RADIO1D: Elastic Representations for Condensed Vision Modeling

2026-07-03

AI Summary

This work addresses the limitations of conventional vision-language models, which rely on fixed-size 2D image patch features and struggle to balance semantic completeness with computational efficiency. The authors propose an elastic one-dimensional visual representation mechanism that compresses images into variable-length, compact token sequences through a multi-teacher knowledge distillation framework combined with an autoencoder architecture. This approach enables on-demand trade-offs between computational cost and accuracy, effectively capturing global semantics with as few as a single token. Integrated with image-text alignment strategies such as SigLIP2, the method significantly reduces computational overhead while simultaneously improving performance on compositional visual retrieval and scene understanding across multiple multimodal benchmarks.

Abstract

This paper challenges the assumption that vision-language models (VLMs) require fixed patch-based 2D vision features. Analyzing fine-tuned vision encoders, we find that representations become increasingly abstract and less spatially coherent during VLM training. Notably, models trained with image-text alignment (such as SigLIP2) develop a small number of specialized tokens that effectively summarize global image content. Building on this, we introduce RADIO1D, which compresses images into a compact, variable-length 1D token sequence using multi-teacher knowledge distillation and an autoencoder design. The resulting representations exhibit strong hierarchical summarization, enabling accurate scene understanding - even with a single token - and support improved composition-aware image retrieval. In VLMs, RADIO1D provides flexible accuracy-efficiency tradeoffs through adjustable token counts, delivering competitive performance on diverse multimodal benchmarks with lower computational overhead and better accuracy.

Problem

• vision-language models
• 2D vision features
• spatial coherence
• representation compression
• efficiency-accuracy tradeoff

Innovation

• RADIO1D
• 1D token sequence
• knowledge distillation
• vision-language models
• hierarchical summarization