In recent years, improvements in decompilers and large language models (LLMs) have made it possible to use powerful techniques for analyzing software binaries, including automated function name recovery. Such powerful binary analysis techniques can be exploited for Man-At-The-End (MATE) attacks, posing a serious threat to the software ecosystem. To protect software against MATE attacks, software obfuscation is widely used. However, it is not clear how effectively code obfuscation can prevent LLM-based MATE attacks, and thus a method for evaluating the resilience of obfuscated code against such attacks is required. This paper proposes a framework for evaluating the resilience of obfuscated code against LLM-based function name recovery MATE attacks. To construct the framework, resilience against the LLM-based function name recovery attack is characterized. The resilience is defined based on the morphological or semantic similarity between the original function name and the recovered function names. In the experiment, the resilience of five well-known obfuscation methods against LLM-based function name recovery attacks (i.e., how effectively these obfuscation methods prevent the prediction of appropriate function names) using the Gemma3-27B-IT, gpt-oss-20b, and llama-4-scout-17b-16e-instruct LLM general-purpose models is evaluated. The results reveal that the obfuscation method based on code virtualization has the highest resilience, while the method based on encoding arithmetic has the lowest resilience. In addition, the resilience of some obfuscation methods varies depending on the model. Furthermore, the analysis of high-resilience obfuscated code indicates that LLM-recovered function names for functions evaluated as high resilience are relatively less related to the implementation of the original functions. This paper makes two contributions: the development of a framework for evaluating the resilience of obfuscated code against LLM-based function name recovery MATE attacks, and an experimental evaluation of the resilience of existing obfuscation methods against such attacks.

Goal-Oriented Requirements Analysis (GORA) is a technique, in which 1) customer requirements are captured as goals, 2) then they are iteratively and incrementally refined and 3) finally a goal model is constructed in order to derive system and software requirements. However, the lack of comprehensive guidelines and methodologies for goal refinement makes it challenging for beginners to construct appropriate models, often resulting in low-quality goal models. This paper focuses on iStar, one of the GORA techniques and proposes a novel approach using GPT-4o to automatically generate iStar models of higher quality from simple problem descriptions written in natural language. Our technique contains two processes. In the first process, we provide GPT-4o with a step-by-step generation procedure and syntax rules for iStar models in order to minimize syntactic violations of a generated iStar model. By using Self-Instruct, without providing a large amount of training data for GPT-4o, our second process detects and corrects semantically low-quality model elements such as ambiguity, omissions and lack of consistency within an iStar model. Our experimental results demonstrated that 76.7% of the iStar models generated in the first process did not include any syntax errors and that in the second process a detection rate of 78.6% and correction rate of 100% were achieved.

Modern software development paradigms, particularly Agile and DevOps methodologies, have compressed development timelines while simultaneously increasing the complexity of security threat landscapes. Traditional threat modeling approaches, notably attack scenario analysis, demand substantial security expertise and extensive time investments that are incompatible with accelerated development cycles. Furthermore, the expanding demand for secure systems has necessitated threat analysis implementation by practitioners lacking comprehensive security backgrounds. This research presents COTTAGE, an automated tool that generates Attack Defense Trees (ADTrees) by leveraging CAPEC and CWE security knowledge repositories, specifically designed to enable effective threat modeling by security non-specialists. Experimental validation involving six participants with limited security expertise revealed that COTTAGE facilitated threat analysis outcomes comparable to expert-level assessment within 30-minute sessions, contrasting with the approximately two-day timeframe typically required by security professionals. Additional validation through DevOps environment case studies confirmed COTTAGE’s capability to support continuous security assessment via automatically generated reference tree structures.

