Hongliang Li

Rui Ma, Qingbo Wu, King Ngi Ngan et al.

Recent years have witnessed the great success of blind image quality assessment (BIQA) in various task-specific scenarios, which present invariable distortion types and evaluation criteria. However, due to the rigid structure and learning framework, they cannot apply to the cross-task BIQA scenario, where the distortion types and evaluation criteria keep changing in practical applications. This paper proposes a scalable incremental learning framework (SILF) that could sequentially conduct BIQA across multiple evaluation tasks with limited memory capacity. More specifically, we develop a dynamic parameter isolation strategy to sequentially update the task-specific parameter subsets, which are non-overlapped with each other. Each parameter subset is temporarily settled to Remember one evaluation preference toward its corresponding task, and the previously settled parameter subsets can be adaptively reused in the following BIQA to achieve better performance based on the task relevance. To suppress the unrestrained expansion of memory capacity in sequential tasks learning, we develop a scalable memory unit by gradually and selectively pruning unimportant neurons from previously settled parameter subsets, which enable us to Forget part of previous experiences and free the limited memory capacity for adapting to the emerging new tasks. Extensive experiments on eleven IQA datasets demonstrate that our proposed method significantly outperforms the other state-of-the-art methods in cross-task BIQA. The source code of the proposed method is available at https://github.com/maruiperfect/SILF.

Zhaofeng Shi, Heqian Qiu, Lanxiao Wang et al.

Efficient adaptation between Egocentric (Ego) and Exocentric (Exo) views is crucial for applications such as human-robot cooperation. However, the success of most existing Ego-Exo adaptation methods relies heavily on target-view data for training, thereby increasing computational and data collection costs. In this paper, we make the first exploration of a Test-time Ego-Exo Adaptation for Action Anticipation (TE$^{2}$A$^{3}$) task, which aims to adjust the source-view-trained model online during test time to anticipate target-view actions. It is challenging for existing Test-Time Adaptation (TTA) methods to address this task due to the multi-action candidates and significant temporal-spatial inter-view gap. Hence, we propose a novel Dual-Clue enhanced Prototype Growing Network (DCPGN), which accumulates multi-label knowledge and integrates cross-modality clues for effective test-time Ego-Exo adaptation and action anticipation. Specifically, we propose a Multi-Label Prototype Growing Module (ML-PGM) to balance multiple positive classes via multi-label assignment and confidence-based reweighting for class-wise memory banks, which are updated by an entropy priority queue strategy. Then, the Dual-Clue Consistency Module (DCCM) introduces a lightweight narrator to generate textual clues indicating action progressions, which complement the visual clues containing various objects. Moreover, we constrain the inferred textual and visual logits to construct dual-clue consistency for temporally and spatially bridging Ego and Exo views. Extensive experiments on the newly proposed EgoMe-anti and the existing EgoExoLearn benchmarks show the effectiveness of our method, which outperforms related state-of-the-art methods by a large margin. Code is available at \href{https://github.com/ZhaofengSHI/DCPGN}{https://github.com/ZhaofengSHI/DCPGN}.

Linfeng Xu, Qingbo Wu, Lili Pan et al.

With the rapid development of wearable cameras, a massive collection of egocentric video for first-person visual perception becomes available. Using egocentric videos to predict first-person activity faces many challenges, including limited field of view, occlusions, and unstable motions. Observing that sensor data from wearable devices facilitates human activity recognition, multi-modal activity recognition is attracting increasing attention. However, the deficiency of related dataset hinders the development of multi-modal deep learning for egocentric activity recognition. Nowadays, deep learning in real world has led to a focus on continual learning that often suffers from catastrophic forgetting. But the catastrophic forgetting problem for egocentric activity recognition, especially in the context of multiple modalities, remains unexplored due to unavailability of dataset. In order to assist this research, we present a multi-modal egocentric activity dataset for continual learning named UESTC-MMEA-CL, which is collected by self-developed glasses integrating a first-person camera and wearable sensors. It contains synchronized data of videos, accelerometers, and gyroscopes, for 32 types of daily activities, performed by 10 participants. Its class types and scale are compared with other publicly available datasets. The statistical analysis of the sensor data is given to show the auxiliary effects for different behaviors. And results of egocentric activity recognition are reported when using separately, and jointly, three modalities: RGB, acceleration, and gyroscope, on a base network architecture. To explore the catastrophic forgetting in continual learning tasks, four baseline methods are extensively evaluated with different multi-modal combinations. We hope the UESTC-MMEA-CL can promote future studies on continual learning for first-person activity recognition in wearable applications.

Wenhui Liao, Jiapeng Wang, Hongliang Li et al.

Text-rich document understanding (TDU) requires comprehensive analysis of documents containing substantial textual content and complex layouts. While Multimodal Large Language Models (MLLMs) have achieved fast progress in this domain, existing approaches either demand significant computational resources or struggle with effective multi-modal integration. In this paper, we introduce DocLayLLM, an efficient multi-modal extension of LLMs specifically designed for TDU. By lightly integrating visual patch tokens and 2D positional tokens into LLMs' input and encoding the document content using the LLMs themselves, we fully take advantage of the document comprehension capability of LLMs and enhance their perception of OCR information. We have also deeply considered the role of chain-of-thought (CoT) and innovatively proposed the techniques of CoT Pre-training and CoT Annealing. Our DocLayLLM can achieve remarkable performances with lightweight training settings, showcasing its efficiency and effectiveness. Experimental results demonstrate that our DocLayLLM outperforms existing OCR-dependent methods and OCR-free competitors. Code and model are available at https://github.com/whlscut/DocLayLLM.

Rui Qi, Fengran Mo, Yufeng Chen et al.

Multilingual retrieval-augmented generation (MRAG) requires models to effectively acquire and integrate beneficial external knowledge from multilingual collections. However, most existing studies employ a unitive process where queries of equivalent semantics across different languages are processed through a single-turn retrieval and subsequent optimization. Such a ``one-size-fits-all'' strategy is often suboptimal in multilingual settings, as the models occur to knowledge bias and conflict during the interaction with the search engine. To alleviate the issues, we propose LcRL, a multilingual search-augmented reinforcement learning framework that integrates a language-coupled Group Relative Policy Optimization into the policy and reward models. We adopt the language-coupled group sampling in the rollout module to reduce knowledge bias, and regularize an auxiliary anti-consistency penalty in the reward models to mitigate the knowledge conflict. Experimental results demonstrate that LcRL not only achieves competitive performance but is also appropriate for various practical scenarios such as constrained training data and retrieval over collections encompassing a large number of languages. Our code is available at https://github.com/Cherry-qwq/LcRL-Open.

Heqian Qiu, Hongliang Li, Taijin Zhao et al.

Crowd understanding has aroused the widespread interest in vision domain due to its important practical significance. Unfortunately, there is no effort to explore crowd understanding in multi-modal domain that bridges natural language and computer vision. Referring expression comprehension (REF) is such a representative multi-modal task. Current REF studies focus more on grounding the target object from multiple distinctive categories in general scenarios. It is difficult to applied to complex real-world crowd understanding. To fill this gap, we propose a new challenging dataset, called RefCrowd, which towards looking for the target person in crowd with referring expressions. It not only requires to sufficiently mine the natural language information, but also requires to carefully focus on subtle differences between the target and a crowd of persons with similar appearance, so as to realize the fine-grained mapping from language to vision. Furthermore, we propose a Fine-grained Multi-modal Attribute Contrastive Network (FMAC) to deal with REF in crowd understanding. It first decomposes the intricate visual and language features into attribute-aware multi-modal features, and then captures discriminative but robustness fine-grained attribute features to effectively distinguish these subtle differences between similar persons. The proposed method outperforms existing state-of-the-art (SoTA) methods on our RefCrowd dataset and existing REF datasets. In addition, we implement an end-to-end REF toolbox for the deeper research in multi-modal domain. Our dataset and code can be available at: \url{https://qiuheqian.github.io/datasets/refcrowd/}.

