Kai Jiang

Chen Zhao, Feng Mi, Xintao Wu et al.

The fairness-aware online learning framework has arisen as a powerful tool for the continual lifelong learning setting. The goal for the learner is to sequentially learn new tasks where they come one after another over time and the learner ensures the statistic parity of the new coming task across different protected sub-populations (e.g. race and gender). A major drawback of existing methods is that they make heavy use of the i.i.d assumption for data and hence provide static regret analysis for the framework. However, low static regret cannot imply a good performance in changing environments where tasks are sampled from heterogeneous distributions. To address the fairness-aware online learning problem in changing environments, in this paper, we first construct a novel regret metric FairSAR by adding long-term fairness constraints onto a strongly adapted loss regret. Furthermore, to determine a good model parameter at each round, we propose a novel adaptive fairness-aware online meta-learning algorithm, namely FairSAOML, which is able to adapt to changing environments in both bias control and model precision. The problem is formulated in the form of a bi-level convex-concave optimization with respect to the model's primal and dual parameters that are associated with the model's accuracy and fairness, respectively. The theoretic analysis provides sub-linear upper bounds for both loss regret and violation of cumulative fairness constraints. Our experimental evaluation on different real-world datasets with settings of changing environments suggests that the proposed FairSAOML significantly outperforms alternatives based on the best prior online learning approaches.

Jintao Zhang, Kaiwen Zheng, Kai Jiang et al.

We introduce TurboDiffusion, a video generation acceleration framework that can speed up end-to-end diffusion generation by 100-200x while maintaining video quality. TurboDiffusion mainly relies on several components for acceleration: (1) Attention acceleration: TurboDiffusion uses low-bit SageAttention and trainable Sparse-Linear Attention (SLA) to speed up attention computation. (2) Step distillation: TurboDiffusion adopts rCM for efficient step distillation. (3) W8A8 quantization: TurboDiffusion quantizes model parameters and activations to 8 bits to accelerate linear layers and compress the model. In addition, TurboDiffusion incorporates several other engineering optimizations. We conduct experiments on the Wan2.2-I2V-14B-720P, Wan2.1-T2V-1.3B-480P, Wan2.1-T2V-14B-720P, and Wan2.1-T2V-14B-480P models. Experimental results show that TurboDiffusion achieves 100-200x speedup for video generation even on a single RTX 5090 GPU, while maintaining comparable video quality. The GitHub repository, which includes model checkpoints and easy-to-use code, is available at https://github.com/thu-ml/TurboDiffusion.

Haoliang Wang, Chen Zhao, Yunhui Guo et al.

Two prevalent types of distributional shifts in machine learning are the covariate shift (as observed across different domains) and the semantic shift (as seen across different classes). Traditional OOD detection techniques typically address only one of these shifts. However, real-world testing environments often present a combination of both covariate and semantic shifts. In this study, we introduce a novel problem, semantic OOD detection across domains, which simultaneously addresses both distributional shifts. To this end, we introduce two regularization strategies: domain generalization regularization, which ensures semantic invariance across domains to counteract the covariate shift, and OOD detection regularization, designed to enhance OOD detection capabilities against the semantic shift through energy bounding. Through rigorous testing on three standard domain generalization benchmarks, our proposed framework showcases its superiority over conventional domain generalization approaches in terms of OOD detection performance. Moreover, it holds its ground by maintaining comparable InD classification accuracy.

Jintao Zhang, Haoxu Wang, Kai Jiang et al. · tsinghua

Sparse-Linear Attention (SLA) combines sparse and linear attention to accelerate diffusion models and has shown strong performance in video generation. However, (i) SLA relies on a heuristic split that assigns computations to the sparse or linear branch based on attention-weight magnitude, which can be suboptimal. Additionally, (ii) after formally analyzing the attention error in SLA, we identify a mismatch between SLA and a direct decomposition into sparse and linear attention. We propose SLA2, which introduces (I) a learnable router that dynamically selects whether each attention computation should use sparse or linear attention, (II) a more faithful and direct sparse-linear attention formulation that uses a learnable ratio to combine the sparse and linear attention branches, and (III) a sparse + low-bit attention design, where low-bit attention is introduced via quantization-aware fine-tuning to reduce quantization error. Experiments show that on video diffusion models, SLA2 can achieve 97% attention sparsity and deliver an 18.6x attention speedup while preserving generation quality.

Jintao Zhang, Kai Jiang, Chendong Xiang et al. · tsinghua

Many training-free sparse attention methods are effective for accelerating diffusion models. Recently, several works suggest that making sparse attention trainable can further increase sparsity while preserving generation quality. We study three key questions: (1) when do the two common masking rules, i.e., Top-k and Top-p, fail, and how can we avoid these failures? (2) why can trainable sparse attention reach higher sparsity than training-free methods? (3) what are the limitations of fine-tuning sparse attention using the diffusion loss, and how can we address them? Based on this analysis, we propose SpargeAttention2, a trainable sparse attention method that achieves high sparsity without degrading generation quality. SpargeAttention2 includes (i) a hybrid masking rule that combines Top-k and Top-p for more robust masking at high sparsity, (ii) an efficient trainable sparse attention implementation, and (iii) a distillation-inspired fine-tuning objective to better preserve generation quality during fine-tuning using sparse attention. Experiments on video diffusion models show that SpargeAttention2 reaches 95% attention sparsity and a 16.2x attention speedup while maintaining generation quality, consistently outperforming prior sparse attention methods.