Past methods for generating personas have primarily relied on manual assumptions, which have made the automated identification of system-level pain points challenging and have often limited the scope to actual inquirers who provided feedback. To address these limitations, we introduce a framework that combines user logs and inquiry data to generate personas in a more automated manner by integrating behavioral patterns and pain points. The framework consists of three phases: clustering users with k-means, analyzing those clusters with PrefixSpan, random forests, and BiLSTM models to identify pain points for both actual and potential inquirers (that is, actual inquirers—users who submitted feedback—and potential inquirers—users who did not reach out to developers regarding issues), and developing personas through large language models. A case study conducted on Japanese B2B SaaS software, involving 574 users and 3.5 million logs, showed that the generated personas were typically easy to conceptualize, and the identified pain points were compelling and actionable, underscoring their significance in guiding software enhancements.

The rapid iteration and frequent updates of modern video games pose significant challenges to the efficiency and specificity of testing. Although automated playtesting methods based on Large Language Models (LLMs) have shown promise, they often lack structured knowledge accumulation mechanisms, making it difficult to conduct precise and efficient testing tailored for incremental game updates. To address this challenge, this paper proposes a KLPEG framework. The framework constructs and maintains a Knowledge Graph(KG) to systematically model game elements, task dependencies, and causal relationships, enabling knowledge accumulation and reuse across versions. Building on this foundation, the framework utilizes LLMs to parse natural language update logs, identify the scope of impact through multi-hop reasoning on the KG, enabling the generation of update-tailored test cases. Experiments in two representative game environments, Overcooked and Minecraft, demonstrate that KLPEG can more accurately locate functionalities affected by updates and complete tests in fewer steps, significantly improving both playtesting effectiveness and efficiency.

Mobile devices are used for various applications. Although graphical user interfaces (GUIs) are commonly employed to operate applications, situations exist in which GUIs cannot be used effectively. Mobile devices also support other types of user interfaces (UIs), including voice UIs and gesture-based UIs, such as shake interfaces. Placeona-a model that identifies the status of the user’s eyes, hands, ears, and mouth when interacting with an application-was proposed to better adapt UIs to the user context. Based on placeona, we propose a method to develop adaptive user interfaces (PlaceonaUI) that detect a user’s current context and dynamically switch to the most appropriate UI. Our method defines specific contexts using the status of the user’s sensory and motor capabilities and establishes rules for UI switching. Furthermore, we developed a support system for creating the PlaceonaUI. The proposed PlaceonaUI and support system were implemented and evaluated to confirm their effectiveness.

Recent advances in Large Language Models (LLMs) have expanded their applicability beyond traditional language processing, creating growing demand for developing LLM-based agents capable of interacting with physical devices in IoT environments. These agents aim to deliver intelligent, context-aware services by leveraging LLMs’ reasoning capabilities and real-world data from diverse IoT sensors and actuators. However, building such agents in practice faces two major challenges. First, the inherent heterogeneity of IoT systems—including differences in device interfaces, data formats, and protocols—makes it difficult for LLM agents to produce accurate and contextually relevant responses. Second, developers are burdened with implementing low-level device logic and integration for each individual service, which hinders rapid development and reuse. To address these issues, we propose an interoperable IoT service platform that facilitates the development of LLM agents. Rather than focusing on agent construction itself, our approach provides a foundational infrastructure that enables seamless integration of LLMs with heterogeneous IoT environments. The platform standardizes access to sensors and actuators through unified interfaces, abstracts device-level differences, and provides a mechanism to easily convert structured sensor and actuator data into natural language representations interpretable by LLMs. Both atomic and composite services can be registered as callable tools, allowing agents to retrieve contextual information and issue control commands via a consistent interface. We implemented the platform in a smart home setting and conducted three evaluations: (1) a scenario-based demonstration of agent behavior, (2) a quantitative analysis of service selection accuracy across multiple LLMs, and (3) a simulation of development effort based on code size. Results show that the platform improves response accuracy, reduces service selection errors—especially for ambiguous requests—and significantly lowers development cost through component reuse. These findings confirm the platform’s effectiveness in bridging language-based reasoning with real-world IoT interaction.

