Xiang Zhang

Minhua Lin, Juncheng Wu, Zijun Wang et al.

LLM agents are increasingly deployed as systems built around editable external harnesses, including prompts, skills, memories and tools, that shape task execution without changing model parameters. Harness self-evolution adapts such agents by updating these harnesses from execution evidence. Yet it remains unclear whether a model's base capability in task-solving predicts its capabilities in harness self-evolution: which models produce useful harness updates, and which actually benefit from them? We analyze two harness self-evolution capabilities: (i) harness-updating, the capability to produce useful persistent harness updates from execution evidence; (ii) harness-benefit, the capability to benefit from updated harnesses during task solving. Our analysis reveals two findings. First, harness-updating is flat in base capability: models from different capability tiers produce harness updates that lead to surprisingly similar gains; even Qwen3.5-9B's updates yield gains comparable to those of Claude Opus~4.6. Second, harness-benefit is non-monotonic in base capability: weak-tier models benefit little from updated harnesses, mid-tier models benefit most, and strong-tier models benefit less than mid-tier. We trace low gains at the weak tier to two failure modes: weak-tier models may fail to activate relevant harness artifacts, or activate them but fail to follow them faithfully. These findings suggest investing capability budget in the task-solving agent rather than the evolver, and targeting harness invocation and long-horizon instruction following in agent training. Our source code is publicly available at https://github.com/A-EVO-Lab/a-evolve/tree/release/harness-evolution.

Xiang Zhang, Ziyuan Zhao, Theodoros Tsiligkaridis et al.

Pre-training on time series poses a unique challenge due to the potential mismatch between pre-training and target domains, such as shifts in temporal dynamics, fast-evolving trends, and long-range and short-cyclic effects, which can lead to poor downstream performance. While domain adaptation methods can mitigate these shifts, most methods need examples directly from the target domain, making them suboptimal for pre-training. To address this challenge, methods need to accommodate target domains with different temporal dynamics and be capable of doing so without seeing any target examples during pre-training. Relative to other modalities, in time series, we expect that time-based and frequency-based representations of the same example are located close together in the time-frequency space. To this end, we posit that time-frequency consistency (TF-C) -- embedding a time-based neighborhood of an example close to its frequency-based neighborhood -- is desirable for pre-training. Motivated by TF-C, we define a decomposable pre-training model, where the self-supervised signal is provided by the distance between time and frequency components, each individually trained by contrastive estimation. We evaluate the new method on eight datasets, including electrodiagnostic testing, human activity recognition, mechanical fault detection, and physical status monitoring. Experiments against eight state-of-the-art methods show that TF-C outperforms baselines by 15.4% (F1 score) on average in one-to-one settings (e.g., fine-tuning an EEG-pretrained model on EMG data) and by 8.4% (precision) in challenging one-to-many settings (e.g., fine-tuning an EEG-pretrained model for either hand-gesture recognition or mechanical fault prediction), reflecting the breadth of scenarios that arise in real-world applications. Code and datasets: https://github.com/mims-harvard/TFC-pretraining.

Kuniaki Saito, Kihyuk Sohn, Xiang Zhang et al.

In Composed Image Retrieval (CIR), a user combines a query image with text to describe their intended target. Existing methods rely on supervised learning of CIR models using labeled triplets consisting of the query image, text specification, and the target image. Labeling such triplets is expensive and hinders broad applicability of CIR. In this work, we propose to study an important task, Zero-Shot Composed Image Retrieval (ZS-CIR), whose goal is to build a CIR model without requiring labeled triplets for training. To this end, we propose a novel method, called Pic2Word, that requires only weakly labeled image-caption pairs and unlabeled image datasets to train. Unlike existing supervised CIR models, our model trained on weakly labeled or unlabeled datasets shows strong generalization across diverse ZS-CIR tasks, e.g., attribute editing, object composition, and domain conversion. Our approach outperforms several supervised CIR methods on the common CIR benchmark, CIRR and Fashion-IQ. Code will be made publicly available at https://github.com/google-research/composed_image_retrieval.

Shijie Zhou, Zhimeng Guo, Charu Aggarwal et al.

Link prediction is an important task that has wide applications in various domains. However, the majority of existing link prediction approaches assume the given graph follows homophily assumption, and designs similarity-based heuristics or representation learning approaches to predict links. However, many real-world graphs are heterophilic graphs, where the homophily assumption does not hold, which challenges existing link prediction methods. Generally, in heterophilic graphs, there are many latent factors causing the link formation, and two linked nodes tend to be similar in one or two factors but might be dissimilar in other factors, leading to low overall similarity. Thus, one way is to learn disentangled representation for each node with each vector capturing the latent representation of a node on one factor, which paves a way to model the link formation in heterophilic graphs, resulting in better node representation learning and link prediction performance. However, the work on this is rather limited. Therefore, in this paper, we study a novel problem of exploring disentangled representation learning for link prediction on heterophilic graphs. We propose a novel framework DisenLink which can learn disentangled representations by modeling the link formation and perform factor-aware message-passing to facilitate link prediction. Extensive experiments on 13 real-world datasets demonstrate the effectiveness of DisenLink for link prediction on both heterophilic and hemophiliac graphs. Our codes are available at https://github.com/sjz5202/DisenLink

Xiang Zhang, Yang Zhang, Lukas Mehl et al.

Accurately modeling soft boundaries, e.g., hair and defocus blur, is a fundamental challenge in stereo conversion due to the ambiguous blending of foreground and background. Existing depth models primarily predict single-layer depth, leading to ambiguity in depth correspondence at soft boundaries. While matting techniques can capture opacity for layered modeling, they often struggle in complex scenes with multiple targets and usually require user intervention. This paper introduces αDepth, a layered representation that decomposes soft boundaries for high-fidelity stereo conversion. Specifically, we first resolve mixed color and depth ambiguity by estimating layered color and depth values at soft boundaries. Considering complex multi-target scenes, we design a Circular Alpha Representation (CAR) that shifts the paradigm from global target extraction to local boundary decomposition. Unlike prior matting methods restricted to a single foreground/background, CAR enables efficient scene-level inference without manual guidance. Extensive evaluations demonstrate that αDepth achieves state-of-the-art performance in stereo conversion, eliminating background bleeding and structural distortions at soft boundaries.

Lei Yu, Bishan Wang, Xiang Zhang et al.

Super-Resolution from a single motion Blurred image (SRB) is a severely ill-posed problem due to the joint degradation of motion blurs and low spatial resolution. In this paper, we employ events to alleviate the burden of SRB and propose an Event-enhanced SRB (E-SRB) algorithm, which can generate a sequence of sharp and clear images with High Resolution (HR) from a single blurry image with Low Resolution (LR). To achieve this end, we formulate an event-enhanced degeneration model to consider the low spatial resolution, motion blurs, and event noises simultaneously. We then build an event-enhanced Sparse Learning Network (eSL-Net++) upon a dual sparse learning scheme where both events and intensity frames are modeled with sparse representations. Furthermore, we propose an event shuffle-and-merge scheme to extend the single-frame SRB to the sequence-frame SRB without any additional training process. Experimental results on synthetic and real-world datasets show that the proposed eSL-Net++ outperforms state-of-the-art methods by a large margin. Datasets, codes, and more results are available at https://github.com/ShinyWang33/eSL-Net-Plusplus.

