Ask Less, See More: Communication-Conditioned Token Pruning for Vehicle-to-Vehicle Cooperative Autonomous Driving with Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) offer a promising paradigm for vehicle-to-vehicle (V2V) cooperative autonomous driving, enabling language-based decision-making in safety-critical occluded scenarios. However, existing V2V–MLLM frameworks rely on dense token-level sharing and fusion, incurring high communication and inference costs. Moreover, conventional V2V perception methods are limited to feature-sharing paradigms without language reasoning, and existing token pruning strategies fail to consider LiDAR-specific spatial structure and multi-agent fusion. To address these limitations, we propose V2V Communication-Conditioned MLLM Framework (V2V-CCM), a dual-stage cooperative communication framework that broadcasts request messages to all agents and uses them to identify redundant visual tokens. Specifically, Question Semantic Message (QSM) encodes global question intent for question-relevant token selection, while Spatial Coverage Message (SCM) summarizes LiDAR features to identify spatially redundant tokens already observed by other agents. Integrated into dual-stage frameworks, V2V-CCM substantially reduces communication and inference costs while preserving question-relevant tokens and removing spatial redundancy. Extensive experiments on V2V-QA and V2V-GoT-QA demonstrate that V2V-CCM consistently outperforms existing pruning methods and achieves state-of-the-art performance.

Multimodal large language models offer a promising foundation for vehicle-to-vehicle cooperative autonomous driving, where multiple vehicles can jointly understand and reason about safety-critical driving situations. This is especially useful when a vehicle’s own sensors are blocked by nearby objects or other road users. However, existing systems often require vehicles to share a large amount of visual information with one another, leading to high communication and computation costs that make real-time deployment difficult. This paper introduces a dual-stage cooperative communication framework that reduces unnecessary information exchange between vehicles. Instead of directly sharing all visual tokens, each vehicle first broadcasts a compact request message that describes the current driving question and summarizes the spatial regions already covered by its sensors. These messages then guide other vehicles to keep information that is relevant to the question while removing visual tokens that are spatially redundant and do not need to be transmitted. Experiments on cooperative driving question-answering benchmarks show that the proposed method improves efficiency while maintaining strong performance. By reducing communication bandwidth, computational cost, and response latency, this work makes multimodal large language models more practical for real-world cooperative autonomous driving.