How Much Information Can a Vision Token Hold? A Scaling Law for Recognition Limits in VLMs
• Authors: Shuxin Zhuang, Zi Liang, Runsheng Yu, Hongzong Li, Rong Feng, Shiqin Tang, Youzhi Zhang
• Status: Preprint 2026