Wang Lu, Jindong Wang, Han Yu et al.

Domain generalization (DG) aims to learn a generalizable model from multiple training domains such that it can perform well on unseen target domains. A popular strategy is to augment training data to benefit generalization through methods such as Mixup~\cite{zhang2018mixup}. While the vanilla Mixup can be directly applied, theoretical and empirical investigations uncover several shortcomings that limit its performance. Firstly, Mixup cannot effectively identify the domain and class information that can be used for learning invariant representations. Secondly, Mixup may introduce synthetic noisy data points via random interpolation, which lowers its discrimination capability. Based on the analysis, we propose a simple yet effective enhancement for Mixup-based DG, namely domain-invariant Feature mIXup (FIX). It learns domain-invariant representations for Mixup. To further enhance discrimination, we leverage existing techniques to enlarge margins among classes to further propose the domain-invariant Feature MIXup with Enhanced Discrimination (FIXED) approach. We present theoretical insights about guarantees on its effectiveness. Extensive experiments on seven public datasets across two modalities including image classification (Digits-DG, PACS, Office-Home) and time series (DSADS, PAMAP2, UCI-HAR, and USC-HAD) demonstrate that our approach significantly outperforms nine state-of-the-art related methods, beating the best performing baseline by 6.5\% on average in terms of test accuracy. Code is available at: https://github.com/jindongwang/transferlearning/tree/master/code/deep/fixed.

Dongsheng Luo, Wei Cheng, Yingheng Wang et al.

Various contrastive learning approaches have been proposed in recent years and achieve significant empirical success. While effective and prevalent, contrastive learning has been less explored for time series data. A key component of contrastive learning is to select appropriate augmentations imposing some priors to construct feasible positive samples, such that an encoder can be trained to learn robust and discriminative representations. Unlike image and language domains where ``desired'' augmented samples can be generated with the rule of thumb guided by prefabricated human priors, the ad-hoc manual selection of time series augmentations is hindered by their diverse and human-unrecognizable temporal structures. How to find the desired augmentations of time series data that are meaningful for given contrastive learning tasks and datasets remains an open question. In this work, we address the problem by encouraging both high \textit{fidelity} and \textit{variety} based upon information theory. A theoretical analysis leads to the criteria for selecting feasible data augmentations. On top of that, we propose a new contrastive learning approach with information-aware augmentations, InfoTS, that adaptively selects optimal augmentations for time series representation learning. Experiments on various datasets show highly competitive performance with up to 12.0\% reduction in MSE on forecasting tasks and up to 3.7\% relative improvement in accuracy on classification tasks over the leading baselines.

Xiang Zhang, Yongwen Su, Subarna Tripathi et al.

In this paper, we present TExt Spotting TRansformers (TESTR), a generic end-to-end text spotting framework using Transformers for text detection and recognition in the wild. TESTR builds upon a single encoder and dual decoders for the joint text-box control point regression and character recognition. Other than most existing literature, our method is free from Region-of-Interest operations and heuristics-driven post-processing procedures; TESTR is particularly effective when dealing with curved text-boxes where special cares are needed for the adaptation of the traditional bounding-box representations. We show our canonical representation of control points suitable for text instances in both Bezier curve and polygon annotations. In addition, we design a bounding-box guided polygon detection (box-to-polygon) process. Experiments on curved and arbitrarily shaped datasets demonstrate state-of-the-art performances of the proposed TESTR algorithm.

Yihe Wang, Yu Han, Haishuai Wang et al.

Contrastive representation learning is crucial in medical time series analysis as it alleviates dependency on labor-intensive, domain-specific, and scarce expert annotations. However, existing contrastive learning methods primarily focus on one single data level, which fails to fully exploit the intricate nature of medical time series. To address this issue, we present COMET, an innovative hierarchical framework that leverages data consistencies at all inherent levels in medical time series. Our meticulously designed model systematically captures data consistency from four potential levels: observation, sample, trial, and patient levels. By developing contrastive loss at multiple levels, we can learn effective representations that preserve comprehensive data consistency, maximizing information utilization in a self-supervised manner. We conduct experiments in the challenging patient-independent setting. We compare COMET against six baselines using three diverse datasets, which include ECG signals for myocardial infarction and EEG signals for Alzheimer's and Parkinson's diseases. The results demonstrate that COMET consistently outperforms all baselines, particularly in setup with 10% and 1% labeled data fractions across all datasets. These results underscore the significant impact of our framework in advancing contrastive representation learning techniques for medical time series. The source code is available at https://github.com/DL4mHealth/COMET.

Minhua Lin, Teng Xiao, Enyan Dai et al.

Graph Contrastive Learning (GCL) has emerged as a popular unsupervised graph representation learning method. However, it has been shown that GCL is vulnerable to adversarial attacks on both the graph structure and node attributes. Although empirical approaches have been proposed to enhance the robustness of GCL, the certifiable robustness of GCL is still remain unexplored. In this paper, we develop the first certifiably robust framework in GCL. Specifically, we first propose a unified criteria to evaluate and certify the robustness of GCL. We then introduce a novel technique, RES (Randomized Edgedrop Smoothing), to ensure certifiable robustness for any GCL model, and this certified robustness can be provably preserved in downstream tasks. Furthermore, an effective training method is proposed for robust GCL. Extensive experiments on real-world datasets demonstrate the effectiveness of our proposed method in providing effective certifiable robustness and enhancing the robustness of any GCL model. The source code of RES is available at https://github.com/ventr1c/RES-GCL.

Jiaxiang Liu, Zhida Feng, Pengyu Zou et al.

We introduce ERNIE-Image, an open-source text-to-image generation model built upon an 8B single-stream DiT architecture. ERNIE-Image aims to bridge the gap between current open-source models and leading closed-source systems through more effective mining of large-scale pre-training data and improved supervision quality throughout training. During pre-training, we adopt a bottom-up data construction pipeline that combines fine-grained image categorization, rich caption annotation, aesthetic assessment, and hierarchical sampling. This strategy reduces data noise while preserving long-tail concepts and detailed real-world knowledge, providing a stronger foundation for complex generation tasks. In the post-training stage, we use a top-down data construction pipeline for high-demand scenarios, diversify prompt annotations to better match real user inputs, and apply a stabilized DPO strategy to align the model with human aesthetic preferences. We further train ERNIE-Image-Turbo for efficient 8-NFE generation and propose MT-DMD to mitigate capability drift during distillation. To make the model easier to use in practical scenarios, we equip it with a lightweight Prompt Enhancer that expands concise user intents into structured visual descriptions. In addition, we develop ERNIE-Image-Aes, an industrial-grade aesthetic model, together with ERNIE-Image-Aes-1K, a human-annotated benchmark for realistic aesthetic evaluation. Extensive qualitative and quantitative experiments show that ERNIE-Image achieves leading performance among open-source models and approaches top-tier commercial models in instruction following, text rendering, and aesthetic quality. We release the trained models and aesthetic resources to facilitate further academic research and technical progress in the AIGC community.

