Abstract:With the rapid development of deep learning, research on deep learning frameworks and hardware has become a crucial direction for advancing the field. Frameworks provide developers with convenient tools for building deep learning models, while hardware delivers powerful computational capabilities. However, limited adaptability and insufficient compatibility between diverse deep learning frameworks and hardware platforms often lead to performance and scalability issues. To address this challenge, deep learning compilation technology has emerged. Models from different frameworks are efficiently mapped to executable code for specific backend devices through a series of intermediate representations and automated transformations. During this process, various deep learning compilation optimization techniques, such as operator fusion, memory optimization, and auto-tuning, are applied. These not only resolve scalability issues but also significantly improve the computational efficiency of the models. This study first reviews the basic concepts and overall process of deep learning compilation, then summarizes and categorizes the common optimization techniques in deep learning compilation, and finally discusses the challenges faced by the field and potential future development directions.

Abstract:Intrusion detection systems rely on prior knowledge, and their input data contain rich semantic information, making them well suited for large language models (LLMs). However, the computational overhead of LLMs and data privacy constraints make their direct deployment difficult. Meanwhile, although small models are convenient for deployment, their performance is limited. To address this issue, this study proposes an iterative model for intrusion detection based on data-driven distillation and performance motivation, termed as RADOM. RADOM performs knowledge distillation through data exchange between LLMs and small models, and improves the quality of LLM-generated data by using incorrect predictions from the small models as iterative feedback. At the same time, this study introduces a feature dimension optimization mechanism, in which LLMs are used to perform feature selection and iterative updating, thereby further improving the classification capability of small models. Experimental results on public datasets and self-collected datasets show that RADOM effectively improves the detection and classification performance of small models for attack behaviors, and the accuracy of small models increases from 67.30% to 96.19%, verifying the effectiveness of the proposed method.

Abstract:Cardinality estimation is a core component of the query optimizer in database management systems (DBMS), and its accuracy directly affects the quality of execution plans. Although existing learning-based cardinality estimation methods outperform traditional methods in certain scenarios, they struggle to maintain consistently high accuracy on highly heterogeneous query workloads, where queries may differ significantly in the number of tables, join patterns, and predicates. Moreover, existing multi-model approaches lack a fine-grained mechanism for selecting models at the query level. Therefore, this study proposes a query routing-based cardinality estimation method, QRCE, whose core ideas include 1) constructing a heterogeneous and scalable Seq2Seq model space, integrating multiple sequence modeling architectures with different structural inductive biases to form a candidate model set; 2) building a gated mixture query router, GMQR, to achieve fine-grained model routing based on query semantics and enable adaptive decision-making for individual queries; 3) proposing a Q-error-based (1+ε)-approximate label strategy, transforming model selection into a supervised learning task, enabling the system to adaptively select the most suitable model based on the semantic features of each query. Experimental results on three benchmark datasets, STATS, JOB-light, and TPC-H, show that QRCE achieves significantly better overall and tail error performance than most baseline methods. Particularly on the structurally complex STATS dataset, QRCE reduces the Q99 error by approximately 53%, effectively avoiding long-tail catastrophic bias. On JOB-light, its Q99 is close to that of the best single model, while maintaining advantages at Q50/Q90/Q95. On TPC-H, the Q99 error is further reduced by about 16%. QRCE demonstrates stronger robustness and adaptability in long-tail queries and complex join scenarios.

Abstract:As deep learning models are increasingly deployed in resource-constrained scenarios, model compression techniques such as pruning, quantization, and knowledge distillation have become key approaches for improving model efficiency and deployability. However, while model compression offers advantages in performance and efficiency, it is also intricately linked to security and privacy issues. On the one hand, compression may alter a model’s representation space or decision boundary, thus introducing new security risks or amplifying existing threats. On the other hand, security and privacy requirements have also shaped the development of model compression techniques, making security-aware compression an increasingly important research direction. This study systematically reviews recent progress in model compression research from a security perspective, focusing on two main aspects. First, it examines the security and privacy risks faced by compressed models, including backdoor attacks, adversarial examples, privacy leakage, and emerging risks in frontier application scenarios. Second, it surveys security-aware model compression methods for adversarial robustness and privacy protection, summarizing representative studies that incorporate security constraints and privacy-preserving mechanisms into the compression process. In addition, this study identifies key issues in security and privacy research on compressed models, provides risk-aware guidance for typical deployment scenarios, recommends benchmark-based evaluation dimensions for assessing the security of compressed models, and discusses future research directions in terms of underlying mechanism analysis and security protection in real-world deployments.

