Abstract:Mainstream large language models adopt an autoregressive generation paradigm, requiring text to be generated token by token, which makes it difficult to meet the demands of efficient inference in practical applications. Diffusion language models achieve parallel decoding through iterative denoising and therefore offer certain theoretical advantages in inference efficiency, making them an important technical direction for improving the inference efficiency of large language models. However, the generation quality of existing diffusion language models still leaves considerable room for improvement. Contrastive decoding is a widely proven effective method for enhancing the generation quality of language models, whose core idea is to calibrate a stronger expert distribution using a weaker amateur distribution. However, existing contrastive decoding methods are primarily designed for autoregressive language models and typically require an additional independent model to provide the amateur distribution, which significantly increases inference overhead and undermines the core theoretical efficiency advantage of diffusion language models. To address this issue, this paper proposes DiffuCD, a contrastive decoding approach designed specifically for diffusion language models. The approach is based on a key observation: during the denoising process of diffusion language models, predictions at earlier timesteps are of lower quality due to limited contextual information, while prediction quality progressively improves as denoising proceeds. This property naturally satisfies the requirement of contrastive decoding for distributions of varying quality. Based on this observation, DiffuCD directly uses the predicted distributions from historical timesteps as amateur distributions and treats the predicted distribution at the current timestep as the expert distribution, thereby calibrating the current distribution. Extensive experiments are conducted on 4 diffusion language models and 6 benchmarks. Experimental results demonstrate that DiffuCD can consistently improve generation quality across various models and tasks, while maintaining nearly identical throughput and memory usage compared to the baseline without contrastive decoding, effectively preserving the theoretical inference efficiency advantage of diffusion language models.

Abstract:Aerospace control systems are characterized by high architectural complexity and tight software-hardware coupling. Their requirements are typically specified in natural language, which inherently suffers from ambiguity and is prone to errors when referencing low-level hardware details. These deficiencies often lead to inconsistencies during requirements propagation and hinder the reuse of software assets across different system models. To address these issues, this study proposes SpaceRDL, a system-level requirements description language specifically designed for aerospace control systems. Its core contribution lies in the introduction of “system resources” as a unified abstraction for entities in the system’s interaction environment and physical devices, enabling engineers to define operations on system resources in a structured and precise manner, thus eliminating ambiguity in the requirements description language while avoiding excessive detail. The language incorporates a system resource repository and a set of atomic operation instructions with well-defined semantics, enabling automated mapping from task intent to software implementation. Furthermore, task requirements specified in SpaceRDL can also serve as reusable assets for cross-project migration. Case studies demonstrate that the proposed language effectively describes the requirements of aerospace control systems, providing practical and viable support for the development of next-generation aerospace software engineering methodologies.

Abstract:The digitalization of the power industry gives rise to massive volumes of structured business data, creating an increasingly urgent demand for efficient natural language-driven data querying in critical business operations. However, power databases typically feature large schemas, ambiguous field semantics, and complex queries, and current query methods therefore continue to rely heavily on manually written SQL. This not only limits efficiency but may also lead to operational or dispatching decision errors. Although large language models (LLMs) and Text-to-SQL techniques provide a new solution to this problem, their applicability remains limited due to the characteristics of power-domain databases. To address these challenges, this study proposes ElectricSQL, a model-agnostic Text-to-SQL framework. The proposed framework improves SQL generation performance for complex power business queries through LLM-based imitation learning and SQL view-based schema transformation, without relying on any specific commercial large language model. In addition, PowerSQL, a Text-to-SQL benchmark dataset for real power business scenarios, is developed. Experimental results show that ElectricSQL significantly outperforms mainstream baselines on the PowerSQL dataset and achieves stable performance gains across multiple types of open-source LLMs. These results verify the effectiveness of the proposed method and provide a reusable solution for building natural language query systems in the power domain and for research on domain-specific Text-to-SQL.

Abstract:With the widespread application of deep learning models in embedded and edge devices, traditional model evaluation methods that rely on the full test set struggle to meet the requirements of low-latency and highly efficient deployment because of their high computational overhead and resource consumption. Existing dataset selection methods often fail to balance evaluation accuracy with high compression ratios, resulting in significant deviations in performance estimation. To address this issue, this study proposes a prediction-uncertainty-aware test set selection (PUA-TSS) method. The proposed method aims to achieve efficient and accurate evaluation of model performance by constructing a highly informative proxy test set. The method first quantifies sample uncertainty using the information entropy of model prediction outputs and ranks the test samples in a task-driven manner. On this basis, an equal-interval uniform sampling strategy is adopted to preserve key distributional characteristics. In particular, it alleviates the loss of rare classes in long-tailed scenarios, thus maintaining evaluation sensitivity and consistency at low sampling rates. Experiments are conducted on benchmark datasets using multiple model architectures across various tasks, including image classification, object detection, and segmentation. The results show that when only 5% of the test samples are retained, the mean absolute deviations between the proxy test set evaluation results and the full test set evaluation results remain within 0.71 percentage points for classification, 2.90 percentage points for detection, and 0.40 percentage points for segmentation. The proposed method significantly outperforms random sampling and existing dataset compression methods. Furthermore, when quantized models with different bit widths are evaluated, the performance evaluation error consistently remains within 5 percentage points, without requiring the proxy set to be reconstructed or full evaluation to be performed repeatedly for each model variant. The results demonstrate that the proposed method substantially reduces the computational cost of model evaluation while providing high accuracy, strong robustness, and good generalization capability, thus offering an effective solution for model evaluation in resource-constrained environments.