Dan Wei, Tiejun Zhou, Yunqing Huang et al.

In this work, we design a multi-category inverse design neural network to map ordered periodic structure to physical parameters. The neural network model consists of two parts, a classifier and Structure-Parameter-Mapping (SPM) subnets. The classifier is used to identify structure, and the SPM subnets are used to predict physical parameters for desired structures. We also present an extensible reciprocal-space data augmentation method to guarantee the rotation and translation invariant of periodic structures. We apply the proposed network model and data augmentation method to two-dimensional diblock copolymers based on the Landau-Brazovskii model. Results show that the multi-category inverse design neural network is high accuracy in predicting physical parameters for desired structures. Moreover, the idea of multi-categorization can also be extended to other inverse design problems.

Han Wang, Jintao Zhang, Kai Jiang et al.

LLM-based Triton kernel generation has attracted significant interest, yet a fundamental empirical question remains unanswered: where does this capability break down, and why? We present KernelBench-X, a benchmark designed to answer this question through category-aware evaluation of correctness and hardware efficiency across 176 tasks in 15 categories. Our systematic comparison of five representative methods yields three main findings. First, task structure determines correctness more than method design. Category explains nearly three times more variance in semantic correctness than method (9.4% vs 3.3% explained deviance), and 72% of Fusion tasks fail across all five methods while Math tasks are solved consistently. Second, iterative refinement improves correctness, but not performance. Across GEAK iterations, compile rate rises from 52.3% to 68.8% while average speedup declines from $1.58\times$ to $1.44\times$; newly rescued kernels consistently underperform persistently correct ones ($1.16\times$ vs $1.58\times$ speedup in round~0$\to$1). Third, correctness does not imply efficiency. 46.6% of correct kernels are slower than the PyTorch eager baseline, and cross-hardware speedup variance reaches $21.4\times$. Besides, quantization remains completely unsolved (0/30 successes) despite non-trivial compilation rates, revealing systematic misunderstanding of numerical computation contracts rather than surface-level syntax errors. These findings suggest that future progress depends on handling global coordination, explicitly modeling numerical precision, and incorporating hardware efficiency into generation. The code is available at https://github.com/BonnieW05/KernelBenchX

Jintao Zhang, Marco Chen, Haoxu Wang et al.

Low-bit attention, such as SageAttention, has emerged as an effective approach for accelerating model inference, but its applicability to training remains poorly understood. In prior work, we introduced SageBwd, a trainable INT8 attention that quantizes six of seven attention matrix multiplications while preserving fine-tuning performance. However, SageBwd exhibited a persistent performance gap to full-precision attention (FPA) during pre-training. In this work, we investigate why this gap occurs and demonstrate that SageBwd matches full-precision attention during pretraining. Through experiments and theoretical analysis, we reach a few important insights and conclusions: (i) QK-norm is necessary for stable training at large tokens per step, (ii) quantization errors primarily arise from the backward-pass score gradient dS, (iii) reducing tokens per step enables SageBwd to match FPA performance in pre-training, and (iv) K-smoothing remains essential for training stability, while Q-smoothing provides limited benefit during pre-training.

Chen Zhao, Kai Jiang, Xintao Wu et al.

The endeavor to preserve the generalization of a fair and invariant classifier across domains, especially in the presence of distribution shifts, becomes a significant and intricate challenge in machine learning. In response to this challenge, numerous effective algorithms have been developed with a focus on addressing the problem of fairness-aware domain generalization. These algorithms are designed to navigate various types of distribution shifts, with a particular emphasis on covariate and dependence shifts. In this context, covariate shift pertains to changes in the marginal distribution of input features, while dependence shift involves alterations in the joint distribution of the label variable and sensitive attributes. In this paper, we introduce a simple but effective approach that aims to learn a fair and invariant classifier by simultaneously addressing both covariate and dependence shifts across domains. We assert the existence of an underlying transformation model can transform data from one domain to another, while preserving the semantics related to non-sensitive attributes and classes. By augmenting various synthetic data domains through the model, we learn a fair and invariant classifier in source domains. This classifier can then be generalized to unknown target domains, maintaining both model prediction and fairness concerns. Extensive empirical studies on four benchmark datasets demonstrate that our approach surpasses state-of-the-art methods.

Kai Jiang, Jiaqi Tang, Qilong Zhai et al.

Quasiperiodic elliptic operators (QEOs) serve as fundamental models in both mathematics and physics, as exemplified by their role in the numerical modeling of one-dimensional photonic quasicrystals. However, distinct from periodic elliptic operators, approximating eigenpairs for QEOs poses significant challenges, particularly in capturing the full spectral structure (notably the continuous spectrum) and deriving convergence guarantees in the absence of compactness. In this paper, we develop a high-accuracy numerical method to compute eigenpairs of QEOs based on the projection method, which embeds quasiperiodic operators into a higher-dimensional periodic torus. To address the non-compactness issue, we construct a directional-derivative Hilbert space along irrational manifolds of a high-dimensional torus and characterize operators equivalent to QEOs within this space. By integrating spectral theory for non-compact operators into the Babuška-Osborn eigenproblem framework, we establish rigorous convergence analysis and prove that our method achieves spectral accuracy. Numerical experiments validate the accuracy and efficiency of the proposed method, including a one-dimensional photonic quasicrystal and two- and three-dimensional QEOs.

