Multimodal Retrieval-Augmented Generation (MRAG) has shown promise in mitigating hallucinations in Multimodal Large Language Models (MLLMs) by incorporating external knowledge. However, existing methods typically adhere to rigid retrieval paradigms by mimicking fixed retrieval trajectories and thus fail to fully exploit the knowledge of different retrieval experts through dynamic interaction based on the model's knowledge needs or evolving reasoning states. To overcome this limitation, we introduce Mixture-of-Retrieval Experts (MoRE), a novel framework that enables MLLMs to collaboratively interact with diverse retrieval experts for more effective knowledge exploitation. Specifically, MoRE learns to dynamically determine which expert to engage with, conditioned on the evolving reasoning state. To effectively train this capability, we propose Stepwise Group Relative Policy Optimization (Step-GRPO), which goes beyond sparse outcome-based supervision by encouraging MLLMs to interact with multiple retrieval experts and synthesize fine-grained rewards, thereby teaching the MLLM to fully coordinate all experts when answering a given query. Experimental results on diverse open-domain QA benchmarks demonstrate the effectiveness of MoRE, achieving average performance gains of over 7% compared to competitive baselines. Notably, MoRE exhibits strong adaptability by dynamically coordinating heterogeneous experts to precisely locate relevant information, validating its capability for robust, reasoning-driven expert collaboration. All codes and data are released on https://github.com/OpenBMB/MoRE.

Few-shot classification is increasingly relevant in emerging applications, such as university course classification in intelligent education systems. University course classification helps students acquire specific skills, comprehend course purposes, and assists departments in defining training goals. However, classifying frontier courses presents challenges due to the absence of labels and descriptions. Few-shot learning addresses this by acquiring meta-knowledge. Heterogeneous graphs (HGs), rich in semantic information, introduce complexities that make few-shot particularly challenging. Addressing this problem, we propose a subgraph-aware convolutional few-shot classification method on HGs (HG-SCC). We first formalize the subgraph sampling strategy for HGs and different views under meta-paths. Then, the layer number adaptive spectral-based graph convolution is designed for personalized node embedding. Furthermore, a high-order convolution operation with classes as nodes is designed to increase the class representation coverage. Modeling subgraph centrality, combined with node features, captures structural information, improving awareness of each sampled subgraph, thus alleviating sparsity in new class labels and enhancing classification accuracy. Euclidean distance-based and task-affected cosine similarity-based classifiers under different meta-paths are proposed, with stacking introduced to blend multiple classifiers based on subgraph features. Experimental results show that our method has high performance in course classification and also outperforms state-of-the-art methods on benchmark datasets.

Knowledge tracing (KT) plays a crucial role in computer-aided education and intelligent tutoring systems, aiming to assess students' knowledge proficiency by predicting their future performance on new questions based on their past response records. While existing deep learning knowledge tracing (DLKT) methods have significantly improved prediction accuracy and achieved state-of-the-art results, they often suffer from a lack of interpretability. To address this limitation, current approaches have explored incorporating psychological influences to achieve more explainable predictions, but they tend to overlook the potential influences of historical responses. In fact, understanding how models make predictions based on response influences can enhance the transparency and trustworthiness of the knowledge tracing process, presenting an opportunity for a new paradigm of interpretable KT. However, measuring unobservable response influences is challenging. In this paper, we resort to counterfactual reasoning that intervenes in each response to answer what if a student had answered a question incorrectly that he/she actually answered correctly, and vice versa. Based on this, we propose RCKT, a novel response influence-based counterfactual knowledge tracing framework. RCKT generates response influences by comparing prediction outcomes from factual sequences and constructed counterfactual sequences after interventions. Additionally, we introduce maximization and inference techniques to leverage accumulated influences from different past responses, further improving the model's performance and credibility. Extensive experimental results demonstrate that our RCKT method outperforms state-of-the-art knowledge tracing methods on four datasets against six baselines, and provides credible interpretations of response influences. The source code is available at https://github.com/JJCui96IRCKT.

Location-based services have attracted significant attentions from both industry and academia, thanks to modern smartphones and mobile Internet. To provide users with gratifications, location-aware publish/subscribe has been recently proposed, which delivers spatio-textual messages of publishers to subscribers whose registered spatio-textual subscriptions are relevant to the messages. Since there could be large numbers of subscriptions, it is necessary to devise an efficient location-aware publish/subscribe system to enable instant message filtering. To this end, in this paper we propose two novel indexing structures, mbrtrie and PKQ. Using the indexes, we devise two filtering algorithms to support fast message filtering. We analyze the complexities of the two filtering algorithms and develop a cost-based model to judiciously select the best filtering algorithm for different scenarios. The experimental results show that our method achieves high performance and significantly outperforms the baseline approaches

Location-based services have been widely adopted in many systems. Existing works employ a pull model or user-initiated model, where a user issues a query to a server which replies with location-aware answers. To provide users with instant replies, a push model or server-initiated model is becoming an inevitable computing model in the next-generation location-based services. In the push model, subscribers register spatio-textual subscriptions to capture their interests, and publishers post spatio-textual messages. This calls for a high-performance location-aware publish/subscribe system to deliver publishers' messages to relevant subscribers. In this paper, we address the research challenges that arise in designing a location-aware publish/subscribe system. We propose an R-tree based index by integrating textual descriptions into R-tree nodes. We devise efficient filtering algorithms and effective pruning techniques to achieve high performance. Our method can support both conjunctive queries and ranking queries. We discuss how to support dynamic updates efficiently. Experimental results show our method achieves high performance which can filter 500 messages in a second for 10 million subscriptions on a commodity computer.

SimRank is a popular and widely-adopted similarity measure to evaluate the similarity between nodes in a graph. It is time and space consuming to compute the SimRank similarities for all pairs of nodes, especially for large graphs. In real-world applications, users are only interested in the most similar pairs. To address this problem, in this paper we study the top-k SimRank-based similarity join problem, which finds k most similar pairs of nodes with the largest SimRank similarities among all possible pairs. To the best of our knowledge, this is the first attempt to address this problem. We encode each node as a vector by summarizing its neighbors and transform the calculation of the SimRank similarity between two nodes to computing the dot product between the corresponding vectors. We devise an efficient two-step framework to compute top- k similar pairs using the vectors. For large graphs, exact algorithms cannot meet the high-performance requirement, and we also devise an approximate algorithm which can efficiently identify top- k similar pairs under user-specified accuracy requirement. Experiments on both real and synthetic datasets show our method achieves high performance and good scalability.

With the exponential increment of data, compression technology becomes an important tool in the field of data management, especially in text management. An increasing pressing challenge is how to efficiently query these massive amounts of sequence data in their compressed format. In this paper we study the problem of answering subsequence-search queries on LZ78 format of texts. We propose the concept of conditional common sub strings of queries to improve query performance. We present a techniques to find minimal conditional common sub strings in compressed text and a local uncompressing technique to verify and locate positions of answers in text. Finally, the experimental results over real data demonstrate the efficiency of our algorithm.