Abstract:Spectre V1 exploits the speculative execution mechanism of processors to steal sensitive data, making detection in closed-source binary programs particularly challenging. Static analysis tends to produce high false-positive rates, while dynamic fuzz testing suffers from limited coverage and noisy reports. To address these issues, this study proposes ConSpec, a binary Spectre V1 detection method. The proposed method first employs a lightweight BERT model to perform semantic modeling of disassembled instruction sequences. By incorporating the assembly-level semantic patterns of Spectre V1, it generates function-level confidence scores to quantify vulnerability suspicion. Based on the confidence scores, functions are then divided into high-, medium-, and low-risk regions. Through whitelist-driven directed instrumentation and tiered configuration, coverage-guided fuzz testing is used to repeatedly execute the instrumented program and collect vulnerability evidence, thus concentrating detection resources on high-risk regions while controlling the overall overhead. To address noisy dynamic validation reports and missed detections caused by limited coverage, a window-level secondary confirmation mechanism is introduced. By matching specific semantic templates and performing local register tracing, this mechanism verifies the key vulnerability-related evidence chain, thus filtering out false positives and compensating for missed detections. Experimental results show that the proposed method fully covers 15 known Spectre V1 exploitation patterns. Tests on five real-world programs, including jsmn, brotli, and OpenSSL, show that, compared with Teapot, ConSpec improves Precision by an average of approximately 24.3 percentage points and the F₂-score by an average of approximately 24.4 percentage points, while reducing the relative false-negative rate overall. The average runtime decreases from 1159.7 min to 774.3 min, a reduction of 33.2%.

Abstract:The openness and fragmentation of the Android ecosystem expose it to continuously evolving and increasingly complex security threats. An increasing number of malicious attackers use runtime analysis techniques to probe application logic and tamper with execution flows, thus bypassing Android security measures and increasing the persistence and impact of malicious activities. Anti-runtime analysis technology detects runtime analysis activities, enabling users and developers to identify such activities promptly and take protective measures, thus safeguarding application integrity and data security. Currently, various anti-runtime analysis methods based on static databases have been proposed. These methods collect threat intelligence on the characteristics of known runtime analysis techniques and match information extracted from devices against this intelligence to detect traces of runtime analysis techniques. However, these methods suffer from “knowledge lag” and provide insufficient analytical depth and result interpretability, making them increasingly ineffective against rapidly evolving runtime analysis techniques. To address these issues, this study proposes and implements AIntel-Agent, a threat intelligence-driven multi-agent Android anti-runtime analysis method. The proposed method uses agent technology to build a dynamic knowledge base. Device fingerprints are then analyzed in depth based on the characteristics of five categories of runtime analysis techniques, including rooting, hooking, debugging, virtual environments, and tampering, thus enabling Android anti-runtime analysis tasks to be completed in an automated and intelligent manner. Through the design of an agent collaboration framework, AIntel-Agent proposes a method for constructing a dynamic knowledge base by obtaining threat intelligence from the Internet in real time and develops an iterative reasoning mechanism that associates device fingerprints with the characteristics of runtime analysis techniques to generate anti-runtime analysis reports for devices. Experimental results based on a dataset of real devices show that AIntel-Agent achieves a precision of 88.5% and a recall of 93.9% on Android devices. Its risk reports contain sufficient evidence and are interpretable, demonstrating its effectiveness in Android anti-runtime analysis.

Abstract:Knowledge graph completion is a crucial approach for updating knowledge graphs. However, knowledge graphs typically contain many complex relation types, which cause the same entity to assume different roles across triples. Moreover, entities inherently possess diverse attributes and semantic features, and different triples may emphasize different semantic aspects of an entity. To address these challenges, this study proposes RinKGC, a role interaction-aware knowledge graph completion method, consisting of three key modules: the role cue generation module, the role interaction mining module, and the role semantic enhancement module. Specifically, the role cue generation module accurately identifies the roles of head entities, relations, and tail entities within triples, enhances interactions between entities and relations, and propagates semantic information across layers via residual connections. The role interaction mining module partitions entity and relation feature dimensions into multiple subspaces and models their interactions within each subspace, enabling efficient perception of diverse semantic features. The role semantic enhancement module leverages complementary information to achieve accurate prediction of target entities. In addition, a Transformer-based backbone is used to reveal long-range semantic dependencies across triples. Extensive experiments on multiple publicly available knowledge graph completion datasets demonstrate that RinKGC achieves superior performance.