Controller synthesis is a formal method approach for automatically generating Labeled Transition System (LTS) controllers that satisfy specified properties. The efficiency of the synthesis process, however, is critically dependent on exploration policies. These policies often rely on fixed rules or strategies learned through reinforcement learning (RL) that consider only a limited set of current features. To address this limitation, this paper introduces GCRL, an approach that enhances RL-based methods by integrating Graph Neural Networks (GNNs). GCRL encodes the history of LTS exploration into a graph structure, allowing it to capture a broader, non-current-based context. In a comparative experiment against state-of-the-art methods, GCRL exhibited superior learning efficiency and generalization across four out of five benchmark domains, except one particular domain characterized by high symmetry and strictly local interactions.

Multimodal Named Entity Recognition (MNER) aims to improve entity prediction by fusing textual and visual information. Most current MNER methods face two main issues: 1) they do not adequately appreciate the necessity of incorporating external knowledge into the model and the presence of knowledge redundancy; 2) during cross-modal fusion, they fail to achieve text-guided multimodal integration, resulting in an excessive introduction of image noise. To address these issues, we propose a novel framework PKTF, which primarily consists of two stages: the prior assisted knowledge generation stage and the multimodal named entity recognition stage. In the first stage, we use InternVL2-8B to generate prior assisted knowledge, aiming to provide additional contextual information for the original text. In the second stage, we design a Text-Max-Directed Fusion Module (TMDF). Specifically, we use gates to modulate the max pooling attention scores guided by the text, in order to obtain text-guided saliency attention scores. We can use this score to maximize the extraction of information favorable to MNER from the image features, while ensuring the dominance of the text. Experimental results show that our method is competitive compared to existing models, and it achieves F1-scores of 75.43% on the Twitter-2015 dataset and 88.74% on the Twitter-2017 dataset, respectively.

We introduce a memory retention mechanism based on semantic compression for LLM-driven test generation, achieving higher consistency and coverage by leveraging past information. We confirmed an average compression ratio of 16%, effectively utilizing both code and execution information.

Code readability is an important aspect of software quality, as it can significantly impact maintenance efforts. LLMs have been used to evaluate code readability. However, developers use different readability criteria for varying skill levels, necessitating the personalization of LLM-based evaluations. This study proposes two methods for calibrating readability evaluations using collaborative filtering and bandit algorithms (BAs). The experimental results demonstrate the need for personalizing LLM-based evaluations. Our methods are effective for these tasks.

Background: A workforce shortage in information technology (IT) has become a pressing issue. To increase the number of women IT professionals, it is crucial to expand the available career options for women students. A previous study demonstrated that gender stereotype is a threat to the performance of women in mathematics tests. This phenomenon is referred to as stereotype threat. Aim: We focused on gender stereotype threat to support women IT students in achieving their best performance. Method: Through a subjective experiment, we examined the relationship between a stereotypical cue and performance in computer programming. Additionally, we investigated the impact of a stereotype threat-removal strategy on programming performance. Result: When analyzing the time required to develop a program, neither the cue nor the strategy showed any significant effect. This result might be influenced by the intrinsic nature of computer programming and the effects of gamification.

Language Models for Code demonstrate significant potential in software engineering, and augmenting language models with various software components has garnered attention. However, the field lacks a comprehensive survey that systematically examines these approaches across different dimensions of code intelligence enhancement. This survey fills this critical gap by providing a comprehensive exploration of augmentation strategies across diverse software engineering tasks. We systematically categorize and analyze a broad spectrum of techniques, encompassing: (1) retrieval-augmented approaches that leverage external code repositories and documentation to enhance contextual understanding, (2) structural knowledge integration methods that incorporate code syntax and semantic patterns into model reasoning, and (3) external tool integration strategies that enable interaction with Application Programming Interfaces (APIs) and development environments. Furthermore, we examine the emerging agent paradigms that transform static code language models into autonomous systems capable of complex software engineering workflows. To our knowledge, this survey establishes the first comprehensive foundation for providing the current landscape of different augmentation techniques for code, providing key research directions for advancing code intelligence.

