

















Authors:Yuchen Feng, Zhenyu Zhang, Naibin Gu, Yilong Chen, Peng Fu, Zheng Lin, Shuohuan Wang, Yu Sun, Hua Wu, Weiping Wang, Haifeng Wang
Abstract:Multimodal large language models (MLLMs) have achieved remarkable progress on various vision-language tasks, yet their visual perception remains limited. Humans, in comparison, perceive complex scenes efficiently by dynamically scanning and focusing on salient regions in a sequential "blink-like" process. Motivated by this strategy, we first investigate whether MLLMs exhibit similar behavior. Our pilot analysis reveals that MLLMs naturally attend to different visual regions across layers and that selectively allocating more computation to salient tokens can enhance visual perception. Building on this insight, we propose Blink, a dynamic visual token resolution framework that emulates the human-inspired process within a single forward pass. Specifically, Blink includes two modules: saliency-guided scanning and dynamic token resolution. It first estimates the saliency of visual tokens in each layer based on the attention map, and extends important tokens through a plug-and-play token super-resolution (TokenSR) module. In the next layer, it drops the extended tokens when they lose focus. This dynamic mechanism balances broad exploration and fine-grained focus, thereby enhancing visual perception adaptively and efficiently. Extensive experiments validate Blink, demonstrating its effectiveness in enhancing visual perception and multimodal understanding.
| Comments: | CVPR 2026 |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV) |
| Cite as: | arXiv:2512.10548 [cs.CV] |
| (or arXiv:2512.10548v3 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2512.10548 arXiv-issued DOI via DataCite |
From: Yuchen Feng [view email]
[v1]
Thu, 11 Dec 2025 11:27:25 UTC (2,863 KB)
[v2]
Wed, 25 Mar 2026 15:36:38 UTC (2,864 KB)
[v3]
Sat, 23 May 2026 16:19:27 UTC (2,864 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理,仅供阅读参考。 原文来自 — 版权归原作者所有。