Xiang Zhang, Lei Yu

Slow shutter speed and long exposure time of frame-based cameras often cause visual blur and loss of inter-frame information, degenerating the overall quality of captured videos. To this end, we present a unified framework of event-based motion deblurring and frame interpolation for blurry video enhancement, where the extremely low latency of events is leveraged to alleviate motion blur and facilitate intermediate frame prediction. Specifically, the mapping relation between blurry frames and sharp latent images is first predicted by a learnable double integral network, and a fusion network is then proposed to refine the coarse results via utilizing the information from consecutive blurry inputs and the concurrent events. By exploring the mutual constraints among blurry frames, latent images, and event streams, we further propose a self-supervised learning framework to enable network training with real-world blurry videos and events. Extensive experiments demonstrate that our method compares favorably against the state-of-the-art approaches and achieves remarkable performance on both synthetic and real-world datasets.

Tianchun Wang, Wei Cheng, Dongsheng Luo et al.

Personalized Federated Learning (PFL) which collaboratively trains a federated model while considering local clients under privacy constraints has attracted much attention. Despite its popularity, it has been observed that existing PFL approaches result in sub-optimal solutions when the joint distribution among local clients diverges. To address this issue, we present Federated Modular Network (FedMN), a novel PFL approach that adaptively selects sub-modules from a module pool to assemble heterogeneous neural architectures for different clients. FedMN adopts a light-weighted routing hypernetwork to model the joint distribution on each client and produce the personalized selection of the module blocks for each client. To reduce the communication burden in existing FL, we develop an efficient way to interact between the clients and the server. We conduct extensive experiments on the real-world test beds and the results show both the effectiveness and efficiency of the proposed FedMN over the baselines.

Kuniaki Saito, Kihyuk Sohn, Xiang Zhang et al.

Image-classification datasets have been used to pretrain image recognition models. Recently, web-scale image-caption datasets have emerged as a source of powerful pretraining alternative. Image-caption datasets are more ``open-domain'', containing a wider variety of scene types and vocabulary words than traditional classification datasets, and models trained on these datasets have demonstrated strong performance on few- and zero-shot recognition tasks. When naively unifying image-classification and -caption dataset, we show that such dataset biases negatively affect pre-training by reducing the generalizability of learned representations and thus jeopardizing zero-shot performance since the unification can tailor the model for the classification dataset, making it vulnerable to the distribution shift from the dataset. In this work, we address the problem by disentangling the dataset bias using prefix tokens that inform a language encoder of the type of the input dataset (e.g., image-classification or caption) at training time. This approach allows the language encoder to share the knowledge from two datasets as well as switch the mode of feature extraction, i.e., image-classification dataset or image-caption dataset tailored mode, where we use image-caption mode in the zero-shot evaluation. Our method is generic and can be easily integrated into existing VL pre-training objectives such as CLIP or UniCL. In experiments, we show that this simple technique improves the performance in zero-shot image recognition accuracy and robustness to the image-level distribution shift.

Ruiqi Wang, Jinyang Huang, Jie Zhang et al.

Depression is a prevalent mental health disorder that significantly impacts individuals' lives and well-being. Early detection and intervention are crucial for effective treatment and management of depression. Recently, there are many end-to-end deep learning methods leveraging the facial expression features for automatic depression detection. However, most current methods overlook the temporal dynamics of facial expressions. Although very recent 3DCNN methods remedy this gap, they introduce more computational cost due to the selection of CNN-based backbones and redundant facial features. To address the above limitations, by considering the timing correlation of facial expressions, we propose a novel framework called FacialPulse, which recognizes depression with high accuracy and speed. By harnessing the bidirectional nature and proficiently addressing long-term dependencies, the Facial Motion Modeling Module (FMMM) is designed in FacialPulse to fully capture temporal features. Since the proposed FMMM has parallel processing capabilities and has the gate mechanism to mitigate gradient vanishing, this module can also significantly boost the training speed. Besides, to effectively use facial landmarks to replace original images to decrease information redundancy, a Facial Landmark Calibration Module (FLCM) is designed to eliminate facial landmark errors to further improve recognition accuracy. Extensive experiments on the AVEC2014 dataset and MMDA dataset (a depression dataset) demonstrate the superiority of FacialPulse on recognition accuracy and speed, with the average MAE (Mean Absolute Error) decreased by 21% compared to baselines, and the recognition speed increased by 100% compared to state-of-the-art methods. Codes are released at https://github.com/volatileee/FacialPulse.

Enyan Dai, Minhua Lin, Xiang Zhang et al.

Graph Neural Networks (GNNs) have achieved promising results in various tasks such as node classification and graph classification. Recent studies find that GNNs are vulnerable to adversarial attacks. However, effective backdoor attacks on graphs are still an open problem. In particular, backdoor attack poisons the graph by attaching triggers and the target class label to a set of nodes in the training graph. The backdoored GNNs trained on the poisoned graph will then be misled to predict test nodes to target class once attached with triggers. Though there are some initial efforts in graph backdoor attacks, our empirical analysis shows that they may require a large attack budget for effective backdoor attacks and the injected triggers can be easily detected and pruned. Therefore, in this paper, we study a novel problem of unnoticeable graph backdoor attacks with limited attack budget. To fully utilize the attack budget, we propose to deliberately select the nodes to inject triggers and target class labels in the poisoning phase. An adaptive trigger generator is deployed to obtain effective triggers that are difficult to be noticed. Extensive experiments on real-world datasets against various defense strategies demonstrate the effectiveness of our proposed method in conducting effective unnoticeable backdoor attacks.

Gaotang Li, Marlena Duda, Xiang Zhang et al.

Brain graphs, which model the structural and functional relationships between brain regions, are crucial in neuroscientific and clinical applications involving graph classification. However, dense brain graphs pose computational challenges including high runtime and memory usage and limited interpretability. In this paper, we investigate effective designs in Graph Neural Networks (GNNs) to sparsify brain graphs by eliminating noisy edges. While prior works remove noisy edges based on explainability or task-irrelevant properties, their effectiveness in enhancing performance with sparsified graphs is not guaranteed. Moreover, existing approaches often overlook collective edge removal across multiple graphs. To address these issues, we introduce an iterative framework to analyze different sparsification models. Our findings are as follows: (i) methods prioritizing interpretability may not be suitable for graph sparsification as they can degrade GNNs' performance in graph classification tasks; (ii) simultaneously learning edge selection with GNN training is more beneficial than post-training; (iii) a shared edge selection across graphs outperforms separate selection for each graph; and (iv) task-relevant gradient information aids in edge selection. Based on these insights, we propose a new model, Interpretable Graph Sparsification (IGS), which enhances graph classification performance by up to 5.1% with 55.0% fewer edges. The retained edges identified by IGS provide neuroscientific interpretations and are supported by well-established literature.

Xiaohong Liu, Guoxing Yang, Yulin Luo et al.