Lei Wang, Qingbo Wu, Desen Yuan et al.

Learning based image quality assessment (IQA) models have obtained impressive performance with the help of reliable subjective quality labels, where mean opinion score (MOS) is the most popular choice. However, in view of the subjective bias of individual annotators, the labor-abundant MOS (LA-MOS) typically requires a large collection of opinion scores from multiple annotators for each image, which significantly increases the learning cost. In this paper, we aim to learn robust IQA models from low-cost MOS (LC-MOS), which only requires very few opinion scores or even a single opinion score for each image. More specifically, we consider the LC-MOS as the noisy observation of LA-MOS and enforce the IQA model learned from LC-MOS to approach the unbiased estimation of LA-MOS. In this way, we represent the subjective bias between LC-MOS and LA-MOS, and the model bias between IQA predictions learned from LC-MOS and LA-MOS (i.e., dual-bias) as two latent variables with unknown parameters. By means of the expectation-maximization based alternating optimization, we can jointly estimate the parameters of the dual-bias, which suppresses the misleading of LC-MOS via a gated dual-bias calibration (GDBC) module. To the best of our knowledge, this is the first exploration of robust IQA model learning from noisy low-cost labels. Theoretical analysis and extensive experiments on four popular IQA datasets show that the proposed method is robust toward different bias rates and annotation numbers and significantly outperforms the other learning based IQA models when only LC-MOS is available. Furthermore, we also achieve comparable performance with respect to the other models learned with LA-MOS.

Hongliang Li, Pingbing Ming, Zhong-ci Shi

We propose two nonconforming finite elements to approximate a boundary value problem arising from strain gradient elasticity, which is a higher-order perturbation of the linearized elastic system. Our elements are H$^2-$nonconforming while H$^1-$conforming. We show both elements converges in the energy norm uniformly with respect to the perturbation parameter.

Yifan Wang, Lin Zhang, Ran Song et al.

Universal domain adaptation (UniDA) aims to transfer the knowledge of common classes from the source domain to the target domain without any prior knowledge on the label set, which requires distinguishing in the target domain the unknown samples from the known ones. Recent methods usually focused on categorizing a target sample into one of the source classes rather than distinguishing known and unknown samples, which ignores the inter-sample affinity between known and unknown samples and may lead to suboptimal performance. Aiming at this issue, we propose a novel UDA framework where such inter-sample affinity is exploited. Specifically, we introduce a knowability-based labeling scheme which can be divided into two steps: 1) Knowability-guided detection of known and unknown samples based on the intrinsic structure of the neighborhoods of samples, where we leverage the first singular vectors of the affinity matrices to obtain the knowability of every target sample. 2) Label refinement based on neighborhood consistency to relabel the target samples, where we refine the labels of each target sample based on its neighborhood consistency of predictions. Then, auxiliary losses based on the two steps are used to reduce the inter-sample affinity between the unknown and the known target samples. Finally, experiments on four public datasets demonstrate that our method significantly outperforms existing state-of-the-art methods.

Shaoxu Cheng, Kanglei Geng, Chiyuan He et al.

Continual Learning (CL) aims to enable Deep Neural Networks (DNNs) to learn new data without forgetting previously learned knowledge. The key to achieving this goal is to avoid confusion at the feature level, i.e., avoiding confusion within old tasks and between new and old tasks. Previous prototype-based CL methods generate pseudo features for old knowledge replay by adding Gaussian noise to the centroids of old classes. However, the distribution in the feature space exhibits anisotropy during the incremental process, which prevents the pseudo features from faithfully reproducing the distribution of old knowledge in the feature space, leading to confusion in classification boundaries within old tasks. To address this issue, we propose the Distribution-Level Memory Recall (DMR) method, which uses a Gaussian mixture model to precisely fit the feature distribution of old knowledge at the distribution level and generate pseudo features in the next stage. Furthermore, resistance to confusion at the distribution level is also crucial for multimodal learning, as the problem of multimodal imbalance results in significant differences in feature responses between different modalities, exacerbating confusion within old tasks in prototype-based CL methods. Therefore, we mitigate the multi-modal imbalance problem by using the Inter-modal Guidance and Intra-modal Mining (IGIM) method to guide weaker modalities with prior information from dominant modalities and further explore useful information within modalities. For the second key, We propose the Confusion Index to quantitatively describe a model's ability to distinguish between new and old tasks, and we use the Incremental Mixup Feature Enhancement (IMFE) method to enhance pseudo features with new sample features, alleviating classification confusion between new and old knowledge.

Wenhui Liao, Hongliang Li, Pengyu Xie et al.

Document parsing is a fundamental task in multimodal understanding, supporting a wide range of downstream applications such as information extraction and intelligent document analysis. Benefiting from strong semantic modeling and robust generalization, VLM-based end-to-end approaches have emerged as the mainstream paradigm in recent years. However, these models often suffer from substantial inference latency, as they must auto-regressively generate long token sequences when processing long-form documents. In this work, motivated by the extremely long outputs and complex layout structures commonly found in document parsing, we propose a training-free and highly efficient acceleration method. Inspired by speculative decoding, we employ a lightweight document parsing pipeline as a draft model to predict batches of future tokens, while the more accurate VLM verifies these draft predictions in parallel. Moreover, we further exploit the layout-structured nature of documents by partitioning each page into independent regions, enabling parallel decoding of each region using the same draft-verify strategy. The final predictions are then assembled according to the natural reading order. Experimental results demonstrate the effectiveness of our approach: on the general-purpose OmniDocBench, our method provides a 2.42x lossless acceleration for the dots.ocr model, and achieves up to 4.89x acceleration on long-document parsing tasks. We will release our code to facilitate reproducibility and future research.

Chiyuan He, Zihuan Qiu, Fanman Meng et al.

Continual learning of pretrained vision-language models (VLMs) is prone to catastrophic forgetting, yet current approaches adapt to new tasks without explicitly preserving the cross-modal semantic geometry inherited from pretraining and previous stages, allowing new-task supervision to induce geometric distortion. We observe that the most pronounced drift tends to concentrate in vulnerable neighborhoods near the old-new semantic interface, where shared visual patterns are easily re-explained by new textual semantics. To address this under an exemplar-free constraint, we propose Semantic Geometry Preservation for Continual Learning (SeGP-CL). SeGP-CL first probes the drift-prone region by constructing a compact set of adversarial anchors with dual-targeted projected gradient descent (DPGD), which drives selected new-task seeds toward old-class semantics while remaining faithful in raw visual space. During training, we preserve cross-modal structure by anchor-guided cross-modal geometry distillation (ACGD), and stabilize the textual reference frame across tasks via a lightweight text semantic-geometry regularization (TSGR). After training, we estimate anchor-induced raw-space drift to transfer old visual prototypes and perform dual-path inference by fusing cross-modal and visual cues. Extensive experiments on five continual learning benchmarks demonstrate that SeGP-CL consistently improves stability and forward transfer, achieving state-of-the-art performance while better preserving semantic geometry of VLMs.

Yilei Qian, Kanglei Geng, Kailong Chen et al.

