Retrieval-Augmented Generation (RAG) plays a crucial role in grounding Large Language Models by leveraging external knowledge, whereas the effectiveness is often compromised by the retrieval of contextually flawed or incomplete information. To address this, knowledge graph-based RAG methods have evolved towards hierarchical structures, organizing knowledge into multi-level summaries. However, these approaches still suffer from two critical, unaddressed challenges: high-level conceptual summaries exist as disconnected ``semantic islands'', lacking the explicit relations needed for cross-community reasoning; and the retrieval process itself remains structurally unaware, often degenerating into an inefficient flat search that fails to exploit the graph's rich topology. To overcome these limitations, we introduce LeanRAG, a framework that features a deeply collaborative design combining knowledge aggregation and retrieval strategies. LeanRAG first employs a novel semantic aggregation algorithm that forms entity clusters and constructs new explicit relations among aggregation-level summaries, creating a fully navigable semantic network. Then, a bottom-up, structure-guided retrieval strategy anchors queries to the most relevant fine-grained entities and then systematically traverses the graph's semantic pathways to gather concise yet contextually comprehensive evidence sets. The LeanRAG can mitigate the substantial overhead associated with path retrieval on graphs and minimize redundant information retrieval. Extensive experiments on four challenging QA benchmarks with different domains demonstrate that LeanRAG significantly outperforms existing methods in response quality while reducing 46% retrieval redundancy.

Low Earth Orbit (LEO) satellite networks are characterized by dynamic network topologies and on-demand service requirements from Internet-of-Things (IoT) applications, which make efficient and intelligent flow scheduling challenging. Conventional schemes rely on static configurations or manual rules, thus making it difficult to capture and respond to diverse service demands. Moreover, they often fail to model task–resource relationships effectively, hindering the generation of real-time, executable scheduling policies. To address these challenges, we propose a knowledge-enhanced, intent-driven flow scheduling (KIFS) framework. Specifically, we design a unified pipeline that first translates user intents into precise Quality of Service (QoS) requirements. It incorporates a network state awareness module to estimate per-link bandwidth and utilization, and constructs a task–resource knowledge graph (KG) to enhance the Deep Q-Network (DQN) agent via state augmentation, action pruning, and reward shaping. Finally, the framework translates the resulting policies into standards-compliant SRv6 configurations for real-time deployment. In simulations, the proposed KIFS framework demonstrates superior performance compared to standard baselines in terms of flow success rate and QoS satisfaction.

Multimodal reasoning in Large Language Models (LLMs) struggles with incomplete knowledge and hallucination artifacts, challenges that textual Knowledge Graphs (KGs) only partially mitigate due to their modality isolation. While Multimodal Knowledge Graphs (MMKGs) promise enhanced cross-modal understanding, their practical construction is impeded by semantic narrowness of manual text annotations and inherent noise in visual-semantic entity linkages. In this paper, we propose Vision-align-to-Language integrated Knowledge Graph (VaLiK), a novel approach for constructing MMKGs that enhances LLMs reasoning through cross-modal information supplementation. Specifically, we cascade pre-trained Vision-Language Models (VLMs) to align image features with text, transforming them into descriptions that encapsulate image-specific information. Furthermore, we developed a cross-modal similarity verification mechanism to quantify semantic consistency, effectively filtering out noise introduced during feature alignment. Even without manually annotated image captions, the refined descriptions alone suffice to construct the MMKG. Compared to conventional MMKGs construction paradigms, our approach achieves substantial storage efficiency gains while maintaining direct entity-to-image linkage capability. Experimental results on multimodal reasoning tasks demonstrate that LLMs augmented with VaLiK outperform previous state-of-the-art models. Our code is published at https://github.com/Wings-Of-Disaster/VaLiK.

Recent advancements in Large Multi-modal Models (LMMs) underscore the importance of scaling by increasing image-text paired data, achieving impressive performance on general tasks. Despite their effectiveness in broad applications, generalist models are primarily trained on web-scale datasets dominated by natural images, resulting in the sacrifice of specialized capabilities for domain-specific tasks that require extensive domain prior knowledge. Moreover, directly integrating expert models tailored for specific domains is challenging due to the representational gap and imbalanced optimization between the generalist model and experts. To address these challenges, we introduce Chimera, a scalable and low-cost multi-modal pipeline designed to boost the ability of existing LMMs with domain-specific experts. Specifically, we design a progressive training strategy to integrate features from expert models into the input of a generalist LMM. To address the imbalanced optimization caused by the well-aligned general visual encoder, we introduce a novel Generalist-Specialist Collaboration Masking (GSCM) mechanism. This results in a versatile model that excels across the chart, table, math, and document domains, achieving state-of-the-art performance on multi-modal reasoning and visual content extraction tasks, both of which are challenging tasks for assessing existing LMMs.