Kai Jiang, Honghao Yang, Yuexian Wang et al.

The mental health assessment of middle school students has always been one of the focuses in the field of education. This paper introduces a new ensemble learning network based on BERT, employing the concept of enhancing model performance by integrating multiple classifiers. We trained a range of BERT-based learners, which combined using the majority voting method. We collect social network text data of middle school students through China's Weibo and apply the method to the task of classifying emotional tendencies in middle school students' social network texts. Experimental results suggest that the ensemble learning network has a better performance than the base model and the performance of the ensemble learning model, consisting of three single-layer BERT models, is barely the same as a three-layer BERT model but requires 11.58% more training time. Therefore, in terms of balancing prediction effect and efficiency, the deeper BERT network should be preferred for training. However, for interpretability, network ensembles can provide acceptable solutions.

Kai Jiang, Zhengyan Shi, Dell Zhang et al.

Class Incremental Learning (CIL) aims to continuously learn new categories while retaining the knowledge of old ones. Pre-trained models (PTMs) show promising capabilities in CIL. However, existing approaches that apply lightweight fine-tuning to backbones still induce parameter drift, thereby compromising the generalization capability of pre-trained models. Parameter drift can be conceptualized as a form of noise that obscures critical patterns learned for previous tasks. However, recent researches have shown that noise is not always harmful. For example, the large number of visual patterns learned from pre-training can be easily abused by a single task, and introducing appropriate noise can suppress some low-correlation features, thus leaving a margin for future tasks. To this end, we propose learning beneficial noise for CIL guided by information theory and propose Mixture of Noise (Min), aiming to mitigate the degradation of backbone generalization from adapting new tasks. Specifically, task-specific noise is learned from high-dimension features of new tasks. Then, a set of weights is adjusted dynamically for optimal mixture of different task noise. Finally, Min embeds the beneficial noise into the intermediate features to mask the response of inefficient patterns. Extensive experiments on six benchmark datasets demonstrate that Min achieves state-of-the-art performance in most incremental settings, with particularly outstanding results in 50-steps incremental settings. This shows the significant potential for beneficial noise in continual learning. Code is available at https://github.com/ASCIIJK/MiN-NeurIPS2025.

Huayu Chen, Kai Jiang, Kaiwen Zheng et al.

Classifier-Free Guidance (CFG) has been a default technique in various visual generative models, yet it requires inference from both conditional and unconditional models during sampling. We propose to build visual models that are free from guided sampling. The resulting algorithm, Guidance-Free Training (GFT), matches the performance of CFG while reducing sampling to a single model, halving the computational cost. Unlike previous distillation-based approaches that rely on pretrained CFG networks, GFT enables training directly from scratch. GFT is simple to implement. It retains the same maximum likelihood objective as CFG and differs mainly in the parameterization of conditional models. Implementing GFT requires only minimal modifications to existing codebases, as most design choices and hyperparameters are directly inherited from CFG. Our extensive experiments across five distinct visual models demonstrate the effectiveness and versatility of GFT. Across domains of diffusion, autoregressive, and masked-prediction modeling, GFT consistently achieves comparable or even lower FID scores, with similar diversity-fidelity trade-offs compared with CFG baselines, all while being guidance-free. Code will be available at https://github.com/thu-ml/GFT.

Jintao Zhang, Haoxu Wang, Kai Jiang et al. · tsinghua

In Diffusion Transformer (DiT) models, particularly for video generation, attention latency is a major bottleneck due to the long sequence length and the quadratic complexity. We find that attention weights can be separated into two parts: a small fraction of large weights with high rank and the remaining weights with very low rank. This naturally suggests applying sparse acceleration to the first part and low-rank acceleration to the second. Based on this finding, we propose SLA (Sparse-Linear Attention), a trainable attention method that fuses sparse and linear attention to accelerate diffusion models. SLA classifies attention weights into critical, marginal, and negligible categories, applying O(N^2) attention to critical weights, O(N) attention to marginal weights, and skipping negligible ones. SLA combines these computations into a single GPU kernel and supports both forward and backward passes. With only a few fine-tuning steps using SLA, DiT models achieve a 20x reduction in attention computation, resulting in significant acceleration without loss of generation quality. Experiments show that SLA reduces attention computation by 95% without degrading end-to-end generation quality, outperforming baseline methods. In addition, we implement an efficient GPU kernel for SLA, which yields a 13.7x speedup in attention computation and a 2.2x end-to-end speedup in video generation on Wan2.1-1.3B. The code is available at https://github.com/thu-ml/SLA.

Jiaqing Zhang, Jie Lei, Weiying Xie et al.

