兰州大学 http:///ir.lzu.edu.cn/:80 2026-09-10T14:26:43Z 2026-09-10T14:26:43Z Y. Bao Y. Lei E. Li L. Chen L. Wei W. Zhao http:///ir.lzu.edu.cn/:80/handle/262010/886303 2026-09-09T09:23:15Z 2026-09-09T09:23:13Z

Title: Redundancy-Aware Graph Clustering Guided Framework for Swarm Intelligence Feature Selection Authors: Y. Bao; Y. Lei; E. Li; L. Chen; L. Wei; W. Zhao Description: Feature selection is an essential preprocessing step in data analysis because it can improve predictive performance while reducing feature redundancy. However, most existing methods either insufficiently exploit inter-feature dependencies or search in an unstructured feature space, which may lead to redundant subsets and inefficient exploration. In particular, swarm intelligence-based feature selection methods lack an explicit mechanism for handling feature correlations during subset construction. To address this issue, this paper proposes a graph-clustering-guided redundancy-aware (GCR) framework for swarm-intelligence-based feature selection. A feature dependency graph is constructed to model pair-wise relationships, and community detection is employed to partition features into redundancy groups. Based on this structured representation, a GCR mechanism is developed to integrate relevance evaluation, redundancy suppression, and diversity promotion during optimization. Experimental results on benchmark datasets demonstrate that the proposed framework improves classification accuracy while selecting informative and less redundant feature subsets.

2026-09-09T09:23:13Z Z. Wang Y. Jiang R. Zhou B. Zhang F. Zhang Z. Xu Y. Zhang J. Wang http:///ir.lzu.edu.cn/:80/handle/262010/886302 2026-09-09T09:13:53Z 2026-09-09T09:13:52Z

Title: DriveCode: Domain Specific Numerical Encoding for LLM-Based Autonomous Driving Authors: Z. Wang; Y. Jiang; R. Zhou; B. Zhang; F. Zhang; Z. Xu; Y. Zhang; J. Wang Description: Large language models (LLMs) have shown great promise for autonomous driving. However, discretizing numbers into tokensmay weaken precise numerical modeling, rely mainly on positional encoding to represent digit place values, and can make it challenging to balance decoding efficiency with numerical precision. These limitations affect both the processing of sensor measurements and the generation of precise control commands, posing a practical challenge for LLM-based autonomous driving systems that require accurate numerical prediction. In this paper, we introduce DriveCode, a novel numerical encoding method that represents numbers as dedicated embeddings rather than discrete text tokens. DriveCode employs a number projector to map numbers into the language model's hidden space, enabling seamless integration with visual and textual features in a unified multimodal sequence. Evaluated on OmniDrive, DriveGPT4, and DriveGPT4-V2 datasets, DriveCode demonstrates superior performance in trajectory prediction and control signal generation, confirming its effectiveness for LLM-based autonomous driving systems.

2026-09-09T09:13:52Z Z. Hao M. Han Y. Wang G. Wang Z. Liu http:///ir.lzu.edu.cn/:80/handle/262010/886301 2026-09-09T09:13:04Z 2026-09-09T09:13:03Z

Title: CoTGait: A Co-training-based Semi-supervised Framework for mmWave Radar Gait Recognition Authors: Z. Hao; M. Han; Y. Wang; G. Wang; Z. Liu Description: In Internet of Things (IoT) application scenarios, millimeter-wave (mmWave) radar–based gait recognition has become an important direction for identity authentication. This is largely because it is wireless and contactless, and it maintains reliable performance through certain obstructions. However, current methods still run into two practical bottlenecks. First, many studies depend on large-scale radar datasets, while collecting and annotating such data is expensive, slow, and labor-intensive. Second, the model lacks sufficient robustness in feature extraction when dealing with samples of low discriminability. To tackle these issues, we propose CoTGait, a semi-supervised mmWave gait recognition method based on a collaborative training strategy. Starting with a small labeled set, we train three heterogeneous models and use them to predict unlabeled data. Notably, we do not treat all pseudo-labels equally; instead, confidence and inter-model consistency metrics are utilized to partition them into high-quality labels and hard samples. To enhance the quality of pseudo-labels, we further introduce a Hard-Sample Adaptive Learning (HSAL) strategy, which is capable of extracting discriminative cues from hard samples. Finally, the dataset is continuously optimized by collaborative training, and the dual-stage attention network ConvSAT is taken as the final network for gait classification. Experiments on our self-collected MMSGait dataset and the public MMRGait-1.0 dataset show accuracies of 94.1% and 91.8%, respectively. Additional results also suggest that CoTGait remains robust across conditions.

