Personalized federated learning for multilingual sentiment analysis poses significant challenges arising from linguistic heterogeneity, non-IID data distributions, and strict privacy requirements. This paper proposes FedPerX, a federated transformer framework that integrates residual adapter-based personalization with adaptive multi-granular differential privacy. The architecture leverages a frozen multilingual backbone (XLM-R) while enabling each client to train lightweight, client-specific adapters. Privacy is enforced through dynamic noise injection at both the feature and adapter levels, calibrated using gradient sensitivity. FedPerX is evaluated on two multilingual benchmarks-MARC and TSMD-spanning structured reviews and informal social media content across more than ten languages. Experimental results demonstrate consistent improvements over seven state-of-the-art baselines, with up to +4.3% gains in macro-F1, a 70% reduction in communication overhead, and the lowest variance in client-level performance. Comprehensive analyses, including fairness, personalization gap, privacy-utility trade-off, and ablation studies, validate the framework's robustness and adaptability. FedPerX advances the design of scalable, personalized, and privacy-preserving models for federated multilingual sentiment analysis.

Data augmentation is crucial for addressing insufficient training data, especially for augmenting positive samples. However, existing methods mostly rely on neural network-based feedback for data augmentation and often overlook the optimization of feature distribution. In this study, we present a practical, distribution-preserving data augmentation pipeline that augments positive samples by optimizing a feature indicator (e.g., two-dimensional entropy), aiming to maintain alignment with the original data distribution. Inspired by the manifold hypothesis, we propose a Manifold Heuristic Optimization Algorithm (MHOA), which augments positive samples by exploring the low-dimensional Euclidean space around object contour pixels instead of the entire decision space. Guided by a "distribution-preservation-first" perspective, our approach explicitly optimizes fidelity to the original data manifold and only retains augmented samples whose feature statistics (e.g., mean, variance) align with the source class. It significantly improves image classification accuracy across neural networks, outperforming state-of-the-art data augmentation methods-especially when the dataset's feature indicator follows a Gaussian distribution. The algorithm's search space, focused on neighborhoods of key feature pixels, is the core driver of its superior performance.

Knowledge distillation (KD) has been successfully applied to dense prediction tasks, and mainstream methods typically boost the student via spatial imitation losses. However, the consecutive downsamplings (e.g., existing in feature pyramid network of detectors) induced in the spatial domain are a type of distortion, hindering the student from analyzing what specific information needs to be imitated, which results in accuracy degradation. To better understand the underlying pattern of corrupted feature maps, we shift our attention to frequency knowledge distillation and propose FreeKD, which determines the optimal localization and extent for the frequency distillation. (1) Frequency Prompts in FreeKD plug into the teacher model, absorbing the semantic frequency context during finetuning. During the distillation period, a pixel-wise frequency mask is generated via Frequency Prompt, to localize those pixels of interest in various frequency bands. (2) A position-aware relational frequency loss is for dense prediction tasks, delivering a high-order spatial enhancement to the student model. While a single distillation loss might not adequately capture both high- and low-frequency signals, especially given their contextual nuances, we enhance FreeKD by introducing a frequency-decoupled strategy. This approach emphasizes relaxed alignment in the high-frequency domain and enforces stronger alignment for low-frequency features. Additionally, we refine the frequency masks by reconstructing the regions of interest based on the student's knowledge, thereby optimizing the distillation process. Extensive experimental results on the widely used COCO, VisDrone, Cityscapes, ADE20K, and COCO-C datasets demonstrate the effectiveness of our proposed framework FreeKD+.

