MVPruner: Dynamic Token Pruning for Accelerating Multi-view Vision-Language Models in Autonomous Driving

A new token pruning method called MVPruner aims to accelerate multi-view vision-language models used in autonomous driving by dynamically discarding unnecessary visual data, according to a preprint posted to the arXiv repository on June 26, 2026 [1]. Vision-Language Models (VLMs) have shown improved generalization and interpretability for autonomous driving systems, but their efficiency is hampered by the long sequences of visual tokens they must process, especially in standard multi-view configurations [1]. Existing approaches to this problem use fixed pruning rates and static importance metrics, which fail to account for how the significance of information from different camera views changes during inference [1]. The authors of the new paper state that their analysis found multi-view VLMs inherently encode task-related view priors in deeper layers and exhibit dynamic information requirements [2]. MVPruner was designed to align its pruning behavior with these dynamic needs through a two-stage process. The first stage allocates pruning budgets based on the information diversity of each view and retains tokens that contribute consistently across stages, while the second stage uses instruction text to guide budget allocation and token selection for task alignment [2]. When applied to the DriveMM model, MVPruner delivered an 87.3% reduction in FLOPs and a 4.97-times speedup in the prefilling phase, while retaining 98.5% accuracy on the DriveLM benchmark [1][2]. The research was shared on arXiv, an open-access repository for electronic preprints that has been in operation since 1991 and now receives about 24,000 submissions per month [6]. The platform also hosts community-developed tools through its arXivLabs framework, which provides features such as bibliographic explorers and code finders on article pages [4][5]. The MVPruner paper appeared alongside these standard arXivLabs integrations, which are developed by third-party collaborators under guidelines that emphasize openness and user data privacy [5].