A research paper proposes ditching the fixed grid of image patches that vision-language models have relied on since the beginning.
The paper, from a team publishing on arXiv, introduces RADIO1D, a system that encodes images into a variable-length sequence of one-dimensional tokens rather than the standard fixed 2D patch grid. It uses multi-teacher knowledge distillation and an autoencoder design to build compact representations. The key finding driving the work: when vision encoders are fine-tuned inside VLMs, their internal representations become more abstract and less tied to spatial position over time. Models trained with image-text alignment - SigLIP2 is the example cited - develop a handful of specialized tokens that essentially summarize the whole image anyway.
That observation matters because it exposes a mismatch at the heart of current VLM design: the architecture forces models to process hundreds of spatial patch tokens, even as the training process pushes representations away from spatial detail and toward global semantics. RADIO1D makes the token count match what the model actually needs. Critically, it allows a single token to support basic scene understanding, with accuracy climbing as more tokens are added - a flexibility that fixed-patch systems cannot offer.
The practical upside is adjustable compute: developers can trade token count against accuracy depending on the task. Vision-language models are already expensive to run, and patch counts are one of the main levers for controlling inference cost. Whether RADIO1D's gains hold outside benchmark conditions - and across the full range of spatially demanding tasks like fine-grained object detection - is the question its authors do not yet answer.