The multiple configuration and multiple data-based (MCMD-based) coarse-grained reconfigurable architectures (CGRAs) are employed as popular domain-specific accelerators due to their efficiency and flexibility. However, the existing MCMD-based CGRAs suffer from low hardware utilization and processing performance of dependent tasks due to the limitations of the memory architecture. In this paper, a parallel memory architecture (denoted as PMA) for MCMD-based CGRAs is introduced, which exploits the three-dimensional adaptive memory partitioning method and task-driven scheduling scheme to improve access performance and reduce hardware overhead of CGRAs. The memory access characteristics are analyzed to determine the memory requirements of MCMD-based CGRAs. Then, a three-dimensional memory partitioning method is proposed, which combines dynamic input/output memory partitioning and static shared memory partitioning to improve memory access performance, storage utilization, and throughput of CGRAs. Moreover, a task-driven scheduling scheme is proposed using priority-based task assignment, timeslot matching, and release methods to schedule the data with efficient memory access and non-conflict pipelining. Finally, an overall architecture of PMA is introduced, and the PMA is designed to validate the efficacy of the proposed approach. Experimental results show that PMA reduces 88.26% access time, lowers 40.3% area overhead, and decreases 48.4% power consumption compared to the existing solutions.

This paper proposes a method for utilizing large language model (LLM) inference results in the state space of deep reinforcement learning (DRL)-based GUI black-box testing for web applications. Previous research on mobile applications has shown that using HTML element interaction information in training improves the exploration performance of DRL agents. However, in web applications, obtaining interaction information can be difficult depending on the browser. Therefore, this study proposes a method that uses an LLM to infer clickability of elements, that is, one example of interaction information, from HTML and utilizes these results as part of the DRL state. Through experiments on a simple web application, we confirmed that using LLM-inferred values as state information improved learning efficiency and accuracy. Similar results were also demonstrated in experiments targeting a complex web application. The findings of this research enable browser-independent automation of GUI black-box testing and present a novel approach to utilizing LLMs for obtaining DRL states.

This paper describes a highly efficient method for the subgraph search problem, a fundamental issue in database query processing. The proposed CodeTree_sub extends the existing CodeTree method from a supergraph search technique to a subgraph search method. Importantly, CodeTree_sub is an index-based approach, despite assertions that the use of indices for the subgraph search problem is inherently inefficient. The proposed method has three advantages over existing techniques. First, unlike existing index-based methods, which extract numerous patterns through enumeration or subgraph mining, CodeTree_sub uses relatively few patterns extracted at random from all available patterns. This significantly reduces the time required to construct the index. Second, rather than simple path and cycle patterns, CodeTree_sub uses the more complex induced subgraphs as patterns, which enables excellent filtering performance from few patterns. Third, by searching for solutions during the filtering step, the need to verify subgraph isomorphisms is removed. We compare the subgraph search performance of CodeTree_sub with existing index- and vertex connectivity approaches, including the state-of-the-art CFQL and VEQ_S, on seven datasets and ten query tasks. The experimental evaluations demonstrate that CodeTree_sub is at least as fast as CFQL and VEQ_S in terms of query processing, filtering, and verification, and can be orders of magnitude faster depending on the dataset and search task. Overall, the results from this study highlight the potential of index-based methods to achieve efficient subgraph search outcomes.