The application of activity recognition in the "AI + Education" field is gaining increasing attention. However, current work mainly focuses on the recognition of activities in manually captured videos and a limited number of activity types, with little attention given to recognizing activities in surveillance images from real classrooms. In real classroom settings, normal teaching activities such as reading, account for a large proportion of samples, while rare non-teaching activities such as eating, continue to appear. This requires a model that can learn non-teaching activities from few samples without forgetting the normal teaching activities, which necessitates fewshot continual learning (FSCL) capability. To address this gap, we constructed a continual learning dataset focused on classroom surveillance image activity recognition called ARIC (Activity Recognition in Classroom). The dataset has advantages such as multiple perspectives, a wide variety of activities, and real-world scenarios, but it also presents challenges like similar activities and imbalanced sample distribution. To overcome these challenges, we designed a few-shot continual learning method that combines supervised contrastive learning (SCL) and an adaptive covariance classifier (ACC). During the base phase, we proposed a SCL approach based on feature augmentation to enhance the model's generalization ability. In the incremental phase, we employed an ACC to more accurately describe the distribution of new classes. Experimental results demonstrate that our method outperforms other existing methods on the ARIC dataset.

Shuai Chen, Fanman Meng, Chenhao Wu et al.

Few-Shot Segmentation (FSS) aims to segment novel classes using only a few annotated images. Despite considerable progress under pixel-wise support annotation, current FSS methods still face three issues: the inflexibility of backbone upgrade without re-training, the inability to uniformly handle various types of annotations (e.g., scribble, bounding box, mask, and text), and the difficulty in accommodating different annotation quantity. To address these issues simultaneously, we propose DiffUp, a novel framework that conceptualizes the FSS task as a conditional generative problem using a diffusion process. For the first issue, we introduce a backbone-agnostic feature transformation module that converts different segmentation cues into unified coarse priors, facilitating seamless backbone upgrade without re-training. For the second issue, due to the varying granularity of transformed priors from diverse annotation types (scribble, bounding box, mask, and text), we conceptualize these multi-granular transformed priors as analogous to noisy intermediates at different steps of a diffusion model. This is implemented via a self-conditioned modulation block coupled with a dual-level quality modulation branch. For the third issue, we incorporate an uncertainty-aware information fusion module to harmonize the variability across zero-shot, one-shot, and many-shot scenarios. Evaluated through rigorous benchmarks, DiffUp significantly outperforms existing FSS models in terms of flexibility and accuracy.

Xingtao Lin, Heqian Qiu, Lanxiao Wang et al.

Recent advancements in prompt tuning have successfully adapted large-scale models like Contrastive Language-Image Pre-trained (CLIP) for downstream tasks such as scene text detection. Typically, text prompt complements the text encoder's input, focusing on global features while neglecting fine-grained details, leading to fine-grained text being ignored in task of scene text detection. In this paper, we propose the region prompt tuning (RPT) method for fine-grained scene text detection, where region text prompt proposed would help focus on fine-grained features. Region prompt tuning method decomposes region text prompt into individual characters and splits visual feature map into region visual tokens, creating a one-to-one correspondence between characters and tokens. This allows a character matches the local features of a token, thereby avoiding the omission of detailed features and fine-grained text. To achieve this, we introduce a sharing position embedding to link each character with its corresponding token and employ a bidirectional distance loss to align each region text prompt character with the target ``text''. To refine the information at fine-grained level, we implement character-token level interactions before and after encoding. Our proposed method combines a general score map from the image-text process with a region score map derived from character-token matching, producing a final score map that could balance the global and local features and be fed into DBNet to detect the text. Experiments on benchmarks like ICDAR2015, TotalText, and CTW1500 demonstrate RPT impressive performance, underscoring its effectiveness for scene text detection.

Mingxin Huang, Hongliang Li, Yuliang Liu et al.

Modularity plays a crucial role in the development and maintenance of complex systems. While end-to-end text spotting efficiently mitigates the issues of error accumulation and sub-optimal performance seen in traditional two-step methodologies, the two-step methods continue to be favored in many competitions and practical settings due to their superior modularity. In this paper, we introduce Bridging Text Spotting, a novel approach that resolves the error accumulation and suboptimal performance issues in two-step methods while retaining modularity. To achieve this, we adopt a well-trained detector and recognizer that are developed and trained independently and then lock their parameters to preserve their already acquired capabilities. Subsequently, we introduce a Bridge that connects the locked detector and recognizer through a zero-initialized neural network. This zero-initialized neural network, initialized with weights set to zeros, ensures seamless integration of the large receptive field features in detection into the locked recognizer. Furthermore, since the fixed detector and recognizer cannot naturally acquire end-to-end optimization features, we adopt the Adapter to facilitate their efficient learning of these features. We demonstrate the effectiveness of the proposed method through extensive experiments: Connecting the latest detector and recognizer through Bridging Text Spotting, we achieved an accuracy of 83.3% on Total-Text, 69.8% on CTW1500, and 89.5% on ICDAR 2015. The code is available at https://github.com/mxin262/Bridging-Text-Spotting.

Mingxin Huang, Dezhi Peng, Hongliang Li et al.

End-to-end scene text spotting, which aims to read the text in natural images, has garnered significant attention in recent years. However, recent state-of-the-art methods usually incorporate detection and recognition simply by sharing the backbone, which does not directly take advantage of the feature interaction between the two tasks. In this paper, we propose a new end-to-end scene text spotting framework termed SwinTextSpotter v2, which seeks to find a better synergy between text detection and recognition. Specifically, we enhance the relationship between two tasks using novel Recognition Conversion and Recognition Alignment modules. Recognition Conversion explicitly guides text localization through recognition loss, while Recognition Alignment dynamically extracts text features for recognition through the detection predictions. This simple yet effective design results in a concise framework that requires neither an additional rectification module nor character-level annotations for the arbitrarily-shaped text. Furthermore, the parameters of the detector are greatly reduced without performance degradation by introducing a Box Selection Schedule. Qualitative and quantitative experiments demonstrate that SwinTextSpotter v2 achieved state-of-the-art performance on various multilingual (English, Chinese, and Vietnamese) benchmarks. The code will be available at \href{https://github.com/mxin262/SwinTextSpotterv2}{SwinTextSpotter v2}.

Mehran Taghian, Yunke Peng, Xing Huang et al.

Large foundation models have become central to modern machine learning, with performance scaling predictably with model size and data. However, training and deploying such models incur substantial computational and memory costs, motivating the development of low-precision training techniques. Recent work has demonstrated that 4-bit floating-point (FP4) formats--such as MXFP4 and NVFP4--can be successfully applied to linear GEMM operations in large language models (LLMs), achieving up to 4x improvements in compute throughput and memory efficiency compared to higher-precision baselines. In this work, we investigate the recently proposed HiFloat4 FP4 format for Huawei Ascend NPUs and systematically compare it with MXFP4 in large-scale training settings. All experiments are conducted on Ascend NPU clusters, with linear and expert GEMM operations performed entirely in FP4 precision. We evaluate both dense architectures (e.g., Pangu and LLaMA-style models) and mixture-of-experts (MoE) models, where both standard linear layers and expert-specific GEMMs operate in FP4. Furthermore, we explore stabilization techniques tailored to FP4 training that significantly reduce numerical degradation, maintaining relative error within 1% of full-precision baselines while preserving the efficiency benefits of 4-bit computation. Our results provide a comprehensive empirical study of FP4 training on NPUs and highlight the practical trade-offs between FP4 formats in large-scale dense and MoE models.

Shimou Ling, Liang Zhang, Jiangwei Zhao et al.

LoRA-based continual learning represents a promising avenue for leveraging pre-trained models in downstream continual learning tasks. Recent studies have shown that orthogonal LoRA tuning effectively mitigates forgetting. However, this work unveils that under orthogonal LoRA tuning, the critical parameters for pre-tasks still change notably after learning post-tasks. To address this problem, we directly propose freezing the most critical parameter matrices in the Vision Transformer (ViT) for pre-tasks before learning post-tasks. In addition, building on orthogonal LoRA tuning, we propose orthogonal LoRA composition (LoRAC) based on QR decomposition, which may further enhance the plasticity of our method. Elaborate ablation studies and extensive comparisons demonstrate the effectiveness of our proposed method. Our results indicate that our method achieves state-of-the-art (SOTA) performance on several well-known continual learning benchmarks. For instance, on the Split CIFAR-100 dataset, our method shows a 6.35\% improvement in accuracy and a 3.24\% reduction in forgetting compared to previous methods. Our code is available at https://github.com/learninginvision/LoRAC-IPC.