Accurate cloud recognition and warning are crucial for various applications, including in-flight support, weather forecasting, and climate research. However, recent deep learning algorithms have predominantly focused on detecting cloud regions in satellite imagery, with insufficient attention to the specificity required for accurate cloud recognition. This limitation inspired us to develop the novel FY-4A-Himawari-8 (FYH) dataset, which includes nine distinct cloud categories and uses precise domain adaptation methods to align 70,419 image-label pairs in terms of projection, temporal resolution, and spatial resolution, thereby facilitating the training of supervised deep learning networks. Given the complexity and diversity of cloud formations, we have thoroughly analyzed the challenges inherent to cloud recognition tasks, examining the intricate characteristics and distribution of the data. To effectively address these challenges, we designed a Distribution-aware Interactive-Attention Network (DIAnet), which preserves pixel-level details through a high-resolution branch and a parallel multi-resolution cross-branch. We also integrated a distribution-aware loss (DAL) to mitigate the imbalance across cloud categories. An Interactive Attention Module (IAM) further enhances the robustness of feature extraction combined with spatial and channel information. Empirical evaluations on the FYH dataset demonstrate that our method outperforms other cloud recognition networks, achieving superior performance in terms of mean Intersection over Union (mIoU). The code for implementing DIAnet is available at https://github.com/icey-zhang/DIAnet.

Mingxiang Cao, Weiying Xie, Xin Zhang et al.

Multi-modal fusion holds great promise for integrating information from different modalities. However, due to a lack of consideration for modal consistency, existing multi-modal fusion methods in the field of remote sensing still face challenges of incomplete semantic information and low computational efficiency in their fusion designs. Inspired by the observation that the visual language pre-training model CLIP can effectively extract strong semantic information from visual features, we propose M$^3$amba, a novel end-to-end CLIP-driven Mamba model for multi-modal fusion to address these challenges. Specifically, we introduce CLIP-driven modality-specific adapters in the fusion architecture to avoid the bias of understanding specific domains caused by direct inference, making the original CLIP encoder modality-specific perception. This unified framework enables minimal training to achieve a comprehensive semantic understanding of different modalities, thereby guiding cross-modal feature fusion. To further enhance the consistent association between modality mappings, a multi-modal Mamba fusion architecture with linear complexity and a cross-attention module Cross-SS2D are designed, which fully considers effective and efficient information interaction to achieve complete fusion. Extensive experiments have shown that M$^3$amba has an average performance improvement of at least 5.98\% compared with the state-of-the-art methods in multi-modal hyperspectral image classification tasks in the remote sensing field, while also demonstrating excellent training efficiency, achieving a double improvement in accuracy and efficiency. The code is released at https://github.com/kaka-Cao/M3amba.

Jiaqing Zhang, Mingxiang Cao, Xue Yang et al.

Visual language models like Contrastive Language-Image Pretraining (CLIP) have shown impressive performance in analyzing natural images with language information. However, these models often encounter challenges when applied to specialized domains such as remote sensing due to the limited availability of image-text pairs for training. To tackle this issue, we introduce DiffCLIP, a novel framework that extends CLIP to effectively convey comprehensive language-driven semantic information for accurate classification of high-dimensional multimodal remote sensing images. DiffCLIP is a few-shot learning method that leverages unlabeled images for pretraining. It employs unsupervised mask diffusion learning to capture the distribution of diverse modalities without requiring labels. The modality-shared image encoder maps multimodal data into a unified subspace, extracting shared features with consistent parameters across modalities. A well-trained image encoder further enhances learning by aligning visual representations with class-label text information from CLIP. By integrating these approaches, DiffCLIP significantly boosts CLIP performance using a minimal number of image-text pairs. We evaluate DiffCLIP on widely used high-dimensional multimodal datasets, demonstrating its effectiveness in addressing few-shot annotated classification tasks. DiffCLIP achieves an overall accuracy improvement of 10.65% across three remote sensing datasets compared with CLIP, while utilizing only 2-shot image-text pairs. The code has been released at https://github.com/icey-zhang/DiffCLIP.

Tianxin Xie, Wentao Lei, Guanjie Huang et al.

Text-to-audio-video (T2AV) generation underpins a wide range of applications demanding realistic audio-visual content, including virtual reality, world modeling, gaming, and filmmaking. However, existing T2AV models remain incapable of generating physically plausible sounds, primarily due to their limited understanding of physical principles. To situate current research progress, we present PhyAVBench, a challenging audio physics-sensitivity benchmark designed to systematically evaluate the audio physics grounding capabilities of existing T2AV models. PhyAVBench comprises 1,000 groups of paired text prompts with controlled physical variables that implicitly induce sound variations, enabling a fine-grained assessment of models' sensitivity to changes in underlying acoustic conditions. We term this evaluation paradigm the Audio-Physics Sensitivity Test (APST). Unlike prior benchmarks that primarily focus on audio-video synchronization, PhyAVBench explicitly evaluates models' understanding of the physical mechanisms underlying sound generation, covering 6 major audio physics dimensions, 4 daily scenarios (music, sound effects, speech, and their mix), and 50 fine-grained test points, ranging from fundamental aspects such as sound diffraction to more complex phenomena, e.g., Helmholtz resonance. Each test point consists of multiple groups of paired prompts, where each prompt is grounded by at least 20 newly recorded or collected real-world videos, thereby minimizing the risk of data leakage during model pre-training. Both prompts and videos are iteratively refined through rigorous human-involved error correction and quality control to ensure high quality. We argue that only models with a genuine grasp of audio-related physical principles can generate physically consistent audio-visual content. We hope PhyAVBench will stimulate future progress in this critical yet largely unexplored domain.