Abstract:Geological disasters pose serious threats to people’s lives, property, and productive activities. With advances in data acquisition and processing technologies, data-driven approaches to geological disaster prevention and control have become a research hotspot. Following the “data-governance-analysis-decision” paradigm, this study systematically proposes a comprehensive framework for data-driven geological disaster prevention and control technologies. At the data foundation level, a triplet framework encompassing research objects, temporal scales, and data modalities is proposed to support the systematic integration of multi-source heterogeneous data. At the data governance level, strategies are systematically summarized to address quality issues including data inconsistency, low resolution, missing and imbalanced data, and uncertainty. At the data analysis level, geological disaster analysis tasks are divided into four categories, including regional risk assessment, image-based disaster identification, surface displacement prediction, and disaster occurrence probability prediction. These tasks are centered around two core questions: “where disasters occur” and “when they occur”. Five key characteristics that data-driven methods urgently need are identified: automation, universality, robustness, explainability, and lightweight design. At the intelligent decision-making level, a three-stage decision framework spanning pre-disaster, during-disaster, and post-disaster phases is constructed, systematically analyzing the core decision-making issues at each stage and clarifying the complete data-driven decision pathway from risk assessment to emergency response and further to recovery and reconstruction. In addition, the study also provides perspectives on future decision-making models based on multi-agent collaboration. Through this review, systematic references are expected to be provided for research and practice in the field of data-driven geological disaster prevention and control, promoting the transformation and development of disaster prevention and mitigation toward digitalization and intelligence.

Abstract:As large language model (LLM)-based intelligent programming assistants gradually reshape software engineering paradigms, AI-generated code has become a critical digital asset, while the resulting ambiguity in intellectual property ownership and risks to software supply chain security are becoming increasingly severe. Code watermarking, serving as a “digital anchor” for copyright ownership confirmation and provenance tracking, faces unique difficulties due to the strict syntactic constraints and low information entropy of programming languages. To address problems in current research, such as fragmented technical approaches and incomplete evaluation systems, this study provides a systematic classification of existing research and develops a theoretical model from a full-lifecycle perspective. First, a novel lifecycle-based classification framework is proposed that categorizes existing techniques into three paradigms: backdoor triggering during training, sampling-bias embedding during inference, and semantic-rephrasing embedding after generation. Their operating mechanisms and evolutionary paths are then analyzed in depth. Second, the constraining relationships among key performance metrics of code watermarking are analyzed in depth, revealing the inherent contradictions and trade-off logic among fidelity, robustness, and watermark capacity under current technological conditions. Furthermore, an adversarial game perspective is introduced to analyze attack methods involving destruction, deception, and detection, and evaluation benchmarks ranging from functional consistency to adversarial traceability capabilities are reviewed. Finally, in view of the limitations of existing techniques in engineering usability and full-lifecycle evaluation, future research directions are outlined, including dynamic protection and ecosystem integration across the entire DevOps pipeline, aiming to promote the transition of relevant technologies from the proof-of-concept stage to industrial-grade application scenarios.

Abstract:With the development of autonomous agent systems, how to maintain structural consistency, support long-range dependency reasoning, and organize multi-stage task execution in complex environments has become an important issue in agent research. However, when handling complex tasks involving long-range dependencies and cross-stage goal coordination, existing agent paradigms still face problems such as difficulty in maintaining global consistency and limited efficiency in strategy organization. To enhance the structural modeling and organizational capabilities of agents, graph structures have been introduced into agent systems, providing structured support for environment representation, task scheduling, and cognitive maintenance. Based on these observations, graph agents are proposed for the first time as a unified analytical framework for agent systems based on the functional roles and embedding positions of graphs. This concept refers to an agent paradigm with graph structures at its core, and related studies are systematically reviewed accordingly. This study focuses on analyzing how graphs, as structural carriers, are embedded into the perception, reasoning, memory, and collaboration processes of agents. Within this analytical framework, this study categorize the functional roles played by graphs in existing studies into three types: graphs for environment modeling, graphs for dependency modeling, and graphs for thought modeling. These three roles respectively characterize the distinct mechanisms by which graph structures contribute to environment representation, task scheduling, and cognitive maintenance. In environment modeling, graphs serve as external structural representation tools to describe entity relations, state transitions, and action spaces, thus providing structured environmental descriptions and a constraint basis for policy generation. In dependency modeling, graphs are used to represent subtask dependencies, collaboration structures, and decision paths, supporting multi-agent coordination, task decomposition, and execution scheduling. In thought modeling, graphs are embedded into internal reasoning and memory-updating processes, forming continuously evolving internal cognitive structures. On this basis, this study further compare the three functional roles in terms of their structural effects, operational costs, and applicability boundaries, and summarize the evaluation metrics, task benchmarks, and baseline settings adopted in existing studies. Finally, the limitations and challenges facing graph agents in open and dynamic environments, complex execution processes, and long-term interaction scenarios are discussed, followed by an outlook on future research directions.