Vision-Language-Action (VLA) models have demonstrated significant potential in complex scene understanding and action reasoning, leading to their increasing adoption in end-to-end autonomous driving systems. However, the long visual tokens of VLA models greatly increase computational costs. Current visual token pruning methods in Vision-Language Models (VLM) rely on either visual token similarity or visual-text attention, but both have shown poor performance in autonomous driving scenarios. Given that human drivers concentrate on relevant foreground areas while driving, we assert that retaining visual tokens containing this foreground information is essential for effective decision-making. Inspired by this, we propose FastDriveVLA, a novel reconstruction-based vision token pruning framework designed specifically for autonomous driving. FastDriveVLA includes a plug-and-play visual token pruner called ReconPruner, which prioritizes foreground information through MAE-style pixel reconstruction. A novel adversarial foreground-background reconstruction strategy is designed to train ReconPruner for the visual encoder of VLA models. Once trained, ReconPruner can be seamlessly applied to different VLA models with the same visual encoder without retraining. To train ReconPruner, we also introduce a large-scale dataset called nuScenes-FG, consisting of 241K image-mask pairs with annotated foreground regions. Our approach achieves state-of-the-art results on the nuScenes open-loop planning benchmark across different pruning ratios.

Clinical studies indicate that the progression of Type 2 Diabetes Mellitus (T2DM) is associated with characteristic alterations in tongue features, which may facilitate non-invasive early detection. However, current deep learning-based tongue imaging approaches for diabetes diagnosis remain constrained by limited datasets, subtle feature variations, dependence on clinical expertise, and the lack of quantitative evaluation. To address these issues, we developed an open-source dataset for T2DM tongue diagnosis (DMT) and benchmarked it using multiple baseline models. Building on DMT, we propose T2Net, a tongue image recognition model for T2DM that simulates the clinical diagnostic process. T2Net comprises four core components: local inspection, pathological clue integration, syndrome identification, and diagnostic confidence estimation. First, T2Net automatically extracts key ROIs by combining large-kernel decomposition with multi-scale learning. Then, a multi-order feature interaction module enables effective fusion of tongue image features across scales to capture pathological clues. Meanwhile, we design a context-aware dynamic aggregation convolution to model long-range dependencies, and propose a flexible focal loss to mimic the diagnostic reasoning process of clinicians, enabling brain-inspired inference. Finally, we propose a clustering-based confidence estimation approach to quantitatively evaluate the reliability of model predictions. Experimental results demonstrate that T2Net achieves highly competitive performance on the DMT dataset, outperforming the second-best baseline by 2.7% in accuracy and 2.0% in F1 score. Moreover, the quantitative evaluation scores are largely consistent with clinical assessments by physicians.

Computational cost of classification is as important as accuracy in on-line classification systems. The computational cost is usually dominated by the cost of computing implicit features of the raw input data. Very few efforts have been made to design classifiers which perform effectively with limited computational power; instead, feature selection is usually employed as a pre-processing step to reduce the cost of running traditional classifiers. We present CoCoST, a novel and effective approach for building classifiers which achieve state-of-the-art classification accuracy, while keeping the expected computational cost of classification low, even without feature selection. CoCost employs a wide range of novel cost-aware decision trees, each of which is tuned to specialize in classifying instances from a subset of the input space, and judiciously consults them depending on the input instance in accordance with a cost-aware meta-classifier. Experimental results on a network flow detection application show that, our approach can achieve better accuracy than classifiers such as SVM and random forests, while achieving 75%-90% reduction in the computational costs.