2026-09-09T09:13:03Z X. Fan H. Xin Z. Wang W. Hou Q. Fei http:///ir.lzu.edu.cn/:80/handle/262010/886300 2026-09-09T09:12:44Z 2026-09-09T09:12:43Z

Title: Assessing the Reliability of LLM-Based Architectural Design Image Generation: A Comprehensive Evaluation Framework and Benchmark (AGEB) Authors: X. Fan; H. Xin; Z. Wang; W. Hou; Q. Fei Description: Text-to-image systems have progressed from research prototypes to widely deployed tools, but high-fidelity imagery alone does not satisfy the requirements of professional architectural design. The Architecture Generation and Evaluation Benchmark (AGEB) is introduced as an end-to-end automated benchmark that assesses the reliability of architectural image generation along four axes: semantic correspondence with the design brief, a prompt-derived circulation proxy, perspective geometry, and no-reference technical quality. AGEB consists of 300 tasks organized into six cognitive levels, a unified generation protocol applied to five representative systems, five independent image generations per prompt, and a four-module evaluation pipeline combining a dual-expert chain-of-thought (COT) reasoning procedure, graph-theoretic circulation analysis, classical geometric consistency checks, and no-reference quality metrics (NIQE, BRISQUE, PIQE, Inception Score). Within this benchmark setting, the large language model is used only as an automated reasoning evaluator in the COT module, not as one of the scored image generators. Under one normalized prompt per task and five generated images per prompt for each system, the five systems exhibit differentiated strengths: GPT-Image-1 obtains the highest COT and circulation-proxy scores, Sora leads on BRISQUE, Inception Score, and PIQE, DALL-E 3 attains the highest perspective consistency, and Midjourney reaches the lowest NIQE. All five systems, however, score below 0.35 on the circulation proxy, indicating limited functional-programme compatibility within the controlled AGEB setting rather than verified pixel-level spatial functionality. Data and code are released at https://github.com/torfqy/Architecture-Generation-and-Evaluation-Benchmark-AGEB-.

2026-09-09T09:12:43Z Y. Li Y. Cheng J. Huang B. Feng J. Zhu X. Li B. Hu http:///ir.lzu.edu.cn/:80/handle/262010/886299 2026-09-09T09:11:59Z 2026-09-09T09:11:58Z

Title: Distill to Diagnose: Toward Stable and Reliable Cross-Modal KD Strategies for Depression Authors: Y. Li; Y. Cheng; J. Huang; B. Feng; J. Zhu; X. Li; B. Hu Description: Multimodal fusion improves depression recognition but incurs high acquisition and deployment costs, limiting its practicality in real-world settings. Cross-modal knowledge distillation (KD) addresses this issue by using multimodal data during training while enabling unimodal inference at test time. However, existing studies mainly focus on architectural design and often overlook the stability and reliability of the transfer process under modality discrepancy and signal quality variation. In this work, we focus on stable and reliable cross-modal KD strategies for depression recognition. Specifically, we observe that large conditional variance in teacher features, conditioned on student-accessible modalities, can destabilize knowledge transfer. To address this issue, we introduce a conditional variance constraint to suppress such uncertainty and improve the stability of teacher-side supervision. In addition, we propose a quality-aware weighting (QAW) strategy to adaptively adjust the contribution of training samples according to modality reliability. These two strategies can be incorporated into the training process to provide more stable and reliable supervisory signals for the student model. Experiments on one synthetic dataset and two real-world multimodal depression datasets (EFVP and AADP) demonstrate consistent improvements in accuracy and F1-score, supporting more effective unimodal depression recognition in resource-constrained healthcare and everyday environments.

