Xuchong Zhang

Zifan Han, Hongbo Sun, Jinglin Xu et al.

Multimodal Large Language Models (MLLMs) have shown strong capabilities in image understanding, motivating recent efforts to extend them to video reasoning. However, existing Video LLMs struggle in online streaming scenarios, where long temporal context must be preserved under strict memory constraints. We propose WAT (Watching Before Thinking), a two-stage framework for online video reasoning. WAT separates processing into a query-independent watching stage and a query-triggered thinking stage. The watching stage builds a hierarchical memory system with a Short-Term Memory (STM) that buffers recent frames and a fixed-capacity Long-Term Memory (LTM) that maintains a diverse summary of historical content using a redundancy-aware eviction policy. In the thinking stage, a context-aware retrieval mechanism combines the query with the current STM context to retrieve relevant historical frames from the LTM for cross-temporal reasoning. To support training for online video tasks, we introduce WAT-85K, a dataset containing streaming-style annotations emphasizing real-time perception, backward tracing, and forecasting. Experiments show that WAT achieves state-of-the-art performance on online video benchmarks, including 77.7% accuracy on StreamingBench and 55.2% on OVO-Bench, outperforming existing open-source online Video LLMs while operating at real-time frame rates.

Yinpeng Dong, Peng Chen, Senyou Deng et al.

The security of artificial intelligence (AI) is an important research area towards safe, reliable, and trustworthy AI systems. To accelerate the research on AI security, the Artificial Intelligence Security Competition (AISC) was organized by the Zhongguancun Laboratory, China Industrial Control Systems Cyber Emergency Response Team, Institute for Artificial Intelligence, Tsinghua University, and RealAI as part of the Zhongguancun International Frontier Technology Innovation Competition (https://www.zgc-aisc.com/en). The competition consists of three tracks, including Deepfake Security Competition, Autonomous Driving Security Competition, and Face Recognition Security Competition. This report will introduce the competition rules of these three tracks and the solutions of top-ranking teams in each track.

Xuchong Zhang, Changfeng Sun, Haoliang Han et al.

Recent studies have demonstrated that object detection networks are usually vulnerable to adversarial examples. Generally, adversarial attacks for object detection can be categorized into targeted and untargeted attacks. Compared with untargeted attacks, targeted attacks present greater challenges and all existing targeted attack methods launch the attack by misleading detectors to mislabel the detected object as a specific wrong label. However, since these methods must depend on the presence of the detected objects within the victim image, they suffer from limitations in attack scenarios and attack success rates. In this paper, we propose a targeted feature space attack method that can mislead detectors to `fabricate' extra designated objects regardless of whether the victim image contains objects or not. Specifically, we introduce a guided image to extract coarse-grained features of the target objects and design an innovative dual attention mechanism to filter out the critical features of the target objects efficiently. The attack performance of the proposed method is evaluated on MS COCO and BDD100K datasets with FasterRCNN and YOLOv5. Evaluation results indicate that the proposed targeted feature space attack method shows significant improvements in terms of image-specific, universality, and generalization attack performance, compared with the previous targeted attack for object detection.

Jingwei Sun, Xuchong Zhang, Changfeng Sun et al.

Multi-View Diffusion Models (MVDMs) enable remarkable improvements in the field of 3D geometric reconstruction, but the issue regarding intellectual property has received increasing attention due to unauthorized imitation. Recently, some works have utilized adversarial attacks to protect copyright. However, all these works focus on single-image generation tasks which only need to consider the inner feature of images. Previous methods are inefficient in attacking MVDMs because they lack the consideration of disrupting the geometric and visual consistency among the generated multi-view images. This paper is the first to address the intellectual property infringement issue arising from MVDMs. Accordingly, we propose a novel latent feature and attention dual erasure attack to disrupt the distribution of latent feature and the consistency across the generated images from multi-view and multi-domain simultaneously. The experiments conducted on SOTA MVDMs indicate that our approach achieves superior performances in terms of attack effectiveness, transferability, and robustness against defense methods. Therefore, this paper provides an efficient solution to protect 3D assets from MVDMs-based 3D geometry reconstruction.

Ningnan Wang, Weihuang Chen, Liming Chen et al.