Zihuan Qiu, Yi Xu, Chiyuan He et al.

Continual model merging integrates independently fine-tuned models sequentially without access to the original training data, offering a scalable and efficient solution for continual learning. However, existing methods face two critical challenges: parameter interference among tasks, which leads to catastrophic forgetting, and limited adaptability to evolving test distributions. To address these issues, we introduce the task of Test-Time Continual Model Merging (TTCMM), which leverages a small set of unlabeled test samples during inference to alleviate parameter conflicts and handle distribution shifts. We propose MINGLE, a novel framework for TTCMM. MINGLE employs a mixture-of-experts architecture with parameter-efficient, low-rank experts, which enhances adaptability to evolving test distributions while dynamically merging models to mitigate conflicts. To further reduce forgetting, we propose Null-Space Constrained Gating, which restricts gating updates to subspaces orthogonal to prior task representations, thereby suppressing activations on old tasks and preserving past knowledge. We further introduce an Adaptive Relaxation Strategy that adjusts constraint strength dynamically based on interference signals observed during test-time adaptation, striking a balance between stability and adaptability. Extensive experiments on standard continual merging benchmarks demonstrate that MINGLE achieves robust generalization, significantly reduces forgetting, and consistently surpasses previous state-of-the-art methods by 7-9% on average across diverse task orders. Our code is available at: https://github.com/zihuanqiu/MINGLE

Peirong Zhang, Jiaxin Zhang, Jiahuan Cao et al.

We propose LGGPT, an LLM-based model tailored for unified layout generation. First, we propose Arbitrary Layout Instruction (ALI) and Universal Layout Response (ULR) as the uniform I/O template. ALI accommodates arbitrary layout generation task inputs across multiple layout domains, enabling LGGPT to unify both task-generic and domain-generic layout generation hitherto unexplored. Collectively, ALI and ULR boast a succinct structure that forgoes superfluous tokens typically found in existing HTML-based formats, facilitating efficient instruction tuning and boosting unified generation performance. In addition, we propose an Interval Quantization Encoding (IQE) strategy that compresses ALI into a more condensed structure. IQE precisely preserves valid layout clues while eliminating the less informative placeholders, facilitating LGGPT to capture complex and variable layout generation conditions during the unified training process. Experimental results demonstrate that LGGPT achieves superior or on par performance compared to existing methods. Notably, LGGPT strikes a prominent balance between proficiency and efficiency with a compact 1.5B parameter LLM, which beats prior 7B or 175B models even in the most extensive and challenging unified scenario. Furthermore, we underscore the necessity of employing LLMs for unified layout generation and suggest that 1.5B could be an optimal parameter size by comparing LLMs of varying scales. Code is available at https://github.com/NiceRingNode/LGGPT.

Mehdi Setayesh, Mahdi Beitollahi, Yasser H. Khalil et al.

Federated multi-task learning (FMTL) seeks to collaboratively train customized models for users with different tasks while preserving data privacy. Most existing approaches assume model congruity (i.e., the use of fully or partially homogeneous models) across users, which limits their applicability in realistic settings. To overcome this limitation, we aim to learn a shared representation space across tasks rather than shared model parameters. To this end, we propose Muscle loss, a novel contrastive learning objective that simultaneously aligns representations from all participating models. Unlike existing multi-view or multi-model contrastive methods, which typically align models pairwise, Muscle loss can effectively capture dependencies across tasks because its minimization is equivalent to the maximization of mutual information among all the models' representations. Building on this principle, we develop FedMuscle, a practical and communication-efficient FMTL algorithm that naturally handles both model and task heterogeneity. Experiments on diverse image and language tasks demonstrate that FedMuscle consistently outperforms state-of-the-art baselines, delivering substantial improvements and robust performance across heterogeneous settings.

Xiang Li, Heqian Qiu, Lanxiao Wang et al.

Error detection is crucial in industrial training, healthcare, and assembly quality control. Most existing work assumes a single-view setting and cannot handle the practical case where a third-person (exo) demonstration is used to assess a first-person (ego) imitation. We formalize Ego$\rightarrow$Exo Imitation Error Detection: given asynchronous, length-mismatched ego and exo videos, the model must localize procedural steps on the ego timeline and decide whether each is erroneous. This setting introduces cross-view domain shift, temporal misalignment, and heavy redundancy. Under a unified protocol, we adapt strong baselines from dense video captioning and temporal action detection and show that they struggle in this cross-view regime. We then propose SAVA-X, an Align-Fuse-Detect framework with (i) view-conditioned adaptive sampling, (ii) scene-adaptive view embeddings, and (iii) bidirectional cross-attention fusion. On the EgoMe benchmark, SAVA-X consistently improves AUPRC and mean tIoU over all baselines, and ablations confirm the complementary benefits of its components. Code is available at https://github.com/jack1ee/SAVAX.

Xiaohan Guo, Yusong Cai, Zejia Liu et al.

Enabling large-scale generative models to continuously learn new visual concepts is essential for personalizing pre-trained models to meet individual user preferences. Existing approaches for continual visual concept learning are constrained by two fundamental challenges: catastrophic forgetting and parameter expansion. In this paper, we propose Redundancy-Removal Mixture of Experts (R^2MoE), a parameter-efficient framework for lifelong visual concept learning that effectively learns new concepts while incurring minimal parameter overhead. Our framework includes three key innovative contributions: First, we propose a mixture-of-experts framework with a routing distillation mechanism that enables experts to acquire concept-specific knowledge while preserving the gating network's routing capability, thereby effectively mitigating catastrophic forgetting. Second, we propose a strategy for eliminating redundant layer-wise experts that reduces the number of expert parameters by fully utilizing previously learned experts. Third, we employ a hierarchical local attention-guided inference approach to mitigate interference between generated visual concepts. Extensive experiments have demonstrated that our method generates images with superior conceptual fidelity compared to the state-of-the-art (SOTA) method, achieving an impressive 87.8\% reduction in forgetting rates and 63.3\% fewer parameters on the CustomConcept 101 dataset. Our code is available at {https://github.com/learninginvision/R2MoE}

Heqian Qiu, Zhaofeng Shi, Lanxiao Wang et al.

In human imitation learning, the imitator typically take the egocentric view as a benchmark, naturally transferring behaviors observed from an exocentric view to their owns, which provides inspiration for researching how robots can more effectively imitate human behavior. However, current research primarily focuses on the basic alignment issues of ego-exo data from different cameras, rather than collecting data from the imitator's perspective, which is inconsistent with the high-level cognitive process. To advance this research, we introduce a novel large-scale egocentric dataset, called EgoMe, which towards following the process of human imitation learning via the imitator's egocentric view in the real world. Our dataset includes 7902 paired exo-ego videos (totaling15804 videos) spanning diverse daily behaviors in various real-world scenarios. For each video pair, one video captures an exocentric view of the imitator observing the demonstrator's actions, while the other captures an egocentric view of the imitator subsequently following those actions. Notably, EgoMe uniquely incorporates exo-ego eye gaze, other multi-modal sensor IMU data and different-level annotations for assisting in establishing correlations between observing and imitating process. We further provide a suit of challenging benchmarks for fully leveraging this data resource and promoting the robot imitation learning research. Extensive analysis demonstrates significant advantages over existing datasets. Our EgoMe dataset and benchmarks are available at https://huggingface.co/datasets/HeqianQiu/EgoMe.