Radiology is a vital and complex component of modern clinical workflow and covers many tasks. Recently, vision-language (VL) foundation models in medicine have shown potential in processing multimodal information, offering a unified solution for various radiology tasks. However, existing studies either pre-trained VL models on natural data or did not fully integrate vision-language architecture and pretraining, often neglecting the unique multimodal complexity in radiology images and their textual contexts. Additionally, their practical applicability in real-world scenarios remains underexplored. Here, we present RadFound, a large and open-source vision-language foundation model tailored for radiology, that is trained on the most extensive dataset of over 8.1 million images and 250,000 image-text pairs, covering 19 major organ systems and 10 imaging modalities. To establish expert-level multimodal perception and generation capabilities, RadFound introduces an enhanced vision encoder to capture intra-image local features and inter-image contextual information, and a unified cross-modal learning design tailored to radiology. To fully assess the models' capability, we construct a benchmark, RadVLBench, including radiology interpretation tasks like medical vision-language question-answering, as well as text generation tasks ranging from captioning to report generation. We also propose a human evaluation framework. When evaluated on the real-world benchmark involving three representative modalities, 2D images (chest X-rays), multi-view images (mammograms), and 3D images (thyroid CT scans), RadFound significantly outperforms other VL foundation models on both quantitative metrics and human evaluation. In summary, the development of RadFound represents an advancement in radiology generalists, demonstrating broad applicability potential for integration into clinical workflows.

Long Lan, Xiao Teng, Jing Zhang et al.

Unsupervised person re-identification is a challenging and promising task in computer vision. Nowadays unsupervised person re-identification methods have achieved great progress by training with pseudo labels. However, how to purify feature and label noise is less explicitly studied in the unsupervised manner. To purify the feature, we take into account two types of additional features from different local views to enrich the feature representation. The proposed multi-view features are carefully integrated into our cluster contrast learning to leverage more discriminative cues that the global feature easily ignored and biased. To purify the label noise, we propose to take advantage of the knowledge of teacher model in an offline scheme. Specifically, we first train a teacher model from noisy pseudo labels, and then use the teacher model to guide the learning of our student model. In our setting, the student model could converge fast with the supervision of the teacher model thus reduce the interference of noisy labels as the teacher model greatly suffered. After carefully handling the noise and bias in the feature learning, our purification modules are proven to be very effective for unsupervised person re-identification. Extensive experiments on three popular person re-identification datasets demonstrate the superiority of our method. Especially, our approach achieves a state-of-the-art accuracy 85.8\% @mAP and 94.5\% @Rank-1 on the challenging Market-1501 benchmark with ResNet-50 under the fully unsupervised setting. The code will be released.

Xiang Zhang, Muhammad Abdul-Mageed, Laks V. S. Lakshmanan

The Transformer architecture excels in a variety of language modeling tasks, outperforming traditional neural architectures such as RNN and LSTM. This is partially due to its elimination of recurrent connections, which allows for parallel training and a smoother flow of gradients. However, this move away from recurrent structures places the Transformer model at the lower end of Chomsky's computational hierarchy, imposing limitations on its computational abilities. Consequently, even advanced Transformer-based models face considerable difficulties in tasks like counting, string reversal, and multiplication. These tasks, though seemingly elementary, require a level of computational complexity that exceeds the capabilities of the Transformer architecture. Concurrently, the emergence of ``Chain of Thought" (CoT) prompting has enabled Transformer-based language models to tackle tasks that were previously impossible or poorly executed. In this work, we thoroughly investigate the influence of recurrent structures in neural models on their reasoning abilities and computability, contrasting the role autoregression plays in the neural models' computational power. We then shed light on how the CoT approach can mimic recurrent computation and act as a bridge between autoregression and recurrence in the context of language models. It is this approximated recurrence that notably improves the model's performance and computational capacity. Moreover, we revisit recent recurrent-based Transformer model designs, focusing on their computational abilities through our proposed concept of ``recurrence-completeness" and identify key theoretical limitations in models like Linear Transformer and RWKV. Through this, we aim to provide insight into the neural model architectures and prompt better model design.

Changcheng Xiao, Qiong Cao, Yujie Zhong et al.

Significant progress has been achieved in multi-object tracking (MOT) through the evolution of detection and re-identification (ReID) techniques. Despite these advancements, accurately tracking objects in scenarios with homogeneous appearance and heterogeneous motion remains a challenge. This challenge arises from two main factors: the insufficient discriminability of ReID features and the predominant utilization of linear motion models in MOT. In this context, we introduce a novel motion-based tracker, MotionTrack, centered around a learnable motion predictor that relies solely on object trajectory information. This predictor comprehensively integrates two levels of granularity in motion features to enhance the modeling of temporal dynamics and facilitate precise future motion prediction for individual objects. Specifically, the proposed approach adopts a self-attention mechanism to capture token-level information and a Dynamic MLP layer to model channel-level features. MotionTrack is a simple, online tracking approach. Our experimental results demonstrate that MotionTrack yields state-of-the-art performance on datasets such as Dancetrack and SportsMOT, characterized by highly complex object motion.

Tencent Hunyuan Team, Ao Liu, Botong Zhou et al. · tencent-ai

As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models.

Yuwei Yin, Jean Kaddour, Xiang Zhang et al.

Data augmentation has been established as an efficacious approach to supplement useful information for low-resource datasets. Traditional augmentation techniques such as noise injection and image transformations have been widely used. In addition, generative data augmentation (GDA) has been shown to produce more diverse and flexible data. While generative adversarial networks (GANs) have been frequently used for GDA, they lack diversity and controllability compared to text-to-image diffusion models. In this paper, we propose TTIDA (Text-to-Text-to-Image Data Augmentation) to leverage the capabilities of large-scale pre-trained Text-to-Text (T2T) and Text-to-Image (T2I) generative models for data augmentation. By conditioning the T2I model on detailed descriptions produced by T2T models, we are able to generate photo-realistic labeled images in a flexible and controllable manner. Experiments on in-domain classification, cross-domain classification, and image captioning tasks show consistent improvements over other data augmentation baselines. Analytical studies in varied settings, including few-shot, long-tail, and adversarial, further reinforce the effectiveness of TTIDA in enhancing performance and increasing robustness.

Tianxiang Zhao, Wenchao Yu, Suhang Wang et al.

Imitation learning has achieved great success in many sequential decision-making tasks, in which a neural agent is learned by imitating collected human demonstrations. However, existing algorithms typically require a large number of high-quality demonstrations that are difficult and expensive to collect. Usually, a trade-off needs to be made between demonstration quality and quantity in practice. Targeting this problem, in this work we consider the imitation of sub-optimal demonstrations, with both a small clean demonstration set and a large noisy set. Some pioneering works have been proposed, but they suffer from many limitations, e.g., assuming a demonstration to be of the same optimality throughout time steps and failing to provide any interpretation w.r.t knowledge learned from the noisy set. Addressing these problems, we propose {\method} by evaluating and imitating at the sub-demonstration level, encoding action primitives of varying quality into different skills. Concretely, {\method} consists of a high-level controller to discover skills and a skill-conditioned module to capture action-taking policies, and is trained following a two-phase pipeline by first discovering skills with all demonstrations and then adapting the controller to only the clean set. A mutual-information-based regularization and a dynamic sub-demonstration optimality estimator are designed to promote disentanglement in the skill space. Extensive experiments are conducted over two gym environments and a real-world healthcare dataset to demonstrate the superiority of {\method} in learning from sub-optimal demonstrations and its improved interpretability by examining learned skills.