Yunsheng Pang, Zijian Liu, Yudong Li et al.

Slate recommendation, which presents users with a ranked item list in a single display, is ubiquitous across mainstream online platforms. Recent advances in generative models have shown significant potential for this task via autoregressive modeling of discrete semantic ID sequences. However, existing methods suffer from three key limitations: entangled item tokenization, inefficient sequential decoding, and the absence of holistic slate planning. These issues often result in substantial inference overhead and inadequate alignment with diverse user preferences and practical business requirements, hindering the industrial deployment of generative slate recommendation systems. In this paper, we propose HiGR, an efficient generative slate recommendation framework that integrates hierarchical planning with listwise preference alignment. First, we design an auto-encoder incorporating residual quantization and contrastive constraints, which tokenizes items into semantically structured IDs to enable controllable generation. Second, HiGR decouples the generation process into two stages: a list-level planning stage to capture global slate intent, and an item-level decoding stage to select specific items, effectively reducing the search space and enabling efficient generation. Third, we introduce a multi-objective and listwise preference alignment mechanism that enhances slate quality by leveraging implicit user feedback. Extensive experiments have validated the effectiveness of our HiGR method. Notably, it outperforms state-of-the-art baselines by over 10\% in offline recommendation quality while achieving a $5\times$ inference speedup. Furthermore, we have deployed HiGR on a commercial platform under Tencent (serving hundreds of millions of users), and online A/B tests show that it increases average watch time and average video plays by 1.22\% and 1.73\%, respectively.

Qun Qiu, Wei Si, Guanghua Ji et al.

In this paper, we develop a set of efficient methods to compute stationary states of the spherical Landau-Brazovskii (LB) model in a discretization-then-optimization way. First, we discretize the spherical LB energy functional into a finite-dimensional energy function by the spherical harmonic expansion. Then five optimization methods are developed to compute stationary states of the discretized energy function, including the accelerated adaptive Bregman proximal gradient, Nesterov, adaptive Nesterov, adaptive nonlinear conjugate gradient and adaptive gradient descent methods. To speed up the convergence, we propose a principal mode analysis (PMA) method to estimate good initial configurations and sphere radius. The PMA method also reveals the relationship between the optimal sphere radius and the dominant degree of spherical harmonics. Numerical experiments show that our approaches significantly reduce the number of iterations and the computational time

Chenglong Bao, Kai Deng, Kai Jiang et al.

We propose an implicit-explicit trust region method for computing second-order stationary points of a class of Landau-type free energy functionals, which correspond to physically (meta-)stable phases. The proposed method is demonstrated through the Landau-Brazovskii (LB) model in this work, while broader applicability to more Landau models of the similar type is straightforwardly extended. The LB energy functional is discretized via the Fourier pseudospectral method, which yields a finite-dimensional nonconvex optimization problem. By exploiting the Hessian structure, specifically, that the interaction potential is diagonal in reciprocal space whereas the bulk energy is diagonal in physical space, we design an adaptive implicit-explicit solver for the trust region subproblem. This solver utilizes the fast Fourier transform to perform efficient matrix-vector products, significantly reducing computational complexity while ensuring provable convergence to the global minimizer of the subproblem. In contrast to existing algorithms that target first-order stationary points, our proposed method can converge to a second-order stationary state, corresponding to a local minimum with theoretical convergence guarantees. Numerical experiments on the LB model demonstrate that the proposed approach efficiently escapes saddle points and significantly outperforms existing first-order schemes. Furthermore, we successfully identify the stable region of the FDDD phase, a structure previously unreported in the LB phase diagram.

Jiaxing Huang, Jingyi Zhang, Kai Jiang et al.

Traditional computer vision generally solves each single task independently by a dedicated model with the task instruction implicitly designed in the model architecture, arising two limitations: (1) it leads to task-specific models, which require multiple models for different tasks and restrict the potential synergies from diverse tasks; (2) it leads to a pre-defined and fixed model interface that has limited interactivity and adaptability in following user' task instructions. To address them, Visual Instruction Tuning (VIT) has been intensively studied recently, which finetunes a large vision model with language as task instructions, aiming to learn from a wide range of vision tasks described by language instructions a general-purpose multimodal model that can follow arbitrary instructions and thus solve arbitrary tasks specified by the user. This work aims to provide a systematic review of visual instruction tuning, covering (1) the background that presents computer vision task paradigms and the development of VIT; (2) the foundations of VIT that introduce commonly used network architectures, visual instruction tuning frameworks and objectives, and evaluation setups and tasks; (3) the commonly used datasets in visual instruction tuning and evaluation; (4) the review of existing VIT methods that categorizes them with a taxonomy according to both the studied vision task and the method design and highlights the major contributions, strengths, and shortcomings of them; (5) the comparison and discussion of VIT methods over various instruction-following benchmarks; (6) several challenges, open directions and possible future works in visual instruction tuning research.

Jiaxing Huang, Kai Jiang, Jingyi Zhang et al.