Hongliang Li, Jiaxin Zhang, Wenhui Liao et al.

Current Multimodal Large Language Model (MLLM) architectures face a critical tradeoff between performance and efficiency: decoder-only architectures achieve higher performance but lower efficiency, while cross-attention-based architectures offer greater efficiency but lower performance. The key distinction lies in how visual tokens are processed. Decoder-only architectures apply self-attention and FFN operations on visual tokens, while cross-attention architectures skip these computations. To investigate whether redundancy exists in this computationally expensive process, we propose a training-free framework for analyzing trained MLLMs. It consists of Probe-Activated Dynamic FFN and Hollow Attention, which enable adjustable reductions in computations for visual tokens, as well as a Layer Ranking Algorithm that prioritizes layers for these reductions. Extensive experiments demonstrate substantial, structured, and clustered redundancy unique to decoder-only MLLMs, offering valuable insights for future MLLM architecture design. Furthermore, by leveraging our reduction framework as a training-free inference acceleration approach, we achieve performance comparable to or better than state-of-the-art methods while remaining compatible with them. Code will be publicly available at https://github.com/L-Hugh/RedundancyLens.

Hongliang Li, Nong Zhang, Zhewen Xu et al.

Understanding and predicting the Madden-Julian Oscillation (MJO) is fundamental for precipitation forecasting and disaster prevention. To date, long-term and accurate MJO prediction has remained a challenge for researchers. Conventional MJO prediction methods using Numerical Weather Prediction (NWP) are resource-intensive, time-consuming, and highly unstable (most NWP methods are sensitive to seasons, with better MJO forecast results in winter). While existing Artificial Neural Network (ANN) methods save resources and speed forecasting, their accuracy never reaches the 28 days predicted by the state-of-the-art NWP method, i.e., the operational forecasts from ECMWF, since neural networks cannot handle climate data effectively. In this paper, we present a Domain Knowledge Embedded Spatio-Temporal Network (DK-STN), a stable neural network model for accurate and efficient MJO forecasting. It combines the benefits of NWP and ANN methods and successfully improves the forecast accuracy of ANN methods while maintaining a high level of efficiency and stability. We begin with a spatial-temporal network (STN) and embed domain knowledge in it using two key methods: (i) applying a domain knowledge enhancement method and (ii) integrating a domain knowledge processing method into network training. We evaluated DK-STN with the 5th generation of ECMWF reanalysis (ERA5) data and compared it with ECMWF. Given 7 days of climate data as input, DK-STN can generate reliable forecasts for the following 28 days in 1-2 seconds, with an error of only 2-3 days in different seasons. DK-STN significantly exceeds ECMWF in that its forecast accuracy is equivalent to ECMWF's, while its efficiency and stability are significantly superior.

Hongliang Li, Jinan Xu, Gengping Cui et al.

The robustness and security of large language models (LLMs) has become a prominent research area. One notable vulnerability is the ability to bypass LLM safeguards by translating harmful queries into rare or underrepresented languages, a simple yet effective method of "jailbreaking" these models. Despite the growing concern, there has been limited research addressing the safeguarding of LLMs in multilingual scenarios, highlighting an urgent need to enhance multilingual safety. In this work, we investigate the correlation between various attack features across different languages and propose Multilingual Collaborative Defense (MCD), a novel learning method that optimizes a continuous, soft safety prompt automatically to facilitate multilingual safeguarding of LLMs. The MCD approach offers three advantages: First, it effectively improves safeguarding performance across multiple languages. Second, MCD maintains strong generalization capabilities while minimizing false refusal rates. Third, MCD mitigates the language safety misalignment caused by imbalances in LLM training corpora. To evaluate the effectiveness of MCD, we manually construct multilingual versions of commonly used jailbreak benchmarks, such as MaliciousInstruct and AdvBench, to assess various safeguarding methods. Additionally, we introduce these datasets in underrepresented (zero-shot) languages to verify the language transferability of MCD. The results demonstrate that MCD outperforms existing approaches in safeguarding against multilingual jailbreak attempts while also exhibiting strong language transfer capabilities. Our code is available at https://github.com/HLiang-Lee/MCD.

Shuai Chen, Fanman Meng, Liming Lei et al.

Few-shot segmentation (FSS) aims to segment new classes using few annotated images. While recent FSS methods have shown considerable improvements by leveraging Segment Anything Model (SAM), they face two critical limitations: insufficient utilization of structural correlations in query images, and significant information loss when converting continuous position priors to discrete point prompts. To address these challenges, we propose CMaP-SAM, a novel framework that introduces contraction mapping theory to optimize position priors for SAM-driven few-shot segmentation. CMaP-SAM consists of three key components: (1) a contraction mapping module that formulates position prior optimization as a Banach contraction mapping with convergence guarantees. This module iteratively refines position priors through pixel-wise structural similarity, generating a converged prior that preserves both semantic guidance from reference images and structural correlations in query images; (2) an adaptive distribution alignment module bridging continuous priors with SAM's binary mask prompt encoder; and (3) a foreground-background decoupled refinement architecture producing accurate final segmentation masks. Extensive experiments demonstrate CMaP-SAM's effectiveness, achieving state-of-the-art performance with 71.1 mIoU on PASCAL-$5^i$ and 56.1 on COCO-$20^i$ datasets. Code is available at https://github.com/Chenfan0206/CMaP-SAM.

Rayan Merghani Ahmed, Adnan Iltaf, Mohamed Elmanna et al.

Accurate segmentation of coronary Digital Subtraction Angiography images is essential to diagnose and treat coronary artery diseases. Despite advances in deep learning, challenges such as high intra-class variance and class imbalance limit precise vessel delineation. Most existing approaches for coronary DSA segmentation cannot address these issues. Also, existing segmentation network's encoders do not directly generate semantic embeddings, which could enable the decoder to reconstruct segmentation masks effectively from these well-defined features. We propose a Supervised Prototypical Contrastive Loss that fuses supervised and prototypical contrastive learning to enhance coronary DSA image segmentation. The supervised contrastive loss enforces semantic embeddings in the encoder, improving feature differentiation. The prototypical contrastive loss allows the model to focus on the foreground class while alleviating the high intra-class variance and class imbalance problems by concentrating only on the hard-to-classify background samples. We implement the proposed SPCL loss within an MSA-UNet3+: a Multi-Scale Attention-Enhanced UNet3+ architecture. The architecture integrates key components: a Multi-Scale Attention Encoder and a Multi-Scale Dilated Bottleneck designed to enhance multi-scale feature extraction and a Contextual Attention Fusion Module built to keep fine-grained details while improving contextual understanding. Experiments on a private coronary DSA dataset show that MSA-UNet3+ outperforms state-of-the-art methods, achieving the highest Dice coefficient and F1-score and significantly reducing ASD and ACD. The developed framework provides clinicians with precise vessel segmentation, enabling accurate identification of coronary stenosis and supporting informed diagnostic and therapeutic decisions. The code will be released at https://github.com/rayanmerghani/MSA-UNet3plus.

Haoran Wei, Qingbo Wu, Hui Li et al.

Due to the lack of natural scene and haze prior information, it is greatly challenging to completely remove the haze from a single image without distorting its visual content. Fortunately, the real-world haze usually presents non-homogeneous distribution, which provides us with many valuable clues in partial well-preserved regions. In this paper, we propose a Non-Homogeneous Haze Removal Network (NHRN) via artificial scene prior and bidimensional graph reasoning. Firstly, we employ the gamma correction iteratively to simulate artificial multiple shots under different exposure conditions, whose haze degrees are different and enrich the underlying scene prior. Secondly, beyond utilizing the local neighboring relationship, we build a bidimensional graph reasoning module to conduct non-local filtering in the spatial and channel dimensions of feature maps, which models their long-range dependency and propagates the natural scene prior between the well-preserved nodes and the nodes contaminated by haze. To the best of our knowledge, this is the first exploration to remove non-homogeneous haze via the graph reasoning based framework. We evaluate our method on different benchmark datasets. The results demonstrate that our method achieves superior performance over many state-of-the-art algorithms for both the single image dehazing and hazy image understanding tasks. The source code of the proposed NHRN is available on https://github.com/whrws/NHRNet.