Yixiao Wang, Yifei Zhang, Mingxiao Huo et al.

The increasing complexity of tasks in robotics demands efficient strategies for multitask and continual learning. Traditional models typically rely on a universal policy for all tasks, facing challenges such as high computational costs and catastrophic forgetting when learning new tasks. To address these issues, we introduce a sparse, reusable, and flexible policy, Sparse Diffusion Policy (SDP). By adopting Mixture of Experts (MoE) within a transformer-based diffusion policy, SDP selectively activates experts and skills, enabling efficient and task-specific learning without retraining the entire model. SDP not only reduces the burden of active parameters but also facilitates the seamless integration and reuse of experts across various tasks. Extensive experiments on diverse tasks in both simulations and real world show that SDP 1) excels in multitask scenarios with negligible increases in active parameters, 2) prevents forgetting in continual learning of new tasks, and 3) enables efficient task transfer, offering a promising solution for advanced robotic applications. Demos and codes can be found in https://forrest-110.github.io/sparse_diffusion_policy/.

Xiang Zhang, Lei Yu, Wen Yang et al.

Event-based motion deblurring has shown promising results by exploiting low-latency events. However, current approaches are limited in their practical usage, as they assume the same spatial resolution of inputs and specific blurriness distributions. This work addresses these limitations and aims to generalize the performance of event-based deblurring in real-world scenarios. We propose a scale-aware network that allows flexible input spatial scales and enables learning from different temporal scales of motion blur. A two-stage self-supervised learning scheme is then developed to fit real-world data distribution. By utilizing the relativity of blurriness, our approach efficiently ensures the restored brightness and structure of latent images and further generalizes deblurring performance to handle varying spatial and temporal scales of motion blur in a self-distillation manner. Our method is extensively evaluated, demonstrating remarkable performance, and we also introduce a real-world dataset consisting of multi-scale blurry frames and events to facilitate research in event-based deblurring.

Xiang Zhang, Taoyue Wang, Xiaotian Li et al.

Contrastive learning has shown promising potential for learning robust representations by utilizing unlabeled data. However, constructing effective positive-negative pairs for contrastive learning on facial behavior datasets remains challenging. This is because such pairs inevitably encode the subject-ID information, and the randomly constructed pairs may push similar facial images away due to the limited number of subjects in facial behavior datasets. To address this issue, we propose to utilize activity descriptions, coarse-grained information provided in some datasets, which can provide high-level semantic information about the image sequences but is often neglected in previous studies. More specifically, we introduce a two-stage Contrastive Learning with Text-Embeded framework for Facial behavior understanding (CLEF). The first stage is a weakly-supervised contrastive learning method that learns representations from positive-negative pairs constructed using coarse-grained activity information. The second stage aims to train the recognition of facial expressions or facial action units by maximizing the similarity between image and the corresponding text label names. The proposed CLEF achieves state-of-the-art performance on three in-the-lab datasets for AU recognition and three in-the-wild datasets for facial expression recognition.

Tianxiang Zhao, Dongsheng Luo, Xiang Zhang et al.

Uncovering rationales behind predictions of graph neural networks (GNNs) has received increasing attention over recent years. Instance-level GNN explanation aims to discover critical input elements, like nodes or edges, that the target GNN relies upon for making predictions. Though various algorithms are proposed, most of them formalize this task by searching the minimal subgraph which can preserve original predictions. However, an inductive bias is deep-rooted in this framework: several subgraphs can result in the same or similar outputs as the original graphs. Consequently, they have the danger of providing spurious explanations and fail to provide consistent explanations. Applying them to explain weakly-performed GNNs would further amplify these issues. To address this problem, we theoretically examine the predictions of GNNs from the causality perspective. Two typical reasons of spurious explanations are identified: confounding effect of latent variables like distribution shift, and causal factors distinct from the original input. Observing that both confounding effects and diverse causal rationales are encoded in internal representations, we propose a simple yet effective countermeasure by aligning embeddings. Concretely, concerning potential shifts in the high-dimensional space, we design a distribution-aware alignment algorithm based on anchors. This new objective is easy to compute and can be incorporated into existing techniques with no or little effort. Theoretical analysis shows that it is in effect optimizing a more faithful explanation objective in design, which further justifies the proposed approach.

Junjie Xu, Enyan Dai, Xiang Zhang et al.

Graph neural networks (GNNs) have achieved great success in various graph problems. However, most GNNs are Message Passing Neural Networks (MPNNs) based on the homophily assumption, where nodes with the same label are connected in graphs. Real-world problems bring us heterophily problems, where nodes with different labels are connected in graphs. MPNNs fail to address the heterophily problem because they mix information from different distributions and are not good at capturing global patterns. Therefore, we investigate a novel Graph Memory Networks model on Heterophilous Graphs (HP-GMN) to the heterophily problem in this paper. In HP-GMN, local information and global patterns are learned by local statistics and the memory to facilitate the prediction. We further propose regularization terms to help the memory learn global information. We conduct extensive experiments to show that our method achieves state-of-the-art performance on both homophilous and heterophilous graphs.

Hao Zhuang, Wenjian Yu, Shih-Hung Weng et al.

We design an algorithmic framework using matrix exponentials for time-domain simulation of power delivery network (PDN). Our framework can reuse factorized matrices to simulate the large-scale linear PDN system with variable stepsizes. In contrast, current conventional PDN simulation solvers have to use fixed step-size approach in order to reuse factorized matrices generated by the expensive matrix decomposition. Based on the proposed exponential integration framework, we design a PDN solver R-MATEX with the flexible time-stepping capability. The key operation of matrix exponential and vector product (MEVP) is computed by the rational Krylov subspace method. To further improve the runtime, we also propose a distributed computing framework DR-MATEX. DR-MATEX reduces Krylov subspace generations caused by frequent breakpoints from a large number of current sources during simulation. By virtue of the superposition property of linear system and scaling invariance property of Krylov subspace, DR-MATEX can divide the whole simulation task into subtasks based on the alignments of breakpoints among those sources. The subtasks are processed in parallel at different computing nodes without any communication during the computation of transient simulation. The final result is obtained by summing up the partial results among all the computing nodes after they finish the assigned subtasks. Therefore, our computation model belongs to the category known as Embarrassingly Parallel model. Experimental results show R-MATEX and DR-MATEX can achieve up to around 14.4X and 98.0X runtime speedups over traditional trapezoidal integration based solver with fixed timestep approach.

Lei Yu, Xiang Zhang, Wei Liao et al.

