Aimon Rahman, Joymallya Chakraborty, Vidit Bansal, Abhishek Dan
This paper proposes mmVB, a lightweight module for scalable multi-image reasoning in multimodal large language models. mmVB assigns query-conditioned per-image token budgets, reducing visual input tokens by 9x. This yields faster inference and lower computational costs while maintaining accuracy.
The proposed method enables efficient multi-image reasoning without retraining the base model, making it practical for real-world applications. This is significant because multi-image reasoning is crucial for tasks like image comparison, counting, and temporal reasoning. The approach also reduces computational costs, making it more accessible.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.