Abstract:Large language models (LLMs) perform well in complex reasoning tasks involving logical operations. However, the internal mechanisms underlying their logical reasoning capabilities remain insufficiently interpretable. Currently, it remains unclear whether neuron subsets closely associated with logical reasoning performance exist within LLMs, what forms of computational operations are performed by these neurons in the network, and how they influence model performance across different tasks. To investigate these key questions concerning the nature of LLM reasoning mechanisms, this study proposes an “identification-intervention-evaluation” framework aimed at systematically analyzing the neuronal operational basis and functional implementation of logical reasoning. A neuron localization method based on gradient importance is used to identify “logic neurons” that are highly sensitive to logical reasoning tasks, and targeted interventions are performed on their activation values to further evaluate their impact on model behavior. The results show that the identified logic-related neuron subsets are mainly concentrated in the middle layers of the network and predominantly located in feed-forward network (FFN) modules. Further analysis indicates that different logical tasks involve both shared neuron components and task-specific differences. Based on these findings, this study explores a lightweight targeted enhancement method for logical reasoning capabilities without modifying model weights. Experimental results show that this method improves the average accuracy of the model on long-chain reasoning tasks in LogicBench by 9.2% and improves performance on logic-sensitive general tasks by 4.57%.

Abstract:The dynamic single-source shortest paths problem is studied in the massively parallel computation (MPC) model, where the memory size of each machine is $ \mathrm{O}({n}^{\alpha }) $. Here, $ n $ is the number of vertices in the graph, and $ \alpha \in (0, 1) $ is a constant. This study primarily considers directed or undirected graphs with integer edge weights, where the ratio between the maximum and minimum edge weights is upper-bounded by $ \left\lceil poly(\log n)\right\rceil $. The graph update operations include single-edge insertion, deletion, and weight change. To address this problem, this study combines the polynomial matrix inverse method, the path decomposition strategy, and parallel linear algebra algorithms. In addition, by designing an efficient parallel polynomial multiplication algorithm, a randomized parallel dynamic single-source shortest paths algorithm is proposed. The algorithm achieves an update round complexity of $ \mathrm{O}({\alpha }^{-1}\log n) $ and a total memory requirement of $ \tilde{\mathrm{O}}({n}^{3-3(3-\omega )/(4-\alpha /2)-\alpha (\omega /2-1)}) $. Here, the update round complexity refers to the number of synchronous iteration rounds required by the algorithm to recompute, after each edge change in the graph, the distances from the source to other affected vertices. $ \tilde{\mathrm{O}}(\cdot ) $ suppresses polylogarithmic factors, and $ \omega $ represents the exponent of the arithmetic complexity of fast matrix multiplication on matrices of size $ \mathrm{O}({n}^{\alpha }) $. Compared with existing parallel static single-source shortest paths algorithms that require at least $ poly(\log n) $ round complexity, the proposed algorithm significantly reduces the round complexity.

Abstract:With the rapid development of diffusion models and generative adversarial networks, AI-generated images are increasingly used in content creation, while also posing new challenges to content trustworthiness and source attribution. Compared with merely distinguishing authentic images from fake ones, accurately identifying the source generative model of generated images has greater practical significance. For instance, this capability can provide important technical support for content source identification and trustworthy governance in scenarios such as generated content platform regulation, tracing of maliciously forged content, and responsibility tracking of generative models. During the synthesis process, different generative models often leave fine-grained high-frequency artifacts at the pixel level, while also producing model-specific differences in generative pattern consistency at the structural and global levels. This cross-scale artifact phenomenon indicates that existing methods relying solely on single-scale or single-semantic-level feature representations are insufficient for simultaneously capturing fine-grained artifacts and global generative style differences, thus limiting attribution performance in complex multi-model scenarios. To address these issues, this study proposes a collaborative pixel-block artifact modeling framework (CPB-AM) for AI-generated image model source attribution. The proposed method adopts a dual-branch architecture to jointly model generative artifacts at different levels. The pixel-level branch extracts local forgery traces through high-frequency enhancement operations and learns fine-grained artifact features associated with the generative domain. The block-level branch partitions the image into local blocks and models inter-block correlations through a frequency-aware attention aggregation mechanism. This attention mechanism adaptively aggregates low-frequency smooth components and explicitly separates structural consistency representations from residual components, enabling more effective characterization of differences among generative models in terms of spatial structure and overall style. Furthermore, for cross-scale fusion, a cross-branch attention-guided fusion module is designed as a guided feature fusion mechanism. By constructing modulation weights from pixel-level artifact features for block-level representations, the module enables fine-grained residual information to guide and constrain the structural modeling process, thus enhancing the complementarity of cross-scale artifact information. Experimental results on datasets containing real images and multiple generative models demonstrate that the proposed method achieves superior classification performance compared with existing approaches in closed-set generative model source attribution tasks. Under open-set experimental settings involving unseen generative models, the proposed method also exhibits strong robustness in both unknown model rejection and attribution discrimination. These results verify the effectiveness of the collaborative pixel-block artifact modeling strategy in improving generative model source attribution accuracy and adaptability to open scenarios.