In multi-issue negotiation problems within multi-agent systems, it is essential to preserve the privacy of each agent’s utility information. Fujita, Ito, and Klein proposed a negotiation protocol based on local search, in which the utility values for each intermediate candidate agreement are concealed using an additive secret sharing scheme. While their protocol aims to protect privacy, the sequence of candidate agreements and their transitions are explicitly revealed during the negotiation process, leaving room for inferring partial utility information. To overcome this limitation, the authors previously extended the protocol to realize fully private hill-climbing-based negotiation under multi-party secure computation (MPC), concealing all intermediate information and revealing only the final agreement. This earlier work was presented at WICT-DM 2024. Building on that foundation, the present study focuses on evaluating the scalability of privacy-preserving negotiation in multi-agent settings, and introduces an extension to support Simulated Annealing (SA) as a representative heuristic method. The proposed protocol securely executes all computations—including candidate generation, utility evaluation, and acceptance decisions—under MPC without revealing any intermediate candidate agreements or comparison results. We evaluate its scalability in negotiations with 2 to 17 agents using a personal computing environment (Apple M3 Pro laptop). The results show that the protocol successfully completes negotiation with 17 agents in approximately 10,659 seconds while maintaining agreement quality. A timer-based breakdown further reveals that random number generation and array updates dominate the execution cost, accounting for more than 70% in large-agent settings, whereas utility computation and acceptance decisions contribute only marginally. These findings demonstrate that MPC can be applied to large-scale multi-agent negotiations with acceptable overhead, and also indicate that future performance improvements should focus on optimizing the identified bottlenecks.

This work presents a 4-row cyclic Enhanced Arbiter PUF (EA-PUF) architecture designed to mitigate spatial delay bias while maintaining high reliability and resistance to modeling attacks. The proposed topology employs a rotation-symmetric four-lane switching network with antipodal pairing, enabling intrinsic gradient cancellation within a single arbiter race. We provide rigorous theoretical foundations through formal bias-cancellation theorems, proving that the 2-row direct-comparison topology cannot cancel gradient bias, while the 4-row design with mirrored placement achieves exact cancellation. A uniqueness proof demonstrates that the antipodal pairing (lanes 0,2 versus 1,3) is the only configuration achieving bias cancellation under mirror-symmetric placement. Hardware implementation on twelve Altera Field Programmable Gate Arrays (FPGAs), spanning heterogeneous and homogeneous device families, demonstrates strong physical symmetry. Timing analyzer of post-synthesis netlists reveals maximum lane delay skew reduced to 37 ps compared to 135 ps in a conventional 2-row Arbiter PUF. The design occupies only 76 LUTs, achieves f_max of 199.12 MHz, and consumes 0.17 mW total thermal power, showing minimal area and power overhead. Across all boards, we measured near-ideal uniqueness (35.58%-57.67%), reliability up to 96.49%, uniformity centered around 53.08%, and balanced bit aliasing (53.09%). Comprehensive entropy analysis yielded Shannon entropy median of 0.992 bits/bit and Min-entropy median of 0.862 bits/bit. Authentication metrics achieved low False Acceptance Rates (1.28%-2.86%) and False Rejection Rates (0.72%-2.37%). Extensive machine learning evaluation using Artificial Neural Networks (ANNs) with depths ranging from 3 to 10 layers demonstrated fundamental modeling resistance, with prediction accuracies plateauing near 54%-57%. These results validate the EA-PUF as a bias-resilient, resource-efficient, and secure candidate for deployment in constrained authentication environments.

Forward-secure digital signatures protect the integrity of past signatures, even if the current signing key is compromised. Among forward-secure signature schemes, the method introduced by Lee et al. [1], based on zero-knowledge succinct non-interactive arguments of knowledge (zk-SNARKs), is particularly notable for achieving constant complexity across all metrics without requiring a predefined maximum time period. However, a naive approach to recursive proof composition results in an excessive amount of redundant computation being repeated for each signing process, which our method reduces significantly. In this paper, we advance a zk-SNARK-based forward-secure signature scheme by significantly improving the efficiency of its signing algorithm. By incorporating commit-and-prove SNARKs, we replace the inner verification process with commit verification within the signing circuit. Furthermore, we employ efficient recursive zk-SNARKs with accumulation and folding schemes to improve the setup and update algorithms. Our implementation demonstrates the practicality of our approach: the signing procedure completes in 0.18 seconds, achieving a 75-fold speedup over the previous scheme, setup time is reduced to 0.71 seconds - over 61 times faster, and public parameters are reduced to 25 MB, more than 16 times smaller.

