The field of multimodal models is moving towards improving inference efficiency and tracking performance. Recent developments focus on reducing computational complexity and suppressing background interference. Notable advancements include novel token pruning and scheduling frameworks, as well as revisiting existing architectures to leverage hidden capabilities. These innovations aim to enhance accuracy and speed in various applications. Noteworthy papers include: Parallel Vision Token Scheduling for Fast and Accurate Multimodal LMMs Inference, which presents a training-free scheduling framework that reduces computational complexity. CPDATrack, a novel tracking framework that suppresses interference from background and distractor tokens while enhancing computational efficiency. SelfMOTR, a tracking transformer that relies on self-generated detection priors, demonstrating strong performance on DanceTrack. Object-Centric Vision Token Pruning for Vision Language Models, a direct and guaranteed approach to select representative vision tokens for high-efficiency yet accuracy-preserving VLM inference.