Deploying Vision Transformers (ViTs) in real-world multi-task learning remains challenging due to their massive computational costs and the difficulty of pruning shared backbones without harming task performance. Single-task pruning often causes destructive interference by discarding weights critical to other tasks, while existing multi-task pruning strategies remain costly and unscalable for billion-parameter models. We propose Multi-LoRA Guided Importance Consensus (LoGIC), a unified framework for efficient and robust multi-task ViT pruning. LoGIC follows a two-phase procedure: (i) task-consistent pruning of LoRA modules, guided by a task-adaptive gating mechanism that balances shared and task-specific contributions while enforcing structured sparsity for deployment; and (ii) cross-task consensus pruning of the frozen ViT backbone, which retains both universally shared and task-specialized capabilities, enabling aggressive sparsity without sacrificing accuracy. Across five diverse vision benchmarks, LoGIC achieves up to 50% structured sparsity while maintaining competitive accuracy and surpassing all baselines.

Few-shot learning (FSL) aims to classify unseen examples (query set) into labeled data (support set) through low-dimensional embeddings. However, the diversity and unpredictability of environments and capture devices make FSL more challenging in real-world applications. In this paper, we propose Dual Support Query Shift (DSQS), a novel challenge in FSL that integrates two key issues: inter-set shifts (between support and query sets) and intra-set shifts (within each set), which significantly hinder model performance. To tackle these challenges, we introduce a Dual Alignment framework (DUAL), whose core insight is that clean features can improve optimal transportation (OT) alignment. Firstly, DUAL leverages a robust embedding function enhanced by a repairer network trained with perturbed and adversarially generated “hard” examples to obtain clean features. Additionally, it incorporates a two-stage OT approach with a negative entropy regularizer, which aligns support set instances, minimizes intra-class distances, and uses query data as anchor nodes to achieve effective distribution alignment. We provide a theoretical bound of DUAL and experimental results on three image datasets, compared against 10 state-of-the-art baselines, showing that DUAL achieves a remarkable average performance improvement of 25.66%. Our code is available at https://github.com/siyang-jiang/DUAL.

Generative document retrieval (GDR) uses pre-trained Transformer-based large language models (LLMs) to extract contextual information and directly predict document identifier token sequences, outperforming traditional document retrieval methods. However, LLMs incur significant computational costs, hindering GDR’s practical application and making inference acceleration essential. Early exiting is one of the conditional computing techniques that expedites LLM inference, but it faces challenges when integrated into GDR due to GDR’s semantically hierarchical structured identifiers, which cause error amplification from premature exits. Moreover, although beam search expands the search space, the hierarchical structure of document identifiers restricts the diversity of initial tokens, leading to inefficiencies. In this work, we introduce Bi-Level Early Exiting for Generative Document Retrieval (BiLEE), comprising Layer Level Early Exiting (LLEE) and Token Level Early Exiting (TLEE). LLEEare designed for hierarchical document identifiers, dynamically escaping from the middle layer of the Transformer calculation based on a data-driven calibrated token threshold. TLEE exiting from unpromising candidate sequences, thus discarding unpromising search beams and enhancing beam search efficiency. Both components dynamically balance the speed-to-accuracy trade-offs for different token positions, doubling GDR’s inference speed and obtaining 13× reduction for FLOPs while maintaining the same level of accuracy. Source code: https://github.com/Rui-Fang/BiLEE.

Efficient matrix operations have been deemed keys to efficient data analysis. Dual-Triangular QR Decomposition (DT-QRD) is a critical component in Tall and skinny QR decomposition (TS-QRD), which is a widely-used matrix operation with various applications, such as data compression and feature extraction. In order to accelerate DT-QRD, in this paper, we propose a new acceleration framework, including Global Acceleration Schemes, and Partially$\boldsymbol{Q}$-rotation Skipping, which utilize the special DT structure in both$\mathbf{Q}$and$\mathbf{R}$matrix to reduce the latency and computation resource. Further, we employ the Systolic-Array Based Architecture (1D & 2D) for implementation to reduce the memory usage. Experimental results manifest that our framework achieves$169.70\times\ (\mathbf{1}\mathbf{D})$and$250.13\times\ (\mathbf{2}\mathbf{D})$speedup.