Causal discovery aims to infer cause-and-effect relationships from observational data, a crucial step beyond statistical correlation. A prominent method for this is the Linear Non-Gaussian Acyclic Model (LiNGAM), which can uniquely identify the causal structure by assuming linear relationships and non-Gaussian noise. LiNGAM-based algorithms typically depend on two key components: a search algorithm to determine the causal ordering of variables, and an independence measure to guide the search. Recent work, LiNGAM-MMI, proposed that replacing the simple greedy search with a global, shortest-path search led to superior performance, particularly when unmeasured common causes (confounders) are present. However, the claim was based on experiments that also modified the independence measure from the original baseline, making it difficult to isolate the source of the improvement. To address this, we perform extensive experiments to test whether the search algorithm was truly the driver of performance, hypothesizing that the choice of independence measure is the dominant factor. In particular, we introduce a unified beam search framework that serves as both an analytical tool to disentangle these components and a practical algorithm with a scalable performance-complexity trade-off. Our simulations comparing the kNN-based Copula Entropy (CopEnt) with the Pairwise Likelihood Ratio (PLR) establish that the independence measure is the dominant factor, to the extent that a simple greedy search with a more effective measure, PLR, outperforms a global search with a less effective one, CopEnt, on general graphs. Furthermore, we find no evidence that the benefit of a more complex search algorithm is specific to handling unmeasured confounders, suggesting it instead serves to overcome general estimation errors arising from finite data. Finally, we demonstrate that the strong performance with CopEnt reported in the previous work was an artifact of a simplistic experimental setup, as its performance advantage is reversed on more realistic and complex structures, including Erdős-Rényi (ER) and Scale-Free (SF) networks.

Reservoir water levels exhibit significant irregular fluctuations due to climate change and intensified water resource exploitation. To improve forecasting accuracy, this study introduces SARIMA-Embedded GRU (SE-GRU), a novel hybrid model that directly embeds statistical time-series components-seasonal buffers, trend differencing gates, and ARMA structures into GRU recurrent units. Unlike conventional two-stage hybrid methods, SE-GRU simultaneously captures linear statistical patterns and nonlinear dependencies through end-to-end optimization, enhancing both accuracy and interpretability. Evaluated on An Khe reservoir data, the proposed model achieves superior performance (MAE = 0.001, MSE = 0.014, RMSE = 0.041), outperforming standalone SARIMA, GRU, and other baseline models, demonstrating the effectiveness of integrated architectural design for reservoir water level forecasting.

Existing parking space line detection systems face persistent challenges under complex environmental conditions, particularly low-light conditions, occlusions, and textured background interference, which lead to compromised accuracy, excessive computational overhead and high latency. To address these limitations, this study proposes PSL-YOLO11, an enhanced lightweight detection framework based on the YOLO11 architecture. The innovation lies in a hybrid dynamic attention mechanism that synergistically combines the Convolutional Block Attention Module (CBAM) and the Global Attention Mechanism (GAM) for superior environmental adaptability. The framework implements multi-scale feature fusion with channel expansion for discriminative feature learning, and adopts GhostNet-based feature compression to reduce computational complexity. Additionally, a key point-driven geometric abstraction algorithm is designed to convert parking lines into quantifiable topological representations. Experimental results on the HNUST-IR dataset demonstrate that PSL-YOLO11 achieves 93.8% mAP and 96.15 FPS, with 22% fewer parameters than YOLO11 while maintaining real-time performance.

Disfluent utterances in spontaneous speech, such as fillers and hesitations, cause recognition errors during automatic speech recognition (ASR). To address this problem, we propose a method called “disfluency labeling”, which replaces disfluent segments in transcription data with one of two labels: # (filler) or @ (hesitation). End-to-end training of the ASR model with such labeled data enables recognition of these disfluent segments as targets, like characters, allowing the extraction of what the speaker intended to say. In evaluation experiments, the proposed disfluency labeling method achieved higher recognition accuracy than the previous proposed ASR methods treating disfluencies, suggesting that explicit learning of disfluency features as labels is effective for improving spontaneous speech recognition.