Segment Anything Models (SAMs) like SEEM and SAM have demonstrated great potential in learning to segment anything. The core design of SAMs lies with Promptable Segmentation, which takes a handcrafted prompt as input and returns the expected segmentation mask. SAMs work with two types of prompts including spatial prompts (e.g., points) and semantic prompts (e.g., texts), which work together to prompt SAMs to segment anything on downstream datasets. Despite the important role of prompts, how to acquire suitable prompts for SAMs is largely under-explored. In this work, we examine the architecture of SAMs and identify two challenges for learning effective prompts for SAMs. To this end, we propose spatial-semantic prompt learning (SSPrompt) that learns effective semantic and spatial prompts for better SAMs. Specifically, SSPrompt introduces spatial prompt learning and semantic prompt learning, which optimize spatial prompts and semantic prompts directly over the embedding space and selectively leverage the knowledge encoded in pre-trained prompt encoders. Extensive experiments show that SSPrompt achieves superior image segmentation performance consistently across multiple widely adopted datasets.

Hui Lin, Danfeng Hong, Shuhang Ge et al.

Remote Sensing Image Captioning (RSIC) presents unique challenges and plays a critical role in applications. Traditional RSIC methods often struggle to produce rich and diverse descriptions. Recently, with advancements in VLMs, efforts have emerged to integrate these models into the remote sensing domain and to introduce descriptive datasets specifically designed to enhance VLM training. This paper proposes RS-MoE, a first Mixture of Expert based VLM specifically customized for remote sensing domain. Unlike traditional MoE models, the core of RS-MoE is the MoE Block, which incorporates a novel Instruction Router and multiple lightweight Large Language Models (LLMs) as expert models. The Instruction Router is designed to generate specific prompts tailored for each corresponding LLM, guiding them to focus on distinct aspects of the RSIC task. This design not only allows each expert LLM to concentrate on a specific subset of the task, thereby enhancing the specificity and accuracy of the generated captions, but also improves the scalability of the model by facilitating parallel processing of sub-tasks. Additionally, we present a two-stage training strategy for tuning our RS-MoE model to prevent performance degradation due to sparsity. We fine-tuned our model on the RSICap dataset using our proposed training strategy. Experimental results on the RSICap dataset, along with evaluations on other traditional datasets where no additional fine-tuning was applied, demonstrate that our model achieves state-of-the-art performance in generating precise and contextually relevant captions. Notably, our RS-MoE-1B variant achieves performance comparable to 13B VLMs, demonstrating the efficiency of our model design. Moreover, our model demonstrates promising generalization capabilities by consistently achieving state-of-the-art performance on the Remote Sensing Visual Question Answering (RSVQA) task.

Chen Zhao, Feng Mi, Xintao Wu et al.

The fairness-aware online learning framework has emerged as a potent tool within the context of continuous lifelong learning. In this scenario, the learner's objective is to progressively acquire new tasks as they arrive over time, while also guaranteeing statistical parity among various protected sub-populations, such as race and gender, when it comes to the newly introduced tasks. A significant limitation of current approaches lies in their heavy reliance on the i.i.d (independent and identically distributed) assumption concerning data, leading to a static regret analysis of the framework. Nevertheless, it's crucial to note that achieving low static regret does not necessarily translate to strong performance in dynamic environments characterized by tasks sampled from diverse distributions. In this paper, to tackle the fairness-aware online learning challenge in evolving settings, we introduce a unique regret measure, FairSAR, by incorporating long-term fairness constraints into a strongly adapted loss regret framework. Moreover, to determine an optimal model parameter at each time step, we introduce an innovative adaptive fairness-aware online meta-learning algorithm, referred to as FairSAOML. This algorithm possesses the ability to adjust to dynamic environments by effectively managing bias control and model accuracy. The problem is framed as a bi-level convex-concave optimization, considering both the model's primal and dual parameters, which pertain to its accuracy and fairness attributes, respectively. Theoretical analysis yields sub-linear upper bounds for both loss regret and the cumulative violation of fairness constraints. Our experimental evaluation on various real-world datasets in dynamic environments demonstrates that our proposed FairSAOML algorithm consistently outperforms alternative approaches rooted in the most advanced prior online learning methods.

Kai Jiang, Chen Zhao, Haoliang Wang et al.

Generalizing to out-of-distribution data while being aware of model fairness is a significant and challenging problem in meta-learning. The goal of this problem is to find a set of fairness-aware invariant parameters of classifier that is trained using data drawn from a family of related training domains with distribution shift on non-sensitive features as well as different levels of dependence between model predictions and sensitive features so that the classifier can achieve good generalization performance on unknown but distinct test domains. To tackle this challenge, existing state-of-the-art methods either address the domain generalization problem but completely ignore learning with fairness or solely specify shifted domains with various fairness levels. This paper introduces an approach to fairness-aware meta-learning that significantly enhances domain generalization capabilities. Our framework, Fairness-Enhanced Meta-Learning for Domain Generalization (FEED), disentangles latent data representations into content, style, and sensitive vectors. This disentanglement facilitates the robust generalization of machine learning models across diverse domains while adhering to fairness constraints. Unlike traditional methods that focus primarily on domain invariance or sensitivity to shifts, our model integrates a fairness-aware invariance criterion directly into the meta-learning process. This integration ensures that the learned parameters uphold fairness consistently, even when domain characteristics vary widely. We validate our approach through extensive experiments across multiple benchmarks, demonstrating not only superior performance in maintaining high accuracy and fairness but also significant improvements over existing state-of-the-art methods in domain generalization tasks.