Embodied visual navigation remains a challenging task, as agents must explore unknown environments with limited knowledge. Existing zero-shot studies have shown that incorporating memory mechanisms to support goal-directed behavior can improve long-horizon planning performance. However, they overlook visual frontier boundaries, which fundamentally dictate future trajectories and observations, and fall short of inferring the relationship between partial visual observations and navigation goals. In this paper, we propose Semantic Cognition Over Potential-based Exploration (SCOPE), a zero-shot framework that explicitly leverages frontier information to drive potential-based exploration, enabling more informed and goal-relevant decisions. SCOPE estimates exploration potential with a Vision-Language Model and organizes it into a spatio-temporal potential graph, capturing boundary dynamics to support long-horizon planning. In addition, SCOPE incorporates a self-reconsideration mechanism that revisits and refines prior decisions, enhancing reliability and reducing overconfident errors. Experimental results on two diverse embodied navigation tasks show that SCOPE outperforms state-of-the-art baselines by 4.6\% in accuracy. Further analysis demonstrates that its core components lead to improved calibration, stronger generalization, and higher decision quality.

Canhui Tang, Zifan Han, Hongbo Sun et al.

Multimodal Large Language Models (MLLMs) have demonstrated significant progress in vision-language tasks, yet they still face challenges when processing long-duration video inputs. The limitation arises from MLLMs' context limit and training costs, necessitating sparse frame sampling before feeding videos into MLLMs. However, building a trainable sampling method remains challenging due to the unsupervised and non-differentiable nature of sparse frame sampling in Video-MLLMs. To address these problems, we propose Temporal Sampling Policy Optimization (TSPO), advancing MLLMs' long-form video-language understanding via reinforcement learning. Specifically, we first propose a trainable event-aware temporal agent, which captures event-query correlation for performing probabilistic keyframe selection. Then, we propose the TSPO reinforcement learning paradigm, which models keyframe selection and language generation as a joint decision-making process, enabling end-to-end group relative optimization for the temporal sampling policy. Furthermore, we propose a dual-style long video training data construction pipeline, balancing comprehensive temporal understanding and key segment localization. Finally, we incorporate rule-based answering accuracy and temporal locating reward mechanisms to optimize the temporal sampling policy. Comprehensive experiments show that our TSPO achieves state-of-the-art performance across multiple long video understanding benchmarks, and shows transferable ability across different cutting-edge Video-MLLMs. Our code is available at https://github.com/Hui-design/TSPO

Cheng Cheng, Lin Song, Yicheng Xiao et al.

Autoregressive (AR) image generators offer a language-model-friendly approach to image generation by predicting discrete image tokens in a causal sequence. However, unlike diffusion models, AR models lack a mechanism to refine previous predictions, limiting their generation quality. In this paper, we introduce TensorAR, a new AR paradigm that reformulates image generation from next-token prediction to next-tensor prediction. By generating overlapping windows of image patches (tensors) in a sliding fashion, TensorAR enables iterative refinement of previously generated content. To prevent information leakage during training, we propose a discrete tensor noising scheme, which perturbs input tokens via codebook-indexed noise. TensorAR is implemented as a plug-and-play module compatible with existing AR models. Extensive experiments on LlamaGEN, Open-MAGVIT2, and RAR demonstrate that TensorAR significantly improves the generation performance of autoregressive models.

He Dai, Xuchong Zhang, Yongli Zhao et al.

Efficient real-time disparity estimation is critical for the application of stereo vision systems in various areas. Recently, stereo network based on coarse-to-fine method has largely relieved the memory constraints and speed limitations of large-scale network models. Nevertheless, all of the previous coarse-to-fine designs employ constant offsets and three or more stages to progressively refine the coarse disparity map, still resulting in unsatisfactory computation accuracy and inference time when deployed on mobile devices. This paper claims that the coarse matching errors can be corrected efficiently with fewer stages as long as more accurate disparity candidates can be provided. Therefore, we propose a dynamic offset prediction module to meet different correction requirements of diverse objects and design an efficient two-stage framework. Besides, we propose a disparity-independent convolution to further improve the performance since it is more consistent with the local statistical characteristics of the compact cost volume. The evaluation results on multiple datasets and platforms clearly demonstrate that, the proposed network outperforms the state-of-the-art lightweight models especially for mobile devices in terms of accuracy and speed. Code will be made available.

Myung Seok Shim, Chenye Zhao, Yang Li et al.

Sensor fusion has wide applications in many domains including health care and autonomous systems. While the advent of deep learning has enabled promising multi-modal fusion of high-level features and end-to-end sensor fusion solutions, existing deep learning based sensor fusion techniques including deep gating architectures are not always resilient, leading to the issue of fusion weight inconsistency. We propose deep multi-modal sensor fusion architectures with enhanced robustness particularly under the presence of sensor failures. At the core of our gating architectures are fusion weight regularization and fusion target learning operating on auxiliary unimodal sensing networks appended to the main fusion model. The proposed regularized gating architectures outperform the existing deep learning architectures with and without gating under both clean and corrupted sensory inputs resulted from sensor failures. The demonstrated improvements are particularly pronounced when one or more multiple sensory modalities are corrupted.