Although synthetic aperture imaging (SAI) can achieve the seeing-through effect by blurring out off-focus foreground occlusions while recovering in-focus occluded scenes from multi-view images, its performance is often deteriorated by dense occlusions and extreme lighting conditions. To address the problem, this paper presents an Event-based SAI (E-SAI) method by relying on the asynchronous events with extremely low latency and high dynamic range acquired by an event camera. Specifically, the collected events are first refocused by a Refocus-Net module to align in-focus events while scattering out off-focus ones. Following that, a hybrid network composed of spiking neural networks (SNNs) and convolutional neural networks (CNNs) is proposed to encode the spatio-temporal information from the refocused events and reconstruct a visual image of the occluded targets. Extensive experiments demonstrate that our proposed E-SAI method can achieve remarkable performance in dealing with very dense occlusions and extreme lighting conditions and produce high-quality images from pure events. Codes and datasets are available at https://dvs-whu.cn/projects/esai/.

Tianxiang Zhao, Dongsheng Luo, Xiang Zhang et al.

Graph serves as a powerful tool for modeling data that has an underlying structure in non-Euclidean space, by encoding relations as edges and entities as nodes. Despite developments in learning from graph-structured data over the years, one obstacle persists: graph imbalance. Although several attempts have been made to target this problem, they are limited to considering only class-level imbalance. In this work, we argue that for graphs, the imbalance is likely to exist at the sub-class topology group level. Due to the flexibility of topology structures, graphs could be highly diverse, and learning a generalizable classification boundary would be difficult. Therefore, several majority topology groups may dominate the learning process, rendering others under-represented. To address this problem, we propose a new framework {\method} and design (1 a topology extractor, which automatically identifies the topology group for each instance with explicit memory cells, (2 a training modulator, which modulates the learning process of the target GNN model to prevent the case of topology-group-wise under-representation. {\method} can be used as a key component in GNN models to improve their performances under the data imbalance setting. Analyses on both topology-level imbalance and the proposed {\method} are provided theoretically, and we empirically verify its effectiveness with both node-level and graph-level classification as the target tasks.

Xiang Zhang, Lijun Yin

Multi-modal learning has been intensified in recent years, especially for applications in facial analysis and action unit detection whilst there still exist two main challenges in terms of 1) relevant feature learning for representation and 2) efficient fusion for multi-modalities. Recently, there are a number of works have shown the effectiveness in utilizing the attention mechanism for AU detection, however, most of them are binding the region of interest (ROI) with features but rarely apply attention between features of each AU. On the other hand, the transformer, which utilizes a more efficient self-attention mechanism, has been widely used in natural language processing and computer vision tasks but is not fully explored in AU detection tasks. In this paper, we propose a novel end-to-end Multi-Head Fused Transformer (MFT) method for AU detection, which learns AU encoding features representation from different modalities by transformer encoder and fuses modalities by another fusion transformer module. Multi-head fusion attention is designed in the fusion transformer module for the effective fusion of multiple modalities. Our approach is evaluated on two public multi-modal AU databases, BP4D, and BP4D+, and the results are superior to the state-of-the-art algorithms and baseline models. We further analyze the performance of AU detection from different modalities.

Xiang Zhang, Huiyuan Yang, Taoyue Wang et al.

Recent studies have focused on utilizing multi-modal data to develop robust models for facial Action Unit (AU) detection. However, the heterogeneity of multi-modal data poses challenges in learning effective representations. One such challenge is extracting relevant features from multiple modalities using a single feature extractor. Moreover, previous studies have not fully explored the potential of multi-modal fusion strategies. In contrast to the extensive work on late fusion, there are limited investigations on early fusion for channel information exploration. This paper presents a novel multi-modal reconstruction network, named Multimodal Channel-Mixing (MCM), as a pre-trained model to learn robust representation for facilitating multi-modal fusion. The approach follows an early fusion setup, integrating a Channel-Mixing module, where two out of five channels are randomly dropped. The dropped channels then are reconstructed from the remaining channels using masked autoencoder. This module not only reduces channel redundancy, but also facilitates multi-modal learning and reconstruction capabilities, resulting in robust feature learning. The encoder is fine-tuned on a downstream task of automatic facial action unit detection. Pre-training experiments were conducted on BP4D+, followed by fine-tuning on BP4D and DISFA to assess the effectiveness and robustness of the proposed framework. The results demonstrate that our method meets and surpasses the performance of state-of-the-art baseline methods.

Minhua Lin, Zhiwei Zhang, Hanqing Lu et al.

Memory-augmented LLM agents maintain external memory banks to support long-horizon interaction, yet most existing systems treat construction, retrieval, and utilization as isolated subroutines. This creates two coupled challenges: strategic blindness on the forward path of the memory cycle, where construction and retrieval are driven by local heuristics rather than explicit strategic reasoning, and sparse, delayed supervision on the backward path, where downstream failures rarely translate into direct repairs of the memory bank. To address these challenges, we propose MemMA, a plug-and-play multi-agent framework that coordinates the memory cycle along both the forward and backward paths. On the forward path, a Meta-Thinker produces structured guidance that steers a Memory Manager during construction and directs a Query Reasoner during iterative retrieval. On the backward path, MemMA introduces in-situ self-evolving memory construction, which synthesizes probe QA pairs, verifies the current memory, and converts failures into repair actions before the memory is finalized. Extensive experiments on LoCoMo show that MemMA consistently outperforms existing baselines across multiple LLM backbones and improves three different storage backends in a plug-and-play manner. Our code is publicly available at https://github.com/ventr1c/memma.

Tianxiang Zhao, Dongsheng Luo, Xiang Zhang et al.

Uncovering rationales behind predictions of graph neural networks (GNNs) has received increasing attention over recent years. Instance-level GNN explanation aims to discover critical input elements, like nodes or edges, that the target GNN relies upon for making predictions. %These identified sub-structures can provide interpretations of GNN's behavior. Though various algorithms are proposed, most of them formalize this task by searching the minimal subgraph which can preserve original predictions. However, an inductive bias is deep-rooted in this framework: several subgraphs can result in the same or similar outputs as the original graphs. Consequently, they have the danger of providing spurious explanations and failing to provide consistent explanations. Applying them to explain weakly-performed GNNs would further amplify these issues. To address this problem, we theoretically examine the predictions of GNNs from the causality perspective. Two typical reasons for spurious explanations are identified: confounding effect of latent variables like distribution shift, and causal factors distinct from the original input. Observing that both confounding effects and diverse causal rationales are encoded in internal representations, \tianxiang{we propose a new explanation framework with an auxiliary alignment loss, which is theoretically proven to be optimizing a more faithful explanation objective intrinsically. Concretely for this alignment loss, a set of different perspectives are explored: anchor-based alignment, distributional alignment based on Gaussian mixture models, mutual-information-based alignment, etc. A comprehensive study is conducted both on the effectiveness of this new framework in terms of explanation faithfulness/consistency and on the advantages of these variants.

Yihe Wang, Taida Li, Yujun Yan et al.