2026-09-09T09:11:58Z H. Chen H. He S. Sun J. Li X. Cui J. Zhu X. Li B. Hu http:///ir.lzu.edu.cn/:80/handle/262010/886298 2026-09-09T09:11:32Z 2026-09-09T09:11:31Z

Title: An Offline Affective Brain-Computer Interface Paradigm with CycleGAN Data Augmentation and Meta-Learning Update Mode Authors: H. Chen; H. He; S. Sun; J. Li; X. Cui; J. Zhu; X. Li; B. Hu Description: Data acquisition and system update are key parts of affective brain-computer interface (aBCI). Facing the issue of data scarcity, supplementing data from real samples of other subjects is a feasible solution. However, only data from subjects with emotional patterns (EP) similar to those of the user can be used for initializing and updating the aBCI. Therefore, we proposed an inter-subject data supplementation solution based on Cycle-Consistent Generative Adversarial Network (CycleGAN). With the introduction of EP metrics, information reconstruction and data augmentation networks can mitigate individual emotional pattern differences (IEPD). Based on these networks, a one-to-one inter-subject data sharing community was established. Even with data showing significant emotional pattern differences, it can still be transformed into data compatible with the user. Given that the system update mode for aBCI remains undefined, we proposed a meta-learning system update mode to learn the user's EP knowledge. Recursive meta-learning is employed to learn emotional information from a multi-task dataset. A temporary meta-learning paradigm for emotion recognition is designed to enable the system to adapt to unknown session tasks. After multi-session updates, an EP record system is used to summarize user-specific emotion-related feature patterns accumulated across sessions. The proposed paradigm is validated with a multi-session sequential learning process, yielding performances of 93.33 ± 7.00% on SEED, 77.08 ± 9.07% on SEED-FRA, and 71.87 ± 7.77% on SEED-GER. This research highlights the importance of offline assistive technologies, providing a novel perspective and an initial framework for the offline operational mechanisms of aBCI.

2026-09-09T09:11:31Z Yao, Wei Liu, Hang Liu, Jizhao Wu, Hongwei Zhang, Xin Wen, Shiping http:///ir.lzu.edu.cn/:80/handle/262010/886297 2026-09-09T08:47:35Z 2026-09-09T08:47:33Z

Title: An Image Segmentation Method Based on Continuous Coupled Neural Network and Parameter Optimization Authors: Yao, Wei; Liu, Hang; Liu, Jizhao; Wu, Hongwei; Zhang, Xin; Wen, Shiping Description: Visual cortex neural networks are widely employed in image segmentation because of their efficient visual feature extraction capability. Among visual cortex neural networks, continuous coupled neural network (CCNN) performs well in processing images because it has a simple structure and few parameters. This paper proposes a simplified CCNN (SCCNN) model and an image segmentation method based on SCCNN. In natural image segmentation experiments, SCCNN preserves more textures and details; in medical image segmentation experiments, the overall performance of SCCNN is nearly 4% better than that of CCNN. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2027.

2026-09-09T08:47:33Z Su, Yuchi Niu, Weina Gong, Jiacheng Yan, Ran Li, Song Liu, Xin Zhang, Xiaosong http:///ir.lzu.edu.cn/:80/handle/262010/886296 2026-09-09T08:47:29Z 2026-09-09T08:47:28Z