To overcome the challenges of dynamic topology and intermittent connectivity in large-scale Low Earth Orbit (LEO) satellite networks, we propose an Intent-Driven routing control framework leveraging Segment Routing over Internet Protocol version 6 (SRv6). The proposed framework enables fine-grained control over routing behaviors while maintaining protocol compatibility through packet-level semantic intent identifiers. The novel contributions include: (1) an Intent-Driven Segment Routing framework supporting diverse routing requirements with packet-level control; (2) an adaptive Polar Region Link Handling Mechanism integrated with dynamic load balancing; and (3) a multi-path-based fault recovery mechanism with rapid convergence characteristics. Experiments on Network Simulator 3 (NS3) platform demonstrate that, compared to traditional routing protocols, the proposed framework reduces end-to-end delay by 64.1% for the Iridium constellation scenario and improves stability by 68.8% in the OneWeb constellation environment.

Traffic flow simulation within the domain of intelligent transportation systems is garnering significant attention, and generating realistic, diverse, and human-like traffic patterns presents critical challenges that must be addressed. Current approaches often hinge on predefined driver models, objective optimization, or reliance on pre-recorded driving datasets, imposing limitations on their scalability, versatility, and adaptability. In this paper, we introduce TrafficMCTS, an innovative framework that harnesses the synergy of group-based Monte Carlo tree search (MCTS) and Social Value Orientation (SVO) to engender a multifaceted traffic flow with varying driving styles and cooperative tendencies. Anchored by a closed-loop architecture, our framework enables vehicles to dynamically adapt to their environment in real time, and ensure feasible collision-free trajectories. Through comprehensive comparisons with state-of-the-art methods, we illuminate the advantages of our approach in terms of computational efficiency, planning success rate, intention completion time, and diversity metrics. Besides, we simulate multiple scenarios to illustrate the effectiveness of the proposed framework and highlight its ability to induce diverse social behaviors within the traffic flow. Finally, we validate the scalability of TrafficMCTS by demonstrating its capability to efficiently simulate diverse traffic scenarios involving numerous interacting vehicles within a complex road network, capturing the intricate dynamics of human-like driving behaviors.

The emergence of Multimodal Large Language Models ((M)LLMs) has ushered in new avenues in artificial intelligence, particularly for autonomous driving by offering enhanced understanding and reasoning capabilities. This paper introduces LimSim++, an extended version of LimSim designed for the application of (M)LLMs in autonomous driving. Acknowledging the limitations of existing simulation platforms, LimSim++ addresses the need for a long-term closed-loop infrastructure supporting continuous learning and improved generalization in autonomous driving. The platform offers extended-duration, multi-scenario simulations, providing crucial information for (M)LLM-driven vehicles. Users can engage in prompt engineering, model evaluation, and framework enhancement, making LimSim++ a versatile tool for research and practice. This paper additionally introduces a baseline (M)LLM-driven framework, systematically validated through quantitative experiments across diverse scenarios. The open-source resources of LimSim++ are available at: https://pjlab-adg.github.io/limsim-plus/.

Class syntax can be used to 1) model temporal or locational evolvement of class labels of feature observation sequences, 2) correct classification errors of static classifiers if feature observations from different classes overlap in feature space, and 3) eliminate redundant features whose discriminative information is already represented in the class syntax. In this paper, we describe a novel method that combines static classifiers with class syntax models for supervised feature subset selection and classification in unified algorithms. Posterior class probabilities given feature observations are first estimated from the output of static classifiers, and then integrated into a parsing algorithm to find an optimal class label sequence for the given feature observation sequence. Finally, both static classifiers and class syntax models are used to search for an optimal subset of features. An optimal feature subset, associated static classifiers, and class syntax models are all learned from training data. We apply this method to logical entity recognition in scanned historical U.S. Food and Drug Administration (FDA) documents containing court case Notices of Judgments (NJs) of different layout styles, and show that the use of class syntax models not only corrects most classification errors of static classifiers, but also significantly reduces the dimensionality of feature observations with negligible impact on classification performance.

Logical entity recognition in heterogeneous collections of document page images remains a challenging problem since the performance of traditional supervised methods degrades dramatically in case of many distinct layout styles. In this paper we present an unsupervised method where layout style information is explicitly used in both training and recognition phases. We represent the layout style, local features, and logical labels of physical regions of a document compactly by an ordered labeled X-Y tree. Style dissimilarity of two document pages is represented by the distance between their respective trees. During the training phase, document pages with true logical labels in training set are classified into distinct layout styles by unsupervised clustering. During the recognition phase, the layout style and logical entities of an input document are recognized simultaneously by matching the input tree to the trees in closest- matched layout style cluster of training set. Experimental results show that our algorithm is robust with both balanced and unbalanced style cluster sizes, zone over-segmentation, zone length variation, and variation in tree representations of the same layout style.