Cross-subject EEG classification typically achieves significantly lower performance than subject-dependent settings. Although this phenomenon has been widely observed in the literature, the underlying causes have not been systematically studied. In this paper, we design a series of controlled experiments to investigate the mechanisms behind the performance drop in cross-subject EEG classification across different EEG tasks. We show that the performance degradation can generally be attributed to two factors: inter-subject variability and shortcut learning. Specifically, multi-class-per-subject EEG classification tasks, such as motor imagery, emotion recognition, and ERP stimulus classification, are mainly affected by inter-subject variability, whereas single-class-per-subject EEG classification tasks, such as brain disease detection, are primarily influenced by shortcut learning based on subject-specific features. These findings provide new insights into the challenges of cross-subject EEG classification and emphasize the importance of appropriate evaluation protocols in EEG research. The code is available at https://github.com/DL4mHealth/EEG-Cross-Subject.

Yihe Wang, Zhiqiao Kang, Bohan Chen et al.

Event-related potential (ERP), a specialized paradigm of electroencephalographic (EEG), reflects neurological responses to external stimuli or events, generally associated with the brain's processing of specific cognitive tasks. ERP plays a critical role in cognitive analysis, the detection of neurological diseases, and the assessment of psychological states. Recent years have seen substantial advances in deep learning-based methods for spontaneous EEG and other non-time-locked task-related EEG signals. However, their effectiveness on ERP data remains underexplored, and many existing ERP studies still rely heavily on manually extracted features. In this paper, we conduct a comprehensive benchmark study that systematically compares traditional manual features (followed by a linear classifier), deep learning models, and pre-trained EEG foundation models for ERP analysis. We establish a unified data preprocessing and training pipeline and evaluate these approaches on two representative tasks, ERP stimulus classification and ERP-based brain disease detection, across 12 publicly available datasets. Furthermore, we investigate various token-embedding strategies within advanced Transformer architectures to identify embedding designs that better suit ERP data. Our study provides a landmark framework to guide method selection and tailored model design for future ERP analysis. The code is available at https://github.com/DL4mHealth/ERP-Benchmark

Tianxiang Zhao, Xiang Zhang, Suhang Wang

In recent years, graph neural networks (GNNs) have achieved state-of-the-art performance for node classification. However, most existing GNNs would suffer from the graph imbalance problem. In many real-world scenarios, node classes are imbalanced, with some majority classes making up most parts of the graph. The message propagation mechanism in GNNs would further amplify the dominance of those majority classes, resulting in sub-optimal classification performance. In this work, we seek to address this problem by generating pseudo instances of minority classes to balance the training data, extending previous over-sampling-based techniques. This task is non-trivial, as those techniques are designed with the assumption that instances are independent. Neglection of relation information would complicate this oversampling process. Furthermore, the node classification task typically takes the semi-supervised setting with only a few labeled nodes, providing insufficient supervision for the generation of minority instances. Generated new nodes of low quality would harm the trained classifier. In this work, we address these difficulties by synthesizing new nodes in a constructed embedding space, which encodes both node attributes and topology information. Furthermore, an edge generator is trained simultaneously to model the graph structure and provide relations for new samples. To further improve the data efficiency, we also explore synthesizing mixed ``in-between'' nodes to utilize nodes from the majority class in this over-sampling process. Experiments on real-world datasets validate the effectiveness of our proposed framework.

Yangguang Wang, Xiang Zhang, Mingyuan Lin et al.

Scene Dynamic Recovery (SDR) by inverting distorted Rolling Shutter (RS) images to an undistorted high frame-rate Global Shutter (GS) video is a severely ill-posed problem due to the missing temporal dynamic information in both RS intra-frame scanlines and inter-frame exposures, particularly when prior knowledge about camera/object motions is unavailable. Commonly used artificial assumptions on scenes/motions and data-specific characteristics are prone to producing sub-optimal solutions in real-world scenarios. To address this challenge, we propose an event-based SDR network within a self-supervised learning paradigm, i.e., SelfUnroll. We leverage the extremely high temporal resolution of event cameras to provide accurate inter/intra-frame dynamic information. Specifically, an Event-based Inter/intra-frame Compensator (E-IC) is proposed to predict the per-pixel dynamic between arbitrary time intervals, including the temporal transition and spatial translation. Exploring connections in terms of RS-RS, RS-GS, and GS-RS, we explicitly formulate mutual constraints with the proposed E-IC, resulting in supervisions without ground-truth GS images. Extensive evaluations over synthetic and real datasets demonstrate that the proposed method achieves state-of-the-art and shows remarkable performance for event-based RS2GS inversion in real-world scenarios. The dataset and code are available at https://w3un.github.io/selfunroll/.

Puyuan Liu, Xiang Zhang, Lili Mou

Sentence summarization aims at compressing a long sentence into a short one that keeps the main gist, and has extensive real-world applications such as headline generation. In previous work, researchers have developed various approaches to improve the ROUGE score, which is the main evaluation metric for summarization, whereas controlling the summary length has not drawn much attention. In our work, we address a new problem of explicit character-level length control for summarization, and propose a dynamic programming algorithm based on the Connectionist Temporal Classification (CTC) model. Results show that our approach not only achieves higher ROUGE scores but also yields more complete sentences.

Dongsheng Luo, Yuchen Bian, Yaowei Yan et al.

Random Walk is a basic algorithm to explore the structure of networks, which can be used in many tasks, such as local community detection and network embedding. Existing random walk methods are based on single networks that contain limited information. In contrast, real data often contain entities with different types or/and from different sources, which are comprehensive and can be better modeled by multiple networks. To take advantage of rich information in multiple networks and make better inferences on entities, in this study, we propose random walk on multiple networks, RWM. RWM is flexible and supports both multiplex networks and general multiple networks, which may form many-to-many node mappings between networks. RWM sends a random walker on each network to obtain the local proximity (i.e., node visiting probabilities) w.r.t. the starting nodes. Walkers with similar visiting probabilities reinforce each other. We theoretically analyze the convergence properties of RWM. Two approximation methods with theoretical performance guarantees are proposed for efficient computation. We apply RWM in link prediction, network embedding, and local community detection. Comprehensive experiments conducted on both synthetic and real-world datasets demonstrate the effectiveness and efficiency of RWM.

Chi Zhang, Xiang Zhang, Mingyuan Lin et al.

Even though the collaboration between traditional and neuromorphic event cameras brings prosperity to frame-event based vision applications, the performance is still confined by the resolution gap crossing two modalities in both spatial and temporal domains. This paper is devoted to bridging the gap by increasing the temporal resolution for images, i.e., motion deblurring, and the spatial resolution for events, i.e., event super-resolving, respectively. To this end, we introduce CrossZoom, a novel unified neural Network (CZ-Net) to jointly recover sharp latent sequences within the exposure period of a blurry input and the corresponding High-Resolution (HR) events. Specifically, we present a multi-scale blur-event fusion architecture that leverages the scale-variant properties and effectively fuses cross-modality information to achieve cross-enhancement. Attention-based adaptive enhancement and cross-interaction prediction modules are devised to alleviate the distortions inherent in Low-Resolution (LR) events and enhance the final results through the prior blur-event complementary information. Furthermore, we propose a new dataset containing HR sharp-blurry images and the corresponding HR-LR event streams to facilitate future research. Extensive qualitative and quantitative experiments on synthetic and real-world datasets demonstrate the effectiveness and robustness of the proposed method. Codes and datasets are released at https://bestrivenzc.github.io/CZ-Net/.