Jiabao Zhao, Jonghan Lim, Hongliang Li et al.

With increasing demand for mass customization, traditional manufacturing robots that rely on rule-based operations lack the flexibility to accommodate customized or new product variants. Human-Robot Collaboration (HRC) has demonstrated potential to improve system adaptability by leveraging human versatility and decision-making capabilities. However, existing HRC frame- works typically depend on predefined perception-manipulation pipelines, limiting their ability to autonomously generate task plans for new product assembly. In this work, we propose CoViLLM, an adaptive human-robot collaborative assembly frame- work that supports the assembly of customized and previously unseen products. CoViLLM combines depth-camera-based localization for object position estimation, human operator classification for identifying new components, and an Large Language Model (LLM) for assembly task planning based on natural language instructions. The framework is validated on the NIST Assembly Task Board for known, customized, and new product cases. Experimental results show that the proposed framework enables flexible collaborative assembly by extending HRC beyond predefined product and task settings.

Kaiyu Huang, Fengran Mo, Xinyu Zhang et al. · tsinghua

The rapid development of Large Language Models (LLMs) demonstrates remarkable multilingual capabilities in natural language processing, attracting global attention in both academia and industry. To mitigate potential discrimination and enhance the overall usability and accessibility for diverse language user groups, it is important for the development of language-fair technology. Despite the breakthroughs of LLMs, the investigation into the multilingual scenario remains insufficient, where a comprehensive survey to summarize recent approaches, developments, limitations, and potential solutions is desirable. To this end, we provide a survey with multiple perspectives on the utilization of LLMs in the multilingual scenario. We first rethink the transitions between previous and current research on pre-trained language models. Then we introduce several perspectives on the multilingualism of LLMs, including training and inference methods, information retrieval, model security, multi-domain with language culture, and usage of datasets. We also discuss the major challenges that arise in these aspects, along with possible solutions. Besides, we highlight future research directions that aim at further enhancing LLMs with multilingualism. The survey aims to help the research community address multilingual problems and provide a comprehensive understanding of the core concepts, key techniques, and latest developments in multilingual natural language processing based on LLMs.

Hongliang Li, Herschel C. Pangborn, Ilya Kovalenko

Demand-side energy management, such as the real-time pricing (RTP) program, offers manufacturers opportunities to reduce energy costs by shifting production to low-price hours. However, this strategy is challenging to implement when machine degradation is considered, as degraded machines have decreased processing capacity and increased energy consumption. Proactive maintenance (PM) can restore machine health but requires production downtime, creating a challenging trade-off: scheduling maintenance during low-price periods sacrifices energy savings opportunities, while deferring maintenance leads to capacity losses and higher energy consumption. To address this challenge, we propose a hierarchical bi-level control framework that jointly optimizes PM planning and runtime production scheduling, considering the machine degradation. A higher-level optimization, with the lower-level model predictive control (MPC) embedded as a sub-problem, determines PM plans that minimize total operational costs under day-ahead RTP. At runtime, the lower-level MPC executes closed-loop production scheduling to minimize energy costs under realized RTP, meeting delivery targets. Simulation results from a lithium-ion battery pack assembly line case study demonstrate that the framework strategically shifts PM away from bottlenecks and high-price hours, meeting daily production targets while reducing energy costs.

Ahmad Rashid, Ruotian Wu, Rongqi Fan et al.

Reward-guided text generation (RGTG) has emerged as a viable alternative to offline reinforcement learning from human feedback (RLHF). RGTG methods can align baseline language models to human preferences without further training like in standard RLHF methods. However, they rely on a reward model to score each candidate token generated by the language model at inference, incurring significant test-time overhead. Additionally, the reward model is usually only trained to score full sequences, which can lead to sub-optimal choices for partial sequences. In this work, we present a novel reward model architecture that is trained, using a Bradley-Terry loss, to prefer the optimal expansion of a sequence with just a \emph{single call} to the reward model at each step of the generation process. That is, a score for all possible candidate tokens is generated simultaneously, leading to efficient inference. We theoretically analyze various RGTG reward models and demonstrate that prior techniques prefer sub-optimal sequences compared to our method during inference. Empirically, our reward model leads to significantly faster inference than other RGTG methods. It requires fewer calls to the reward model and performs competitively compared to previous RGTG and offline RLHF methods.

Xiang Li, Heqian Qiu, Lanxiao Wang et al.

With the rapid development of artificial intelligence technologies and wearable devices, egocentric vision understanding has emerged as a new and challenging research direction, gradually attracting widespread attention from both academia and industry. Egocentric vision captures visual and multimodal data through cameras or sensors worn on the human body, offering a unique perspective that simulates human visual experiences. This paper provides a comprehensive survey of the research on egocentric vision understanding, systematically analyzing the components of egocentric scenes and categorizing the tasks into four main areas: subject understanding, object understanding, environment understanding, and hybrid understanding. We explore in detail the sub-tasks within each category. We also summarize the main challenges and trends currently existing in the field. Furthermore, this paper presents an overview of high-quality egocentric vision datasets, offering valuable resources for future research. By summarizing the latest advancements, we anticipate the broad applications of egocentric vision technologies in fields such as augmented reality, virtual reality, and embodied intelligence, and propose future research directions based on the latest developments in the field.

Peirong Zhang, Yuliang Liu, Songxuan Lai et al.

Handwriting verification has stood as a steadfast identity authentication method for decades. However, this technique risks potential privacy breaches due to the inclusion of personal information in handwritten biometrics such as signatures. To address this concern, we propose using the Random Digit String (RDS) for privacy-preserving handwriting verification. This approach allows users to authenticate themselves by writing an arbitrary digit sequence, effectively ensuring privacy protection. To evaluate the effectiveness of RDS, we construct a new HRDS4BV dataset composed of online naturally handwritten RDS. Unlike conventional handwriting, RDS encompasses unconstrained and variable content, posing significant challenges for modeling consistent personal writing style. To surmount this, we propose the Pattern Attentive VErification Network (PAVENet), along with a Discriminative Pattern Mining (DPM) module. DPM adaptively enhances the recognition of consistent and discriminative writing patterns, thus refining handwriting style representation. Through comprehensive evaluations, we scrutinize the applicability of online RDS verification and showcase a pronounced outperformance of our model over existing methods. Furthermore, we discover a noteworthy forgery phenomenon that deviates from prior findings and discuss its positive impact in countering malicious impostor attacks. Substantially, our work underscores the feasibility of privacy-preserving biometric verification and propels the prospects of its broader acceptance and application.

Linfeng Xu, Fanman Meng, Qingbo Wu et al.

The application of activity recognition in the ``AI + Education" field is gaining increasing attention. However, current work mainly focuses on the recognition of activities in manually captured videos and a limited number of activity types, with little attention given to recognizing activities in surveillance images from real classrooms. Activity recognition in classroom surveillance images faces multiple challenges, such as class imbalance and high activity similarity. To address this gap, we constructed a novel multimodal dataset focused on classroom surveillance image activity recognition called ARIC (Activity Recognition In Classroom). The ARIC dataset has advantages of multiple perspectives, 32 activity categories, three modalities, and real-world classroom scenarios. In addition to the general activity recognition tasks, we also provide settings for continual learning and few-shot continual learning. We hope that the ARIC dataset can act as a facilitator for future analysis and research for open teaching scenarios. You can download preliminary data from https://ivipclab.github.io/publication_ARIC/ARIC.