Title: RlDecompiler: Enhancing LLM-based Decompilation via Reinforcement Learning with a Multi-Faceted Reward Function Authors: Su, Yuchi; Niu, Weina; Gong, Jiacheng; Yan, Ran; Li, Song; Liu, Xin; Zhang, Xiaosong Description: Decompiling binary code into human-readable, high-level source code is a core challenge in reverse engineering. While traditional methods often rely on brittle, pattern-based heuristics, the advent of Large Language Models (LLMs) offers a more flexible and robust approach. However, current LLM-based decompilation efforts are often limited by their training methodologies, which typically treat the task as a simple sequence-to-sequence translation and struggle to enforce the functional correctness of the output. To address these issues, this paper proposes an innovative framework for training LLMs to perform high-fidelity decompilation. A core contribution of our work is a novel data processing pipeline that enriches the model's input. This pipeline integrates Ghidra-based static analysis to directly embed crucial context, such as static resources (strings, floating-point numbers) and relabeled basic blocks - from the binary into an LLM-friendly prompt. Building on this enriched input, we employ reinforcement learning fine-tuning guided by a multi-faceted reward function that comprehensively evaluates syntactic correctness, AST similarity, compilability, and functional correctness via test cases. Using this framework, we trained the RlDecompiler family of models (1.3B and 3B). Experimental results demonstrate that RlDecompiler achieves state-of-the-art performance, and its generated code quality is also higher than that of the baseline models. The RlDecompiler 1.3B and 3B models achieve rerunnable rates of 27.96% and 40.70%, respectively, outperforming existing baselines. The code is available at https://github.com/ri-char/rldecompile. © 2026 Copyright held by the owner/author(s).

2026-09-09T08:47:28Z Cheng, Jianjun Ren, Bo Ma, Zhixin http:///ir.lzu.edu.cn/:80/handle/262010/886295 2026-09-09T08:47:27Z 2026-09-09T08:47:25Z

Title: Shortest Path Tree Maximize Node Influence Propagation in Complex Networks Authors: Cheng, Jianjun; Ren, Bo; Ma, Zhixin Description: Influence maximization has been one hot topic in the field of network analysis in recent years. The essence of influence maximization is to find a given number, k, seed nodes, so that their influence can diffuse to as many other nodes as possible in the network. Considering the timeliness and the value attenuation of information in the transmission, we believe that information spreading along the shortest path from one node to another is the fastest way between the pair of nodes. The shortest paths originated from any node form a shortest-path tree, and using the shortest-path tree to estimate the influence of the root node is an effective approach. Motivated by this, we propose an influence maximization algorithm based on the shortest-path tree in this paper, we name it SPTIM (acronym for Shortest Path Tree based Influence Maximization). Firstly, for every node, the shortest paths are identified to form its shortest-path tree, then its influence is estimated according to the rules developed in this paper. Afterwards, the node with the largest estimated influence is selected as a seed. Then, To alleviate the problem of influence overlapping, we attenuate the one- and two-hop neighbors of the selected seed with different proportions considering the impact of community structure. Next, the node with the largest influence is selected as another seed accordingly. This procedure is repeated until the number of selected nodes reaches k. The experimental results on some real networks under the IC, LT, and SIR models show that the proposed method, SPTIM, is competitive in terms of performance and efficiency. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.

2026-09-09T08:47:25Z Wen, Jingkai Guo, Lan Zhou, Qingguo http:///ir.lzu.edu.cn/:80/handle/262010/886294 2026-09-09T08:47:23Z 2026-09-09T08:47:22Z

Title: Few-Shot Semantic Segmentation via Latent Knowledge Mining and Dense Feature Alignment Authors: Wen, Jingkai; Guo, Lan; Zhou, Qingguo Description: Few-shot semantic segmentation aims to segment novel classes with minimal annotated support samples. Existing methods extract global or local prototypes for feature matching, but such methods often ignore the rich latent knowledge embedded in the deep feature space of pre-trained vision models. When facing complex background interference and significant intra-class morphological variations, relying solely on limited explicit appearance prototypes makes it difficult to maintain robust semantic consistency. To address this limitation, we propose a latent knowledge activation mechanism that dynamically extracts category-relevant deep priors from the backbone through support guidance, enriching feature representation. Subsequently, to address the spatial misalignment between support and query images, we design a contextual dense alignment module to guide the activated latent knowledge to achieve pixel-level precise matching and information transfer within the query image features. This cross-level mechanism effectively suppresses the semantic interference of complex background noise and translates abstract latent features into explicit local boundary segmentation clues. We conduct extensive experiments on standard few-shot segmentation benchmark datasets. Both qualitative and quantitative evaluation results show that our method significantly outperforms existing advanced baseline models, not only substantially improving the detail fidelity of target edges but also demonstrating outstanding cross-category generalization capabilities. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2027.

2026-09-09T08:47:22Z