How Much Information Can a Vision Token Hold? A Scaling Law for Recognition Limits in VLMs

2026-01-28

AI Summary

This study investigates the theoretical and practical limits of visual tokens in vision-language models (VLMs) for encoding textual information. Through a stress test that incrementally increases the number of characters embedded in images, the authors uncover a three-phase transition behavior—stable, transitional, and collapse—in token recognition performance as information density rises. They propose, for the first time, a universal probabilistic scaling law that integrates average token load with visual density, demonstrating its validity across multiple mainstream VLMs. This work provides both theoretical grounding and empirical guidance for compressing high-density visual contexts, clearly delineating the critical trade-off boundary between efficiency and accuracy in visual language understanding.

Technology Category

• Computer Vision: Language and Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Recent vision-centric approaches have made significant strides in long-context modeling. Represented by DeepSeek-OCR, these models encode rendered text into continuous vision tokens, achieving high compression rates without sacrificing recognition precision. However, viewing the vision encoder as a lossy channel with finite representational capacity raises a fundamental question: what is the information upper bound of visual tokens? To investigate this limit, we conduct controlled stress tests by progressively increasing the information quantity (character count) within an image. We observe a distinct phase-transition phenomenon characterized by three regimes: a near-perfect Stable Phase, an Instability Phase marked by increased error variance, and a total Collapse Phase. We analyze the mechanical origins of these transitions and identify key factors. Furthermore, we formulate a probabilistic scaling law that unifies average vision token load and visual density into a latent difficulty metric. Extensive experiments across various Vision-Language Models demonstrate the universality of this scaling law, providing critical empirical guidance for optimizing the efficiency-accuracy trade-off in visual context compression.

Problem

• vision token
• information capacity
• recognition limits
• scaling law
• visual context compression

Innovation

• vision token capacity
• scaling law
• phase transition
• visual context compression
• Vision-Language Models