Huiyu Xiong, Lanxiao Wang, Heqian Qiu et al.

To address the problem of catastrophic forgetting due to the invisibility of old categories in sequential input, existing work based on relatively simple categorization tasks has made some progress. In contrast, video captioning is a more complex task in multimodal scenario, which has not been explored in the field of incremental learning. After identifying this stability-plasticity problem when analyzing video with sequential input, we originally propose a method to Mitigate Catastrophic Forgetting in class-incremental learning for multimodal Video Captioning (MCF-VC). As for effectively maintaining good performance on old tasks at the macro level, we design Fine-grained Sensitivity Selection (FgSS) based on the Mask of Linear's Parameters and Fisher Sensitivity to pick useful knowledge from old tasks. Further, in order to better constrain the knowledge characteristics of old and new tasks at the specific feature level, we have created the Two-stage Knowledge Distillation (TsKD), which is able to learn the new task well while weighing the old task. Specifically, we design two distillation losses, which constrain the cross modal semantic information of semantic attention feature map and the textual information of the final outputs respectively, so that the inter-model and intra-model stylized knowledge of the old class is retained while learning the new class. In order to illustrate the ability of our model to resist forgetting, we designed a metric CIDER_t to detect the stage forgetting rate. Our experiments on the public dataset MSR-VTT show that the proposed method significantly resists the forgetting of previous tasks without replaying old samples, and performs well on the new task.

Jian Jiao, Yu Dai, Hefei Mei et al.

Recent video class-incremental learning usually excessively pursues the accuracy of the newly seen classes and relies on memory sets to mitigate catastrophic forgetting of the old classes. However, limited storage only allows storing a few representative videos. So we propose SNRO, which slightly shifts the features of new classes to remember old classes. Specifically, SNRO contains Examples Sparse(ES) and Early Break(EB). ES decimates at a lower sample rate to build memory sets and uses interpolation to align those sparse frames in the future. By this, SNRO stores more examples under the same memory consumption and forces the model to focus on low-semantic features which are harder to be forgotten. EB terminates the training at a small epoch, preventing the model from overstretching into the high-semantic space of the current task. Experiments on UCF101, HMDB51, and UESTC-MMEA-CL datasets show that SNRO performs better than other approaches while consuming the same memory consumption.

Hefei Mei, Taijin Zhao, Shiyuan Tang et al.

Few-shot object detection (FSOD) aims to achieve object detection only using a few novel class training data. Most of the existing methods usually adopt a transfer-learning strategy to construct the novel class distribution by transferring the base class knowledge. However, this direct way easily results in confusion between the novel class and other similar categories in the decision space. To address the problem, we propose generating local reverse samples (LRSamples) in Prototype Reference Frames to adaptively adjust the center position and boundary range of the novel class distribution to learn more discriminative novel class samples for FSOD. Firstly, we propose a Center Calibration Variance Augmentation (CCVA) module, which contains the selection rule of LRSamples, the generator of LRSamples, and augmentation on the calibrated distribution centers. Specifically, we design an intra-class feature converter (IFC) as the generator of CCVA to learn the selecting rule. By transferring the knowledge of IFC from the base training to fine-tuning, the IFC generates plentiful novel samples to calibrate the novel class distribution. Moreover, we propose a Feature Density Boundary Optimization (FDBO) module to adaptively adjust the importance of samples depending on their distance from the decision boundary. It can emphasize the importance of the high-density area of the similar class (closer decision boundary area) and reduce the weight of the low-density area of the similar class (farther decision boundary area), thus optimizing a clearer decision boundary for each category. We conduct extensive experiments to demonstrate the effectiveness of our proposed method. Our method achieves consistent improvement on the Pascal VOC and MS COCO datasets based on DeFRCN and MFDC baselines.

Zihuan Qiu, Yi Xu, Fanman Meng et al.

Class Incremental Learning (CIL) aims to sequentially acquire knowledge of new classes without forgetting previously learned ones. Despite recent progress, current CIL methods still exhibit significant performance gaps compared to their oracle counterparts-models trained with full access to historical data. Inspired by recent insights on Linear Mode Connectivity (LMC), we revisit the geometric properties of oracle solutions in CIL and uncover a fundamental observation: these oracle solutions typically maintain low-loss linear connections to the optimum of previous tasks. Motivated by this finding, we propose Increment Vector Transformation (IVT), a novel plug-and-play framework designed to mitigate catastrophic forgetting during training. Rather than directly following CIL updates, IVT periodically teleports the model parameters to transformed solutions that preserve linear connectivity to previous task optimum. By maintaining low-loss along these connecting paths, IVT effectively ensures stable performance on previously learned tasks. The transformation is efficiently approximated using diagonal Fisher Information Matrices, making IVT suitable for both exemplar-free and exemplar-based scenarios, and compatible with various initialization strategies. Extensive experiments on CIFAR-100, FGVCAircraft, ImageNet-Subset, and ImageNet-Full demonstrate that IVT consistently enhances the performance of strong CIL baselines. Specifically, on CIFAR-100, IVT improves the last accuracy of the PASS baseline by +5.12% and reduces forgetting by 2.54%. For the CLIP-pre-trained SLCA baseline on FGVCAircraft, IVT yields gains of +14.93% in average accuracy and +21.95% in last accuracy. The code will be released.

Zihuan Qiu, Lei Wang, Yang Cao et al.

Data-free continual model merging (DFCMM) aims to fuse independently fine-tuned models into a single backbone that evolves with incoming tasks without accessing task data. This paper formulate two fundamental desiderata for DFCMM: transparency, avoiding interference with earlier tasks, and fidelity, adapting faithfully to each new task. This poses a challenge that existing approaches fail to address: how to bridge data-level desiderata with parameter-space optimization to ensure transparency and fidelity in the absence of task data. To this end, we propose NUFILT (NUll-space FILTering), a data-free framework that directly links these desiderata to optimization. Our key observation is that task vectors approximately align with representation subspaces, providing structural surrogates for enforcing transparency and fidelity. Accordingly, we design a null-space projector that preserves prior responses by filtering out overlapping components of new task vectors, thereby ensuring transparency, and a lightweight LoRA adapter that injects complementary task-specific signals, enabling fidelity in adapting to new tasks. The adapter is trained with a projection-based surrogate loss to retain consistency with previous knowledge while introducing novel directions. This joint filtering-adaptation process allows the backbone to absorb new knowledge while retaining existing behaviors, and the updates are finally fused back in a layer-wise linear fashion without extra parameters or inference cost. Theoretically, we establish approximate subspace alignment guarantees that justify null-space filtering. Empirically, NUFILT achieves state-of-the-art performance with minimal forgetting on both vision and NLP benchmarks, improving average accuracy by 4-7% over OPCM and WUDI-Merging, while narrowing the gap to fine-tuning and reducing computation overhead.

Yasser H. Khalil, Mehdi Setayesh, Hongliang Li

Machine unlearning (MU) aims to remove the influence of specific "forget" data from a trained model while preserving its knowledge of the remaining "retain" data. Existing MU methods based on label manipulation or model weight perturbations often achieve limited unlearning effectiveness. To address this, we introduce CoUn, a novel MU framework inspired by the observation that a model retrained from scratch using only retain data classifies forget data based on their semantic similarity to the retain data. CoUn emulates this behavior by adjusting learned data representations through contrastive learning (CL) and supervised learning, applied exclusively to retain data. Specifically, CoUn (1) leverages semantic similarity between data samples to indirectly adjust forget representations using CL, and (2) maintains retain representations within their respective clusters through supervised learning. Extensive experiments across various datasets and model architectures show that CoUn consistently outperforms state-of-the-art MU baselines in unlearning effectiveness. Additionally, integrating our CL module into existing baselines empowers their unlearning effectiveness.