Style classification of document page images is crucial for logical structure analysis of heterogeneous collections of documents. Both layout and contextual features contain significant information about document styles. Most existing methods are supervised methods in which specific document models or classifiers are learned from a training set of document page images with known class labels. In this paper, we propose an unsupervised classification method that involves no training or manual selection of algorithm parameters. In particular, we first represent each document page as an ordered labeled X-Y tree. A tree matching algorithm is then used to compute style dissimilarity between two document pages. We propose a set of tree edit cost functions based on Karl Pearson distance between two multivariate feature observations, which is robust to the over-segmentation problem and zone length variations of same logical entities. Finally, the K-medoids algorithm is used to find an optimal grouping of the trees into K clusters, each of which corresponds to a distinct document style. We evaluate our algorithm on test datasets with different cluster sizes and degrees of style similarity. Experimental results show our algorithm achieved an average classification accuracy of 95.69% over six datasets consisting of 150 pages of 11 different styles.

Structural information about a document is essential for structured query processing, indexing, and retrieval. A document page can be partitioned into a hierarchy of homogeneous regions such as columns, paragraphs, etc.; these regions are called physical components, and define the physical layout of the page. In this paper we develop a class of models for the physical layouts of technical paper title pages. We model physical layout using hidden semiMarkov models for directional projections of page regions, and a stochastic attributed K-d tree grammar model for the 2D hierarchical structure of these regions. We use the models to generate sets of synthetic title page images of three distinctive styles, which we use in controlled experiments on page structure analysis.

Image segmentation is an important component of any document image analysis system. While many segmentation algorithms exist in the literature, very few i) allow users to specify the physical style, and ii) incorporate user-specified style information into the algorithm's objective function that is to be minimized. We describe a segmentation algorithm that models a document's physical structure as a hierarchical structure where each node describes a region of the document using a stochastic regular grammar. The exact form of the hierarchy and the stochastic language is specified by the user, while the probabilities associated with the transitions are estimated from groundtruth data. We demonstrate the segmentation algorithm on images of bilingual dictionaries.

Parsing bilingual dictionaries is important for building cross-language retrieval systems and speech recognition algorithms. We describe a general purpose algorithm that can be easily modified to convert printed bilingual dictionaries in various layouts and language pairs into electronic/symbolic lexicons. In a previous paper [SPIE Document Recognition and Retrieval, San Jose, January 2002], we described an algorithm for segmenting the physical layout of dictionaries into columns and lines. In this paper we assume that the physical lines are given then recognize the lines that constitute a dictionary item. Furthermore, the algorithm simultaneously recognizes the logical structure within the dictionary items (head-word, pronounciation, part of speech and definition). We demonstrate our algorithm on 30 scanned Chinese-English dictionary pages which include more than 2500 lexicon items.

While numerous page segmentation algorithms have been proposed in the literature, there is lack of comparative evaluation of these algorithms. In the existing performance evaluation methods, two crucial components are usually missing: 1) automatic training of algorithms with free parameters and 2) statistical and error analysis of experimental results. We use the following five-step methodology to quantitatively compare the performance of page segmentation algorithms: 1) first, we create mutually exclusive training and test data sets with groundtruth, 2) we then select a meaningful and computable performance metric, 3) an optimization procedure is then used to search automatically for the optimal parameter values of the segmentation algorithms on the training data set, 4) the segmentation algorithms are then evaluated on the test data set, and, finally, 5) a statistical and error analysis is performed to give the statistical significance of the experimental results. In particular, instead of the ad hoc and manual approach typically used in the literature for training algorithms, we pose the automatic training of algorithms as an optimization problem and use the simplex algorithm to search for the optimal parameter value. A paired-model statistical analysis and an error analysis are then conducted to provide confidence intervals for the experimental results of the algorithms. This methodology is applied to the evaluation of live page segmentation algorithms of which, three are representative research algorithms and the other two are well-known commercial products, on 978 images from the University of Washington III data set. It is found that the performance indices of the Voronoi, Docstrum, and Caere segmentation algorithms are not significantly different from each other, but they are significantly better than that of ScanSoft's segmentation algorithm, which, in turn, is significantly better than that of X-Y cut.

Most page segmentation algorithms have user-specifiable free parameters. However, algorithm designers typically do not provide a quantitative/rigorous method for choosing values for these parameters. The free parameter values can affect the segmentation result quite drastically and are very dependent on the particular dataset that the algorithm is being used on. We present an automatic training method for choosing free parameters of page segmentation algorithms. The automatic training problem is posed as a multivariate non-smooth function optimization problem. An efficient direct search method-simplex method-is used to solve this optimization problem. This training method is used applied to the training of Kise's page segmentation algorithm. It is found that a set of optimal parameter values and their corresponding performance index can be found using relatively few function evaluations. The UW III dataset was used for conducting our experiments.