Retrieval-Augmented Generation (RAG) over dynamically growing long-text corpora demands indexing mechanisms that are both efficient and update-friendly. Prior graph- and hierarchy-augmented RAG systems often rely on LLM-based summarization or entity–relation extraction during indexing and maintenance, incurring high construction latency and substantial token cost. We propose MCHRAG, a hierarchical semantic routing framework based on a Multi-Centroid Routing Tree designed for streaming updates. MCHRAG first partitions embeddings into coarse hash buckets via hyperplane locality-sensitive hashing, then builds a coarse-to-fine routing hierarchy in the embedding space, where each tree node (including leaf buckets) is represented by a small set of vector prototypes with radius constraints. Both retrieval and incremental insertion follow the same top-down routing to select a small set of leaf buckets, followed by exact similarity reranking within routed candidates, enabling bounded candidate size with high recall. By eliminating LLM-dependent indexing and maintenance, MCHRAG achieves zero indexing and update token consumption while substantially reducing construction and update latency. Extensive experiments on long-context and multi-document QA benchmarks show that MCHRAG matches or improves retrieval quality, especially recall under distractor-heavy settings, with significantly lower build and update costs.

Deep cross-modal hashing has achieved remarkable success in cross-modal retrieval due to its fast retrieval speed and low storage cost, but it highly vulnerable to adversarial attacks. Mainstream defense methods rely on adversarial training, which often induces a robustness–standard performance trade-off, leading to the learning of only limited robust features and degraded standard performance. To address this issue, we propose a Spatial-Frequency Domain Complementary Learning (SFCL) framework to overcome these two challenges by: 1) exploiting the complementarity of spatial and frequency features to learn more comprehensive and adversarially robust features, addressing the limited robustness of existing defenses; 2) by supplementing frequency-domain information, it avoids the performance degradation commonly caused by adversarial training. Specifically, SFCL consists of two modules: a Spatial-Frequency Robust Gating (SFRG) module, which selects robust features and strengthens complementarity via a conditional mutual information-based loss; and a Robustness-Aware Feature Fusion (RAFF) module, which performs bidirectional feature interaction and fusion. Extensive experiments demonstrate significant robustness gains over existing state-of-the-art methods, along with improved standard performance.

Deep hashing is widely used in large-scale image retrieval systems due to its efficient retrieval performance. However, its susceptibility to adversarial attacks limits its security in practical applications. Adversarial training is the most effective method for improving robustness, but it often leads to a significant trade-off between robustness and retrieval accuracy. In this paper, we conduct spectral analysis and find that generating high-quality hash codes requires wide-frequency response models, whereas adversarial training forces the model into spectral collapse, degrading it to a low-frequency response model and weakening its discriminability. To address this issue, we propose a Spectral-Adaptive Adversarial Hashing (SAAH) framework, which selectively preserves discriminative and task-relevant frequency components while suppressing adversarially unstable ones, enabling robust hashing without sacrificing retrieval performance. Extensive experiments on benchmark datasets demonstrate that SAAH consistently achieves a superior balance between retrieval accuracy and adversarial robustness, achieving the best performance in both retrieval accuracy and robustness compared with existing robust hashing methods.

Vision-Language Models (VLMs) play a crucial role in the advancement of Artificial General Intelligence (AGI). As AGI rapidly evolves, addressing security concerns has emerged as one of the most significant challenges for VLMs. In this paper, we present extensive experiments that expose the vulnerabilities of conventional adaptation methods for VLMs, highlighting significant security risks. Moreover, as VLMs grow in size, the application of traditional adversarial adaptation techniques incurs substantial computational costs. To address these issues, we propose a parameter-efficient adversarial adaptation method called AdvLoRA based on Low-Rank Adaptation. We investigate and reveal the inherent low-rank properties involved in adversarial adaptation for VLMs. Different from LoRA, we enhance the efficiency and robustness of adversarial adaptation by introducing a novel reparameterization method that leverages parameter clustering and alignment. Additionally, we propose an adaptive parameter update strategy to further bolster robustness. These innovations enable our AdvLoRA to mitigate issues related to model security and resource wastage. Extensive experiments confirm the effectiveness and efficiency of AdvLoRA.

The aim of cross-modal retrieval is to search for flexible results across different types of multimedia data. However, the labeled data is usually limited and not well paired with different modalities in practical applications. These issues are not well addressed in the existing works, which cannot consider the semantic information about unlabeled and unpaired data, synchronously. Self-training is a well-known strategy to handle semi-supervised problems. Motivated by the self-training, this paper proposes a self-training-based cross-modal hashing framework (STCH) to tackle the semi-supervised and semi-paired challenges. In the framework, graph neural networks are used to capture potential intra-modality and inter-modality similarities to produce pseudo labels. Then the inconsistent pseudo labels of different modalities are refined with a heuristic filter to enhance the model robustness. To train STCH, we propose an alternating learning strategy to conduct the self-train by predicting pseudo labels during the training procedure, which can be seamlessly incorporated into semi-supervised and supervised learning. In this way, the proposed method can leverage sufficient semantic information to enhance the semi-supervised effect and address the semi-paired problem. Experiments on the real-world datasets demonstrate that our approach outperforms related methods on hash cross-modal retrieval.