Changsheng Xu

Few-Shot Class-Incremental Learning (FSCIL) aims to continually learn new knowledge about few-shot novel classes while retaining the previously learned knowledge of old classes. With the rapid development of vision-language models (e.g., CLIP) known for their strong transferability, several studies have explored applying them to FSCIL. However, they generally retain the original architecture, where the visual and textual encoders operate independently. This limits the impact of textual semantics on the discriminative capability of the visual branch and somewhat restricts the model’s performance in FSCIL. In this article, we argue that integrating class-level semantics from the CLIP textual branch into the visual branch can provide a more robust representational foundation for FSCIL. To this end, we propose LEAD, a CLIP-based FSCIL method that explicitly leverages textual semantic guidance to enhance the discriminability and generalizability of the visual branch, improving adaptability to few-shot novel classes while mitigating interference with old ones. Specifically, the Conditional Semantic Constructor (CSC) adaptively converts textual semantic cues into class-relevant guidance, and the Semantic-Enhanced Attention (SEA) module explicitly incorporates semantic guidance into the image encoder to strengthen semantic awareness. A text-guided prototype generation strategy incrementally expands the text-guided prototype space, enhancing the distinction between different classes while promoting model stability in incremental sessions. Extensive experiments on three benchmark datasets validate the effectiveness of LEAD and its competitive performance against existing FSCIL methods.

Due to the impressive zero-shot capabilities, pre-trained vision-language models (e.g., CLIP), have attracted widespread attention and adoption across various domains. Nonetheless, CLIP has been observed to be susceptible to adversarial examples. Through experimental analysis, we have observed a phenomenon wherein adversarial perturbations induce shifts in text-guided attention. Building upon this observation, we propose a simple yet effective strategy: Text-Guided Attention for Zero-Shot Robustness (TGA-ZSR). This framework incorporates two components: Local Attention Refinement Module and Global Attention Constraint Module. Our goal is to maintain the generalization of the CLIP model and enhance its adversarial robustness: The Local Attention Refinement Module aligns the text-guided attention obtained from the target model via adversarial examples with the text-guided attention acquired from the original model via clean examples. This alignment enhances the model's robustness. Additionally, the Global Attention Constraint Module acquires text-guided attention from both the target and original models using clean examples. Its objective is to maintain model performance on clean samples while enhancing overall robustness. However, we observe that the method occasionally focuses on irrelevant or spurious features, which can lead to suboptimal performance and undermine its robustness in certain scenarios. To overcome this limitation, we further propose a novel approach called Complementary Text-Guided Attention (Comp-TGA). This method integrates two types of foreground attention: attention guided by the class prompt and reversed attention driven by the non-class prompt. These complementary attention mechanisms allow the model to capture a more comprehensive and accurate representation of the foreground. The experiments validate that TGA-ZSR and Comp-TGA yield 9.58% and 11.95% improvements respectively, in zero-shot robust accuracy over the current state-of-the-art techniques across 16 datasets.

Conventional Video Large Language Models (Video-LLMs) operate under an offline paradigm: they ingest a complete video before producing any response. This post-hoc design fundamentally limits their ability to serve as real-time interactors that perceive, reason, and communicate while the video is still unfolding. Streaming video understanding breaks this constraint by requiring models to process frames as they arrive and generate temporally grounded responses on the fly, thereby transforming passive narrators into active, always-on interactors. In this survey, we present a comprehensive and structured overview of recent advances in streaming video understanding, with a particular focus on Video-LLM-based approaches. We categorize existing methods into two complementary paradigms: proactive streaming models, which autonomously determine when to initiate responses in dynamic visual environments, and reactive streaming models, which focus on how to maintain long-term context and deliver efficient inference upon user queries over unbounded video streams. Together, these two paradigms characterize the core capabilities required for online video interactors: temporal awareness for timely engagement and scalable memory for sustained reasoning. We further provide a systematic review of emerging benchmarks and datasets designed to evaluate streaming-specific capabilities, including multi-turn dialogue, real-time captioning, and proactive response timing. Finally, we identify the key open challenges and outline promising future directions toward building truly interactive, efficient, and scalable streaming video systems. An accompanying resource repository is maintained at https://github.com/sotayang/Awesome-Streaming-Video-Understanding.

CLIP (Contrastive Language-Image Pre-training) has demonstrated outstanding capability in aligning visual and textual representations, which in turn has supported numerous follow-up models and downstream applications. However, their widespread deployment also exposes them to critical security vulnerabilities, particularly under backdoor attacks. Existing backdoor attack methods typically bind a single target class through one-time training or fine-tuning, which significantly limits their threat potential and practicality. To address this limitation, we propose a novel multi-target backdoor attack framework, the Multiple Target Classes by Rapid Transformation Network (MRTNet), which enables simultaneous attacks on multiple target classes in any CLIP model and facilitates rapid switching between them, all without requiring retraining or fine-tuning of the victim model. First, we propose a novel stealthy multi-target trigger generation, which can stealthily activate multi-target attacks without being perceptible to human eyes. Second, we propose a multi-target rapid transformation network, which introduces a decoupled trigger recognizer external to the encoder and a multiple target feature rebinding mechanism, enabling efficient switching between multiple target classes without retraining the base model. Experimental results demonstrate that our method achieves both high stealthiness and strong attack performance, with success rates exceeding 97. 7% in different settings. We hope that this research provides new insights into the security vulnerabilities of CLIP and its downstream applications.

In the rapidly evolving landscape of digital communication, researchers are increasingly focusing on sign language generation as a means to bridge the communication divide for the Deaf and Hard-of-Hearing communities worldwide. Modeling an efficient sign language video generation framework presents two main challenges: ensuring temporal coherence and handling diverse, complex data distributions, especially given the limited availability of annotated sign video data. To address these challenges, we propose SignMoDP, a sign language video generation model that incorporates a mixture of diffusions for universal sign language generation. We improve temporal coherence by integrating the State Space Model (SSM) into the video diffusion framework, ensuring smooth frame transitions and a natural flow of sign language. By utilizing video-to-patch conversion and a multilingual text encoder, the Sign State Space Model efficiently handles long sequences and complex tasks. Additionally, we introduce the Mixture of Diffusion Paths module to manage diverse data distributions, utilizing expert routing strategies for distinct data regions. We propose a Patch Mixture of Experts (Patch MoE), optimized with low-rank strategies, to further ensure robustness and adaptability in sign language video generation. Evaluations on RWTH-2014 , AUTSL, CSL-Daily, WLASL and LSFB benchmarks demonstrate state-of-the-art performance, positioning SignMoDP as a universal solution for scalable, inclusive sign language generation. More details can be found on our project page: https://github.com/mingtiannihao/SignMoDP.