Point cloud denoising is an essential preprocessing step in 3D vision tasks such as autonomous driving, robotic perception, and 3D reconstruction. However, uneven point density, complex geometry, and uncontrolled noise make it difficult for existing methods to balance accuracy and efficiency. To address this challenge, we propose a lightweight multi-scale point cloud denoising network (LPCDN) that enhances inference efficiency while maintaining high denoising quality. LPCDN adopts complementary large- and small-scale branches: large-scale modeling improves robustness to density variations, whereas small-scale modeling preserves fine-grained geometric structures. Random sampling is applied in dense local regions to reduce computation, while representative sampling is employed for large-scale structures to retain global consistency. A multi-scale feature fusion mechanism with channel attention further strengthens adaptability and representation capability. Extensive experiments on both synthetic and real-world datasets show that LPCDN achieves competitive accuracy compared with state-of-the-art methods, while significantly accelerating inference. In addition, the integration of Automatic Mixed Precision (AMP) in training and inference reduces runtime by about 30% without accuracy loss, confirming the effectiveness and deployability of the proposed approach.

This paper proposes a method for correcting character-recognition errors in Japanese handwritten answers to writing-type questions from exercise books for automated scoring. We created a model to correct character-recognition errors by fine-tuning the Text-to-Text Transfer Transformer using pairs of automatically recognized data from handwritten answers and their manual corrections. The data comprised handwritten Japanese answers from 185 junior high school students to writing-type questions in a Japanese language task. In addition, we augmented the training data using the five best results of the character-recognition model with confidence scores to learn additional patterns of recognition errors. The experimental results revealed that the answers corrected by the proposed method were closer to the actual answers than those before the correction and data augmentation was effective for the correction model. We further conducted an empirical study on how error correction in Japanese handwritten answer recognition affects automated grading accuracy of a zero-shot automated grading system using GPT-4o and GPT-4o-mini. Experimental results showed that error correction improved grading accuracy by up to 13 percentage points. Moreover, a detailed analysis of grading cases elucidated how error correction influences grading outcomes. These findings underscore the potential of error correction methods to enhance the accuracy of automated grading for descriptive answers.

With the expanding constellation scale and growing demands for Quality of Service (QoS) assurance, satellite networks are evolving toward multi-layer architectures. Traditional single-layer designs fail to ensure link stability and service continuity for emerging applications. This paper proposes a cross-layer communication design for large-scale Low Earth Orbit (LEO) satellite networks oriented toward QoS optimization. Specifically, we establish a dynamic topology model for large-scale LEO satellite networks and define the multitask scenario. The routing problem is formulated as an Integer Linear Program (ILP) to maximum the average link utilization ratio under unsplittable flow constraints. We propose the Deep Deterministic Policy Gradient-Based QoS Routing Algorithm (DDPG-QoSRA) to ensure both path reliability and the transmission performance of critical services. Finally, the simulations show the effectiveness and superiority of the proposed algorithm.

With the commercialization of 5G, the era of hyper-connectivity is accelerating its evolution. The cloud-edge collaborative networks (CECN) has become a key infrastructure supporting the intelligent society. Its open and dynamic nature has also introduced continuously evolving security threats. Traditional static defense has limitations in responding to dynamic attacks and achieving cross-domain resource collaboration. We propose a deception defense method based on two-layer game and multi-agent reinforcement learning. We analyze the network state evolution process. We construct a two-layer architecture that integrates cooperative game and Stackelberg game, and we design a strategy solving method based on multi-agent reinforcement learning (MARL) to achieve adaptive optimization of defense strategy. The simulation results show that the proposed method can effectively improve the overall defense effectiveness and maintain fairness within the defense alliance.