LoopVL: Recurrent Visual Intelligence

2026-09-29

AI Summary

This study addresses the limited computational depth of non-recurrent vision-language models (VLMs) by exploring the scalability of recurrent Transformers. We propose Module-Loop and Model-Loop architectures that achieve deep multimodal computation through iterative updates of a unified vision-language state with shared parameters. Furthermore, we reveal a "visual insight" phenomenon, demonstrating that parameter-sharing mechanisms can overcome the performance bottlenecks inherent in conventional architectures. The project encompasses the complete pipeline from scratch pre-training to post-training. Extensive experiments show that our approach surpasses both comparable and larger-scale non-recurrent models on multimodal understanding and visual reasoning benchmarks, thereby validating the effectiveness of recurrent modeling for VLMs.

Abstract

We introduce LoopVL to study whether Loop Transformers can be effectively extended to vision- language models. LoopVL combines Module-Loop and Model-Loop computation to iteratively update a unified vision-language state through shared modules. We train LoopVL from scratch through language pre-training, multimodal training, and post-training. LoopVL outperforms a range of similarly sized and larger non-recurrent models on multimodal understanding and visual reasoning benchmarks. We also observe Visual Aha Moments in LoopVL, characterized by pronounced shifts in visual attention across loops. LoopVL provides practical evidence for recurrent vision-language modeling and offers an intuitive perspective on how shared parameters can support deeper multimodal computation over continuously evolving visual-language states.

Problem

• Loop Transformers
• Vision-Language Models
• Recurrent Computation
• Visual Reasoning
• Multimodal Understanding

Innovation

• Recurrent Vision-Language Model
• Loop Transformer
• Shared Modules
• Visual Aha Moments
• Multimodal Reasoning