Visual-spatial understanding, defined as the ability to infer object relationships and scene layouts from visual inputs, is fundamental to downstream tasks such as robotic navigation and embodied interaction. However, pre-trained vision-language models (VLMs) remain constrained by spatial uncertainty stemming from inherently 2D observations and by the scarcity of data for 3D spatial understanding. To address these limitations, we proposed a novel framework, SpaceEra, in the NeurIPS 2025 Spotlight paper. Although it achieved significant performance gains, we further observed that its effectiveness is hindered by insufficient input from scanning videos and weak reasoning constraints. To tackle these newly emerged challenges, we extend the original framework into a comprehensive system, termed SpaceEra++, which spans data construction, model design, training optimization, and prompting inference. Specifically, to alleviate input insufficiency, we introduce ScenePick, a frame sampling strategy that balances spatial coverage with object semantics to produce compact yet comprehensive scene representations. In addition, to enhance spatial reasoning, we develop SpaceAlign, which enforces pairwise object constraints by jointly exploiting absolute coordinates and relative spatial relations, thereby aligning optimization with spatial accuracy. Extensive experiments across multiple benchmarks demonstrate consistent improvements over strong baselines, while ablation studies validate both the individual and joint contributions of each component, and further analyses provide guidance for future research.

Zero-shot proxies, also known as training-free metrics, are widely adopted to reduce the computational overhead in neural network evaluation for scenarios such as Neural Architecture Search (NAS), as they do not require any training. Existing zero-shot metrics have several limitations, including weak correlation with the true performance and poor generalisation across different networks or downstream tasks. For example, most of these metrics apply only to either convolutional neural networks (CNNs) or Transformers, but not both. To address these limitations, we propose Sample-Wise Activation Patterns (SWAP), and its derivative, SWAP-Score, a novel and highly effective zero-shot metric. SWAP-Score is broadly applicable across both architecture families and task domains, demonstrating strong predictive performance in the majority of tasks. This metric measures the expressivity of neural networks over a mini-batch of samples, showing a high correlation with the neural networks' ground-truth performance. For both CNNs and Transformers, the SWAP-Score outperforms existing zero-shot metrics across computer vision and natural language processing tasks. For instance, Spearman's correlation coefficient between the SWAP-Score and CIFAR-10 validation accuracy for DARTS CNNs is 0.93, and 0.71 for FlexiBERT Transformers on GLUE tasks. Moreover, SWAP-Score is label-independent, hence can be applied at the pre-training stage of language models to estimate their performance for downstream tasks. When applied to NAS, SWAP-empowered NAS, SWAP-NAS can achieve competitive performance using only approximately 6 and 9 minutes of GPU time, on CIFAR-10 and ImageNet respectively. Our code is available at: https://github.com/pym1024/SWAP_Universal.

Time series, characterized by a sequence of data points organized in a discrete-time order, are ubiquitous in real-world scenarios. Unlike other data modalities, time series present unique challenges in learning and modeling due to their intricate and dynamic nature, including the entanglement of nonlinear patterns and time-variant trends. Recent years have witnessed remarkable breakthroughs in time series analysis, with techniques shifting from traditional statistical methods to contemporary deep learning models. In this paper, we delve into the design of deep time series models across various analysis tasks and review the existing literature from two perspectives: basic modules and model architectures. Further, we develop and release Time Series Library (TSLib) as a fair benchmark of deep time series models for diverse analysis tasks. TSLib implements 41 prominent models, including both small- and large-scale time series models, covers 30 datasets from different domains, and supports 5 prevalent analysis tasks. Based on TSLib, we evaluate 16 popular deep time series models and 6 advanced time series foundation models. Empirical findings indicate that models with specific structures are apt only at distinct analytical tasks, providing insights for research and adoption of deep time series models. Code and datasets are available at https://github.com/thuml/Time-Series-Library.

To the best of our knowledge, this paper is the first to integrate fractal signal processing with vision graph neural networks, establishing a new graph representation learning paradigm consistent with fractal dynamics. Building on this foundation, we propose a Fractal-domain Vision Graph Neural Network (FD-ViG). Specifically, FD-ViG includes: (i) a Fractal-Domain Learning Module that maps images into the fractal-domain using local Hölder exponents and the Singularity Power Spectrum (SPS), enabling fractal-spatial feature fusion; (ii) a Fractal Graph Construction Module that adaptively generates a topology by combining semantic attention with fractal similarity in the fractal feature space; and (iii) a Graph Propagation Module with power-law multi-scale propagation to realize cross-scale diffusion and aggregation, enabling coupled texture-structure learning. Experiments on UCMerced, RSSCN7, and SIRI-WHU achieve overall accuracies of 91.75%, 89.52%, and 92.78%, respectively. Compared with representative vision graph models such as ViG, WiGNet, and ViHGNN, our method achieves consistent improvements over prior methods across all three datasets, while remaining lightweight (2.6M parameters). Moreover, despite having far fewer parameters than ResNet-18, our model yields competitive or better performance on two datasets, and further demonstrates strong generalization ability in cross-dataset evaluation on SAR imagery. This work provides a principled and effective bridge between fractal theory and graph deep learning, benefiting interpretable remote sensing scene understanding under complex textures and structures.