Xiang Zhang, Jiaqi Wei, Yuejin Yang et al.

Chain-of-Thought (CoT) prompting has significantly advanced task-solving capabilities in natural language processing with large language models. Unlike standard prompting, CoT encourages the model to generate intermediate reasoning steps, non-answer tokens, that help guide the model toward more accurate final outputs. These intermediate steps enable more complex reasoning processes such as error correction, memory management, future planning, and self-reflection. However, applying CoT to non-natural language domains, such as protein and RNA language models, is not yet possible, primarily due to the limited expressiveness of their token spaces (e.g., amino acid tokens). In this work, we propose and define the concept of language expressiveness: the ability of a given language, using its tokens and grammar, to encode information. We show that the limited expressiveness of protein language severely restricts the applicability of CoT-style reasoning. To overcome this, we introduce reflection pretraining, for the first time in a biological sequence model, which enables the model to engage in intermediate reasoning through the generation of auxiliary "thinking tokens" beyond simple answer tokens. Theoretically, we demonstrate that our augmented token set significantly enhances biological language expressiveness, thereby improving the overall reasoning capacity of the model. Experimentally, our pretraining approach teaches protein models to self-correct and leads to substantial performance gains compared to standard pretraining.

Xiang Zhang, Bingxin Ke, Hayko Riemenschneider et al.

By training over large-scale datasets, zero-shot monocular depth estimation (MDE) methods show robust performance in the wild but often suffer from insufficient detail. Although recent diffusion-based MDE approaches exhibit a superior ability to extract details, they struggle in geometrically complex scenes that challenge their geometry prior, trained on less diverse 3D data. To leverage the complementary merits of both worlds, we propose BetterDepth to achieve geometrically correct affine-invariant MDE while capturing fine details. Specifically, BetterDepth is a conditional diffusion-based refiner that takes the prediction from pre-trained MDE models as depth conditioning, in which the global depth layout is well-captured, and iteratively refines details based on the input image. For the training of such a refiner, we propose global pre-alignment and local patch masking methods to ensure BetterDepth remains faithful to the depth conditioning while learning to add fine-grained scene details. With efficient training on small-scale synthetic datasets, BetterDepth achieves state-of-the-art zero-shot MDE performance on diverse public datasets and on in-the-wild scenes. Moreover, BetterDepth can improve the performance of other MDE models in a plug-and-play manner without further re-training.

Zhangyi Cheng, Xiang Zhang, Lei Yu et al.

This paper aims at demystifying a single motion-blurred image with events and revealing temporally continuous scene dynamics encrypted behind motion blurs. To achieve this end, an Implicit Video Function (IVF) is learned to represent a single motion blurred image with concurrent events, enabling the latent sharp image restoration of arbitrary timestamps in the range of imaging exposures. Specifically, a dual attention transformer is proposed to efficiently leverage merits from both modalities, i.e., the high temporal resolution of event features and the smoothness of image features, alleviating temporal ambiguities while suppressing the event noise. The proposed network is trained only with the supervision of ground-truth images of limited referenced timestamps. Motion- and texture-guided supervisions are employed simultaneously to enhance restorations of the non-referenced timestamps and improve the overall sharpness. Experiments on synthetic, semi-synthetic, and real-world datasets demonstrate that our proposed method outperforms state-of-the-art methods by a large margin in terms of both objective PSNR and SSIM measurements and subjective evaluations.

Yihe Wang, Nadia Mammone, Darina Petrovsky et al.

Electroencephalography (EEG) has emerged as a cost-effective and efficient tool to support neurologists in the detection of Alzheimer's Disease (AD). However, most existing approaches rely heavily on manual feature engineering or data transformation. While such techniques may provide benefits when working with small-scale datasets, they often lead to information loss and distortion when applied to large-scale data, ultimately limiting model performance. Moreover, the limited subject scale and demographic diversity of datasets used in prior studies hinder comprehensive evaluation of model robustness and generalizability, thus restricting their applicability in real-world clinical settings. To address these challenges, we propose ADformer, a novel multi-granularity spatial-temporal transformer designed to capture both temporal and spatial features from raw EEG signals, enabling effective end-to-end representation learning. Our model introduces multi-granularity embedding strategies across both spatial and temporal dimensions, leveraging a two-stage intra-inter granularity self-attention mechanism to learn both local patterns within each granularity and global dependencies across granularities. We evaluate ADformer on 4 large-scale datasets comprising a total of 1,713 subjects, representing one of the largest corpora for EEG-based AD detection to date, under a cross-validated, subject-independent setting. Experimental results demonstrate that ADformer consistently outperforms existing methods, achieving subject-level F1 scores of 92.82%, 89.83%, 67.99%, and 83.98% on the 4 datasets, respectively, in distinguishing AD from healthy control (HC) subjects.

Xiang Zhang, Tengfei Wang, Fang Xu et al.

Visual localization in large-scale UAV scenarios is a critical capability for autonomous systems, yet it remains challenging due to geometric complexity and environmental variations. While 3D Gaussian Splatting (3DGS) has emerged as a promising scene representation, existing 3DGS-based visual localization methods struggle with robust pose initialization and sensitivity to rendering artifacts in large-scale settings. To address these limitations, we propose LSGS-Loc, a novel visual localization pipeline tailored for large-scale 3DGS scenes. Specifically, we introduce a scale-aware pose initialization strategy that combines scene-agnostic relative pose estimation with explicit 3DGS scale constraints, enabling geometrically grounded localization without scene-specific training. Furthermore, in the pose refinement, to mitigate the impact of reconstruction artifacts such as blur and floaters, we develop a Laplacian-based reliability masking mechanism that guides photometric refinement toward high-quality regions. Extensive experiments on large-scale UAV benchmarks demonstrate that our method achieves state-of-the-art accuracy and robustness for unordered image queries, significantly outperforming existing 3DGS-based approaches. Code is available at: https://github.com/xzhang-z/LSGS-Loc

Maolin Luo, Xiang Zhang

Clinical evidence encompasses the associations and impacts between patients, interventions (such as drugs or physiotherapy), problems, and outcomes. The goal of recommending clinical evidence is to provide medical practitioners with relevant information to support their decision-making processes and to generate new evidence. Our specific task focuses on recommending evidence based on clinical problems. However, the direct connections between certain clinical problems and related evidence are often sparse, creating a challenge of link sparsity. Additionally, to recommend appropriate evidence, it is essential to jointly exploit both topological relationships among evidence and textual information describing them. To address these challenges, we define two knowledge graphs: an Evidence Co-reference Graph and an Evidence Text Graph, to represent the topological and linguistic relations among evidential elements, respectively. We also introduce a multi-channel heterogeneous learning model and a fusional attention mechanism to handle the co-reference-text heterogeneity in evidence recommendation. Our experiments demonstrate that our model outperforms state-of-the-art methods on open data.