Yidan Liu, Weiying Xie, Kai Jiang et al.

The majority of existing hyperspectral anomaly detection (HAD) methods use the low-rank representation (LRR) model to separate the background and anomaly components, where the anomaly component is optimized by handcrafted sparse priors (e.g., $\ell_{2,1}$-norm). However, this may not be ideal since they overlook the spatial structure present in anomalies and make the detection result largely dependent on manually set sparsity. To tackle these problems, we redefine the optimization criterion for the anomaly component in the LRR model with a self-supervised network called self-supervised anomaly prior (SAP). This prior is obtained by the pretext task of self-supervised learning, which is customized to learn the characteristics of hyperspectral anomalies. Specifically, this pretext task is a classification task to distinguish the original hyperspectral image (HSI) and the pseudo-anomaly HSI, where the pseudo-anomaly is generated from the original HSI and designed as a prism with arbitrary polygon bases and arbitrary spectral bands. In addition, a dual-purified strategy is proposed to provide a more refined background representation with an enriched background dictionary, facilitating the separation of anomalies from complex backgrounds. Extensive experiments on various hyperspectral datasets demonstrate that the proposed SAP offers a more accurate and interpretable solution than other advanced HAD methods.

Kai Jiang, Jiaxing Huang, Weiying Xie et al.

Camera-only Bird's Eye View (BEV) has demonstrated great potential in environment perception in a 3D space. However, most existing studies were conducted under a supervised setup which cannot scale well while handling various new data. Unsupervised domain adaptive BEV, which effective learning from various unlabelled target data, is far under-explored. In this work, we design DA-BEV, the first domain adaptive camera-only BEV framework that addresses domain adaptive BEV challenges by exploiting the complementary nature of image-view features and BEV features. DA-BEV introduces the idea of query into the domain adaptation framework to derive useful information from image-view and BEV features. It consists of two query-based designs, namely, query-based adversarial learning (QAL) and query-based self-training (QST), which exploits image-view features or BEV features to regularize the adaptation of the other. Extensive experiments show that DA-BEV achieves superior domain adaptive BEV perception performance consistently across multiple datasets and tasks such as 3D object detection and 3D scene segmentation.

Kai Jiang, Jiaxing Huang, Weiying Xie et al.

Large-vocabulary object detectors (LVDs) aim to detect objects of many categories, which learn super objectness features and can locate objects accurately while applied to various downstream data. However, LVDs often struggle in recognizing the located objects due to domain discrepancy in data distribution and object vocabulary. At the other end, recent vision-language foundation models such as CLIP demonstrate superior open-vocabulary recognition capability. This paper presents KGD, a Knowledge Graph Distillation technique that exploits the implicit knowledge graphs (KG) in CLIP for effectively adapting LVDs to various downstream domains. KGD consists of two consecutive stages: 1) KG extraction that employs CLIP to encode downstream domain data as nodes and their feature distances as edges, constructing KG that inherits the rich semantic relations in CLIP explicitly; and 2) KG encapsulation that transfers the extracted KG into LVDs to enable accurate cross-domain object classification. In addition, KGD can extract both visual and textual KG independently, providing complementary vision and language knowledge for object localization and object classification in detection tasks over various downstream domains. Experiments over multiple widely adopted detection benchmarks show that KGD outperforms the state-of-the-art consistently by large margins.

Kaiwen Zheng, Huayu Chen, Haotian Ye et al.

Online reinforcement learning (RL) has been central to post-training language models, but its extension to diffusion models remains challenging due to intractable likelihoods. Recent works discretize the reverse sampling process to enable GRPO-style training, yet they inherit fundamental drawbacks, including solver restrictions, forward-reverse inconsistency, and complicated integration with classifier-free guidance (CFG). We introduce Diffusion Negative-aware FineTuning (DiffusionNFT), a new online RL paradigm that optimizes diffusion models directly on the forward process via flow matching. DiffusionNFT contrasts positive and negative generations to define an implicit policy improvement direction, naturally incorporating reinforcement signals into the supervised learning objective. This formulation enables training with arbitrary black-box solvers, eliminates the need for likelihood estimation, and requires only clean images rather than sampling trajectories for policy optimization. DiffusionNFT is up to $25\times$ more efficient than FlowGRPO in head-to-head comparisons, while being CFG-free. For instance, DiffusionNFT improves the GenEval score from 0.24 to 0.98 within 1k steps, while FlowGRPO achieves 0.95 with over 5k steps and additional CFG employment. By leveraging multiple reward models, DiffusionNFT significantly boosts the performance of SD3.5-Medium in every benchmark tested.

Kai Jiang, Jiaxing Huang

Autoregressive models have demonstrated great performance in natural language processing (NLP) with impressive scalability, adaptability and generalizability. Inspired by their notable success in NLP field, autoregressive models have been intensively investigated recently for computer vision, which perform next-token predictions by representing visual data as visual tokens and enables autoregressive modelling for a wide range of vision tasks, ranging from visual generation and visual understanding to the very recent multimodal generation that unifies visual generation and understanding with a single autoregressive model. This paper provides a systematic review of vision autoregressive models, including the development of a taxonomy of existing methods and highlighting their major contributions, strengths, and limitations, covering various vision tasks such as image generation, video generation, image editing, motion generation, medical image analysis, 3D generation, robotic manipulation, unified multimodal generation, etc. Besides, we investigate and analyze the latest advancements in autoregressive models, including thorough benchmarking and discussion of existing methods across various evaluation datasets. Finally, we outline key challenges and promising directions for future research, offering a roadmap to guide further advancements in vision autoregressive models.