Abstract:Geological disaster knowledge services play a critical role in emergency command, public safety, and science communication. However, existing large language model (LLM)-based question-answering systems still have limitations in multi-source heterogeneous knowledge fusion, complex task planning, and highly trustworthy reasoning. To address these issues, this study proposes a multi-agent collaborative reasoning framework for geological disaster knowledge services. Driven by explicit task states, the proposed framework decomposes complex queries into subtasks such as retrieval, analysis, generation, and verification through a role-based agent collaboration mechanism and ensures the traceability of the reasoning process and the controllability of results through task write-back and closed-loop consistency verification. At the data level, the framework supports unified semantic access to knowledge from multiple sources, such as popular science and domain-specific knowledge, laws and regulations, and disaster records. At the reasoning level, a hierarchical reasoning process based on collaborative execution and quality control is constructed. Experimental results demonstrate that the proposed framework outperforms the baseline method, which is based on a single model combined with external retrieval, in terms of domain knowledge coverage, answer consistency, and collaborative reasoning capability, and shows more stable performance in high-reliability tasks involving laws, regulations, and disaster statistics. This study provides a feasible systematic solution for the application of multi-agent LLMs in disaster emergency knowledge services.

Abstract:Double-block-length hash functions are a classical approach for amplifying the concrete security of hash functions. This construction has been proven to achieve optimal quantum collision resistance under certain conditions. However, whether double-block-length constructions can still achieve optimal concrete security in the stronger and more applicable quantum security model of the collapsing property remains an open question. To conduct a further study on this issue, this study considers the collapsing property, which extends the notion of collision resistance in the quantum setting. This study focuses on the collapsing security of Nandi’s double-block-length construction based on a random oracle. This study proposes that when the permutations $ \pi $ within the construction is composed completely of a number of c-cycle permutations (i.e., for any $ {x}\in{{\{0, 1\}}}^{{m}} $, $ \pi^{{a}}{(x) =x} $ if and only if $ {a=c} $), the collapsing security of this construction is optimal. Constructing a permutation $ \pi $ composed solely of c-cycle permutations is straightforward. Optimal collapsing security implies that when the output size of the random oracle is n bits, the adversary can effectively distinguish between the two states, measuring the hash value of a quantum superposition of messages and measuring the message superposition itself, only after making at least $ \text{O(}{{2}}^{{2n/3}}\text{)} $ queries. The proposed optimal construction can also be extended by the Merkle-Damg?rd construction. The extended hash function retains the collapsing property. Therefore, this study provides a theoretical foundation for the design of collapsing hash functions in the future.

Abstract:With the increasing complexity of cyber attacks, the collection and integration of threat intelligence face challenges of fragmentation. In this context, enterprises have begun building private threat intelligence knowledge graph systems. However, traditional construction methods suffer from inefficiency and limited effectiveness due to inaccurate extraction of entities and their relationships. To address these issues, this study proposes OPFA: a general threat intelligence knowledge graph construction framework based on model context protocol (MCP) agents. Through dynamic semantic subject positioning, large language model prompt templates are fine-tuned to drive agents to extract threat entities and their relationships. These threat entities and relationships are automatically created as nodes and edges in the knowledge graph, and threat intelligence is then linked through entity attribute value matching to form a complete knowledge graph. The system subsequently retrieves MCP resources for critical threat entities (such as vulnerabilities and malicious samples) to achieve customized knowledge graph expansion. Experimental results demonstrate that the agents not only improve Precision (97.22% and 97.83%, respectively), Recall (90.91% and 95.52%, respectively), and F1-score (94.44% and 96.66%, respectively) in threat entity and entity relationship extraction, but also effectively enhance the construction efficiency of the knowledge graph.