Erxin Yu, Jing Li, Ming Liao, Qi Zhu, Boyang Xue, Minghui Xu, Baojun Wang, Lanqing Hong, Fei Mi, Lifeng Shang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Deep-learning-based image inpainting technology has achieved remarkable visual consistency but is vulnerable to malicious use. Existing detection methods overlook semantic inconsistencies between targets and backgrounds, leading to ambiguous results due to low discriminability. To tackle these challenges, we draw inspiration from human strategies in visual tasks, which involve initially assigning uncertainty across the entire input and subsequently concentrating on highly uncertain regions using prior knowledge like boundary information. Building on this, we propose a Dual Information Guided Network (DIGNet). It combines object-background semantic modulation with uncertainty to precisely locate inpainting regions. This is the first work to address inpainting prediction inaccuracies by considering both edge uncertainty and semantic inconsistency. DIGNet consists of three key parts: the Edge Uncertainty Awareness Module (EUAM), the Edge Correction Module (ECM) based on semantic differences, and the Dual Information Guided Interaction Module (DIGIM). We use semantic inconsistency to get edge constraints and quantify uncertainty as feature variance to guide mainstream feature maps. The DIGIM effectively fuses guide information for accurate predictions. Comprehensive experiments show that our method outperforms existing CNN-based approaches. Specifically, it improves the F1 Score by at least 0.31% and the IOU by at least 0.19% on multiple datasets.

As large language models (LLMs) constantly evolve, ensuring their safety remains a critical research issue.Previous red teaming approaches for LLM safety have primarily focused on single prompt attack or goal hijacking.To the best of our knowledge, we are the first to study LLM safety in multi-turn dialogue coreference.We created a dataset of 1, 400 questions across 14 categories, each featuring multi-turn coreference safety attacks.We then conducted detailed evaluations on five widely used open-source LLMs.The results indicated that under multi-turn coreference safety attacks, the highest attack successful rate was 56% with the LLaMA2-Chat-7b model, while the lowest was 13.9% with the Mistral-7B-Instruct model.These findings highlight the safety vulnerabilities in LLMs during dialogue coreference interactions.Warning: This paper may contain offensive language or harmful content. 1

Point cloud-based place recognition can be used for global localization in large-scale scenes and loop-closure detection in simultaneous localization and mapping (SLAM) systems in the absence of GPS. Current learning-based approaches aim to extract global and local features from 3D point clouds to encode them as descriptors for point cloud retrieval. The key problems are that the occlusion of point clouds by dynamic objects in the scene affects the point cloud structure, a single perceptual field of the network cannot adequately extract point cloud features, and the correlation between features is not fully utilized. To overcome this, we propose a novel network called CapsLoc3D. We first use the static point cloud generation module to remove the occlusion effects of dynamic objects, and then obtain the point cloud descriptors by processing with the CapsLoc3D network which contains the point spatial transformation module, multi-scale feature fusion module, Capsnet module and a GeM Pooling layer. After validation using the Oxford RobotCar, KITTI, and NEU datasets, experiments show that our method performs better and also has good generalization performance and computational efficiency compared with current state-of-the-art algorithms.

Loop closure detection has the potential to correct the drift of trajectories and build a global consistent map in LiDAR SLAM, however it remains a challenging problem in outdoor environment due to the sparsity of 3D point clouds data, large-scale scenes and moving objects. Inspired by the way humans perceive the environment through recognizing objects and identifying their relations, this paper presents a novel descriptor that contains semantic and topological information for loop closure detection. Unlike most existing methods that extract features from the raw point clouds or use all semantic objects, we directly discard point clouds representing pedestrians and vehicles after semantic segmentation. Then, we propose a semantic topological graph representation from the remaining point clouds and convert this graph into a descriptor. Additionally, we propose a two-stage algorithm for matching descriptors to efficiently determine the loop. Our method has been extensively evaluated using the KITTI dataset and outperforms state-of-the-art methods, especially in the challenging situations such as viewpoint changes and dynamic scenes.

Ming Liao, Jing Li, Haisong Zhang, Lingzhi Wang, Xixin Wu, Kam-Fai Wong. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Conventional topic models are ineffective for topic extraction from microblog messages since the lack of structure and context among the posts renders poor message-level word co-occurrence patterns.In this work, we organize microblog posts as conversation trees based on reposting and replying relations, which enrich context information to alleviate data sparseness.Our model generates words according to topic dependencies derived from the conversation structures.In specific, we differentiate messages as leader messages, which initiate key aspects of previously focused topics or shift the focus to different topics, and follower messages that do not introduce any new information but simply echo topics from the messages that they repost or reply.Our model captures the different extents that leader and follower messages may contain the key topical words, thus further enhances the quality of the induced topics.The results of thorough experiments demonstrate the effectiveness of our proposed model.

The Optical Burst-Train Switching (OBTS) network efficiently transports multi-granular subwavelength constant-bit-rate (CBR) and variable-bitrate (VBR) traffic streams over an all-optical network with sparse wavelength convertibility while minimizing access blocking of traffic streams. Sub-wavelengt

This paper presents a novel architecture of multi-granular stream optical burst switching (MGS-OBS), which enables an all-optical network to support sub-wavelength streaming traffic of diverse bandwidth granularities with guaranteed end-to-end delay and data loss rate. It shapes a sub-wavelength stream into a periodic burst train; and allows variable bit rate (VBR) streaming traffic with same source and destination to share the reserved bandwidth through statistical multiplexing. By adopting periodic advanced scheduling of burst trains, MGS-OBS substantially decreases recursive signaling processing in supporting sub-wavelength streaming traffic. MGS-OBS employs burst train reservation blocking recovery by time-shifting a blocked burst train. Performance of MGS-OBS is analyzed in terms of wavelength utilization and data loss rate, and the stream transmission failure rate. It is found that the performance of the MGS-OBS is compared favorably with other OBS schemes that target for efficient support of bursty traffic.