Shuai Chen, Fanman Meng, Chunjin Yang et al.

Cross-Domain Few-Shot Segmentation (CD-FSS) remains challenging due to limited data and domain shifts. Recent foundation models like the Segment Anything Model (SAM) have shown remarkable zero-shot generalization capability in general segmentation tasks, making it a promising solution for few-shot scenarios. However, adapting SAM to CD-FSS faces two critical challenges: reliance on manual prompt and limited cross-domain ability. Therefore, we propose the Composable Meta-Prompt (CMP) framework that introduces three key modules: (i) the Reference Complement and Transformation (RCT) module for semantic expansion, (ii) the Composable Meta-Prompt Generation (CMPG) module for automated meta-prompt synthesis, and (iii) the Frequency-Aware Interaction (FAI) module for domain discrepancy mitigation. Evaluations across four cross-domain datasets demonstrate CMP's state-of-the-art performance, achieving 71.8\% and 74.5\% mIoU in 1-shot and 5-shot scenarios respectively.

Shuai Chen, Fanman Meng, Xiwei Zhang et al.

This paper presents DFR (Decompose, Fuse and Reconstruct), a novel framework that addresses the fundamental challenge of effectively utilizing multi-modal guidance in few-shot segmentation (FSS). While existing approaches primarily rely on visual support samples or textual descriptions, their single or dual-modal paradigms limit exploitation of rich perceptual information available in real-world scenarios. To overcome this limitation, the proposed approach leverages the Segment Anything Model (SAM) to systematically integrate visual, textual, and audio modalities for enhanced semantic understanding. The DFR framework introduces three key innovations: 1) Multi-modal Decompose: a hierarchical decomposition scheme that extracts visual region proposals via SAM, expands textual semantics into fine-grained descriptors, and processes audio features for contextual enrichment; 2) Multi-modal Contrastive Fuse: a fusion strategy employing contrastive learning to maintain consistency across visual, textual, and audio modalities while enabling dynamic semantic interactions between foreground and background features; 3) Dual-path Reconstruct: an adaptive integration mechanism combining semantic guidance from tri-modal fused tokens with geometric cues from multi-modal location priors. Extensive experiments across visual, textual, and audio modalities under both synthetic and real settings demonstrate DFR's substantial performance improvements over state-of-the-art methods.

Shenxi Liu, Kan Li, Mingyang Zhao et al.

With the rapid progress of artificial intelligence (AI) in multi-modal understanding, there is increasing potential for video comprehension technologies to support professional domains such as medical education. However, existing benchmarks suffer from two primary limitations: (1) Linguistic Singularity: they are largely confined to English, neglecting the need for multilingual resources; and (2) Shallow Reasoning: their questions are often designed for surface-level information retrieval, failing to properly assess deep multi-modal integration. To address these limitations, we present M3-Med, the first benchmark for Multi-lingual, Multi-modal, and Multi-hop reasoning in Medical instructional video understanding. M3-Med consists of medical questions paired with corresponding video segments, annotated by a team of medical experts. A key innovation of M3-Med is its multi-hop reasoning task, which requires a model to first locate a key entity in the text, then find corresponding visual evidence in the video, and finally synthesize information across both modalities to derive the answer. This design moves beyond simple text matching and poses a substantial challenge to a model's deep cross-modal understanding capabilities. We define two tasks: Temporal Answer Grounding in Single Video (TAGSV) and Temporal Answer Grounding in Video Corpus (TAGVC). We evaluated several state-of-the-art models and Large Language Models (LLMs) on M3-Med. The results reveal a significant performance gap between all models and human experts, especially on the complex multi-hop questions where model performance drops sharply. M3-Med effectively highlights the current limitations of AI models in deep cross-modal reasoning within specialized domains and provides a new direction for future research.

Taijin Zhao, Heqian Qiu, Yu Dai et al.

Few-shot object detection (FSOD) aims to detect objects with limited samples for novel classes, while relying on abundant data for base classes. Existing FSOD approaches, predominantly built on the Faster R-CNN detector, entangle objectness recognition and foreground classification within shared feature spaces. This paradigm inherently establishes class-specific objectness criteria and suffers from unrepresentative novel class samples. To resolve this limitation, we propose a Uniform Orthogonal Feature Space (UOFS) optimization framework. First, UOFS decouples the feature space into two orthogonal components, where magnitude encodes objectness and angle encodes classification. This decoupling enables transferring class-agnostic objectness knowledge from base classes to novel classes. Moreover, implementing the disentanglement requires careful attention to two challenges: (1) Base set images contain unlabeled foreground instances, causing confusion between potential novel class instances and backgrounds. (2) Angular optimization depends exclusively on base class foreground instances, inducing overfitting of angular distributions to base classes. To address these challenges, we propose a Hybrid Background Optimization (HBO) strategy: (1) Constructing a pure background base set by removing unlabeled instances in original images to provide unbiased magnitude-based objectness supervision. (2) Incorporating unlabeled foreground instances in the original base set into angular optimization to enhance distribution uniformity. Additionally, we propose a Spatial-wise Attention Disentanglement and Association (SADA) module to address task conflicts between class-agnostic and class-specific tasks. Experiments demonstrate that our method significantly outperforms existing approaches based on entangled feature spaces.

Yukun Chen, Zihuan Qiu, Fanman Meng et al.

Unlike traditional Multimodal Class-Incremental Learning (MCIL) methods that focus only on vision and text, this paper explores MCIL across vision, audio and text modalities, addressing challenges in integrating complementary information and mitigating catastrophic forgetting. To tackle these issues, we propose an MCIL method based on multimodal pre-trained models. Firstly, a Multimodal Incremental Feature Extractor (MIFE) based on Mixture-of-Experts (MoE) structure is introduced to achieve effective incremental fine-tuning for AudioCLIP. Secondly, to enhance feature discriminability and generalization, we propose an Adaptive Audio-Visual Fusion Module (AAVFM) that includes a masking threshold mechanism and a dynamic feature fusion mechanism, along with a strategy to enhance text diversity. Thirdly, a novel multimodal class-incremental contrastive training loss is proposed to optimize cross-modal alignment in MCIL. Finally, two MCIL-specific evaluation metrics are introduced for comprehensive assessment. Extensive experiments on three multimodal datasets validate the effectiveness of our method.

Chiyuan He, Zihuan Qiu, Fanman Meng et al.

Continual learning of vision-language models (VLMs) focuses on leveraging cross-modal pretrained knowledge to incrementally adapt to expanding downstream tasks and datasets, while tackling the challenge of knowledge forgetting. Existing research often focuses on connecting visual features with specific class text in downstream tasks, overlooking the latent relationships between general and specialized knowledge. Our findings reveal that forcing models to optimize inappropriate visual-text matches exacerbates forgetting of VLM's recognition ability. To tackle this issue, we propose DesCLIP, which leverages general attribute (GA) descriptions to guide the understanding of specific class objects, enabling VLMs to establish robust vision-GA-class trilateral associations rather than relying solely on vision-class connections. Specifically, we introduce a language assistant to generate concrete GA description candidates via proper request prompts. Then, an anchor-based embedding filter is designed to obtain highly relevant GA description embeddings, which are leveraged as the paired text embeddings for visual-textual instance matching, thereby tuning the visual encoder. Correspondingly, the class text embeddings are gradually calibrated to align with these shared GA description embeddings. Extensive experiments demonstrate the advancements and efficacy of our proposed method, with comprehensive empirical evaluations highlighting its superior performance in VLM-based recognition compared to existing continual learning methods.