Multi-view 3D reconstruction, namely, structure-from-motion followed by multi-view stereo, is a fundamental component of 3D computer vision. In general, multi-view 3D reconstruction suffers from an unknown scale ambiguity unless a reference object of known size is present in the scene. In this article, we show that multi-view images captured using a dual-pixel (DP) sensor can automatically resolve the scale ambiguity, without requiring a reference object or prior calibration. Specifically, the defocus blur observed in DP images provides sufficient information to determine the absolute scale when paired with depth maps (up to scale) recovered from multi-view 3D reconstruction. Based on this observation, we develop a simple yet effective linear method to estimate the absolute scale, followed by the intensity-based optimization stage that aligns the left and right DP images by shifting them back toward each other using cross-view blur kernels. Experiments demonstrate the effectiveness of the proposed approach across diverse scenes captured with different cameras and lenses. Code and data are available at https://github.com/lilika-makabe/dp-sfm-tpami.git.

In this paper, we addressed the limitation of relying solely on distribution alignment and source-domain empirical risk minimization in Unsupervised Domain Adaptation (UDA). Our information-theoretic analysis showed that this standard adversarial-based framework neglects the discriminability of target-domain features, leading to suboptimal performance. To bridge this theoretical-practical gap, we defined "good representation learning" as guaranteeing both transferability and discriminability, and proved that an additional loss term targeting target-domain discriminability is necessary. Building on these insights, we proposed a novel adversarial-based UDA framework that explicitly integrates a domain alignment objective with a discriminability-enhancing constraint. Instantiated as Domain-Invariant Representation Learning with Global and Local Consistency (RLGLC), our method leverages Asymmetrically-Relaxed Wasserstein of Wasserstein Distance (AR-WWD) to address class imbalance and semantic dimension weighting, and employs a local consistency mechanism to preserve fine-grained target-domain discriminative information. Extensive experiments across multiple benchmark datasets demonstrate that RLGLC consistently surpasses state-of-the-art methods, confirming the value of our theoretical perspective and underscoring the necessity of enforcing both transferability and discriminability in adversarial-based UDA.

Multimodal large language models (MLLMs) offer immense potential for biomedical AI, yet current applications remain limited to coarse-grained image understanding and basic textual queries-falling short of the fine-grained reasoning required in clinical contexts. In this work, we present a comprehensive solution spanning data, model, and training innovations to advance pixel-level multimodal intelligence in biomedicine. First, we construct MeCoVQA, a new visual-language benchmark that spans eight medical imaging modalities and four core tasks, supporting both spatially-grounded reasoning and fine grained diagnostic comprehension. Building on this, we introduce MedPLIB, an end-to-end biomedical MLLM equipped with pixel level visual understanding. MedPLIB supports diverse multi modal tasks-including VQA, point- and region-based querying, grounding, and segmentation-through unified modeling. To further accommodate the heterogeneous nature of biomedical tasks, we design a task-specialized Mixture-of-Experts (MoE) architecture, where each expert is tailored to a specific task and jointly optimized via unified fine-tuning. This modular design accommodates diverse biomedical tasks while maintaining a unified and efficient architecture. By integrating retrieval-augmented generation (RAG) and in-context learning (ICL), MedPLIB also demonstrates strong generalization on out-of-distribution (OOD) medical image segmentation. Experiments across multiple benchmarks show that MedPLIB sets a new state-of-the-art on biomedical vision-language tasks; notably, it outperforms the best existing small and large models by 19.7 and 15.6 mDice in zero shot pixel-level grounding, highlighting its clinical utility and generalization strength. Code and data are publicly available at GitHub: https://github.com/ShawnHuang497/MedPLIB.