Jiaxing Huang, Jingyi Zhang, Kai Jiang et al.

Recent studies on generalizable object detection have attracted increasing attention with additional weak supervision from large-scale datasets with image-level labels. However, weakly-supervised detection learning often suffers from image-to-box label mismatch, i.e., image-level labels do not convey precise object information. We design Language Hierarchical Self-training (LHST) that introduces language hierarchy into weakly-supervised detector training for learning more generalizable detectors. LHST expands the image-level labels with language hierarchy and enables co-regularization between the expanded labels and self-training. Specifically, the expanded labels regularize self-training by providing richer supervision and mitigating the image-to-box label mismatch, while self-training allows assessing and selecting the expanded labels according to the predicted reliability. In addition, we design language hierarchical prompt generation that introduces language hierarchy into prompt generation which helps bridge the vocabulary gaps between training and testing. Extensive experiments show that the proposed techniques achieve superior generalization performance consistently across 14 widely studied object detection datasets.

Kai Jiang, Siqi Huang, Xiangyu Chen et al.

Continual learning in visual understanding aims to deal with catastrophic forgetting in Multimodal Large Language Models (MLLMs). MLLMs deployed on devices have to continuously adapt to dynamic scenarios in downstream tasks, such as variations in background and perspective, to effectively perform complex visual tasks. To this end, we construct a multimodal visual understanding dataset (MSVQA) encompassing four different scenarios and perspectives including high altitude, underwater, low altitude and indoor, to investigate the catastrophic forgetting in MLLMs under the dynamics of scenario shifts in real-world data streams. Furthermore, we propose mUltimodal coNtInual learning with MLLMs From multi-scenarIo pERspectives (UNIFIER) to address visual discrepancies while learning different scenarios. Specifically, it decouples the visual information from different scenarios into distinct branches within each vision block and projects them into the same feature space. A consistency constraint is imposed on the features of each branch to maintain the stability of visual representations across scenarios. Extensive experiments on the MSVQA dataset demonstrate that UNIFIER effectively alleviates forgetting of cross-scenario tasks and achieves knowledge accumulation within the same scenario.

Chen Zhao, Feng Mi, Xintao Wu et al.

In the problem of online learning for changing environments, data are sequentially received one after another over time, and their distribution assumptions may vary frequently. Although existing methods demonstrate the effectiveness of their learning algorithms by providing a tight bound on either dynamic regret or adaptive regret, most of them completely ignore learning with model fairness, defined as the statistical parity across different sub-population (e.g., race and gender). Another drawback is that when adapting to a new environment, an online learner needs to update model parameters with a global change, which is costly and inefficient. Inspired by the sparse mechanism shift hypothesis, we claim that changing environments in online learning can be attributed to partial changes in learned parameters that are specific to environments and the rest remain invariant to changing environments. To this end, in this paper, we propose a novel algorithm under the assumption that data collected at each time can be disentangled with two representations, an environment-invariant semantic factor and an environment-specific variation factor. The semantic factor is further used for fair prediction under a group fairness constraint. To evaluate the sequence of model parameters generated by the learner, a novel regret is proposed in which it takes a mixed form of dynamic and static regret metrics followed by a fairness-aware long-term constraint. The detailed analysis provides theoretical guarantees for loss regret and violation of cumulative fairness constraints. Empirical evaluations on real-world datasets demonstrate our proposed method sequentially outperforms baseline methods in model accuracy and fairness.

Kai Jiang, Xiangyue Liu, Zheng Ju et al.

Compared with MS-COCO, the dataset for the competition has a larger proportion of large objects which area is greater than 96x96 pixels. As getting fine boundaries is vitally important for large object segmentation, Mask R-CNN with PointRend is selected as the base segmentation framework to output high-quality object boundaries. Besides, a better engine that integrates ResNeSt, FPN and DCNv2, and a range of effective tricks that including multi-scale training and test time augmentation are applied to improve segmentation performance. Our best performance is an ensemble of four models (three PointRend-based models and SOLOv2), which won the 2nd place in IJCAI-PRICAI 3D AI Challenge 2020: Instance Segmentation.

Kai Jiang, XiaoLong Qin

Reinforcement learning usually uses the feedback rewards of environmental to train agents. But the rewards in the actual environment are sparse, and even some environments will not rewards. Most of the current methods are difficult to get good performance in sparse reward or non-reward environments. Although using shaped rewards is effective when solving sparse reward tasks, it is limited to specific problems and learning is also susceptible to local optima. We propose a model-free method that does not rely on environmental rewards to solve the problem of sparse rewards in the general environment. Our method use the minimum number of transitions between states as the distance to replace the rewards of environmental, and proposes a goal-distance gradient to achieve policy improvement. We also introduce a bridge point planning method based on the characteristics of our method to improve exploration efficiency, thereby solving more complex tasks. Experiments show that our method performs better on sparse reward and local optimal problems in complex environments than previous work.