While large language models (LLMs) show considerable promise across various fields, they have notable limitations in handling multi-document question answering (Multi-doc QA) tasks. The first challenge is long-range dependency modeling, where LLMs struggle to focus on key information in long texts, which weakens important semantic connections. Second, most LLMs suffer from the ''lost-in-the-middle'' issue, where they have difficulty processing information in the middle of long inputs. Current solutions either truncate global dependencies or demand costly finetuning, ultimately lacking a universal and simple solution for these challenges. To resolve these limitations, we propose Dual-Stage Adaptive Sharpening (DSAS) containing two modules. (i) The Contextual Gate Weighting (CGW) module alleviates ''lost-in-the-middle'' by assessing paragraph relevance through layer-wise attention tracking and position-aware weighting. (ii) The Reciprocal Attention Suppression (RAS) module enhances focus on critical paragraphs by suppressing information exchange between key and irrelevant texts, thus mitigating the limitations in long-range dependency modeling. Extensive experiments on four benchmarks demonstrate DSAS's efficacy across mainstream LLMs (Llama, Qwen, Mistral, and Deepseek), with an average F1-score improvement of 4.2% in Multi-doc QA tasks on Llama-3.1-8B-Instruct and Qwen2.5-14B-Instruct. Ablation studies confirm the essential contributions of both the CGW and RAS modules. In addition, detailed discussions in the Appendix further validate the robustness and scalability of DSAS.

Magnetic Resonance Imaging (MRI) plays a critical role in clinical diagnostics, but its broader application is constrained by prolonged acquisition times, especially under high acceleration factors. We therefore propose a self-supervised MRI reconstruction framework that integrates latent diffusion models with a parallel dual-network architecture, aiming to achieving high-quality image reconstruction under highly accelerated conditions. Within this framework, a dual-branch network is designed to process distinct subsets of under-sampled k-space data. This design enables the capture of complementary information across subsets while effectively mitigating overfitting. Both branches are guided by reconstruction and difference losses to enforce consistency across unobserved regions, enabling accurate recovery of fully sampled images. A solid mathematical foundation is formulated to offer theoretical guarantees regarding the reliable approximation of fully sampled data under specific constraints. Experimental results demonstrate that integrating parallel dual-network selfsupervision with latent diffusion advances self-supervised MRI reconstruction, highlighting its potential for clinically feasible high-fidelity MRI under high acceleration.

CPS (Cyber-Physical Systems) are physical and engineered systems featuring a tight combination of computation and physical processes by communication networks. CPS are mainly applied in some critical domains, so it is very essential to ensure the correctness of CPS. As a very important method for verifying system correctness, formal verification has been successfully applied in the verification of CPS; however, the high theoretical level of formal modeling techniques in formal verification and the lack of generality of formal models make it difficult to integrate formal verification with enterprise standard system development process. In this paper, we model CPS by HybridUML, an extension of UML which is now the de facto standard modeling language in system development practice, and then transform HybridUML model into the input language of theorem prover KeYmaera-QHP(Quantified Hybrid Program), and finally verify the QHP code with KeYmaera. When transforming a HybridUML model, we define the metamodels of HybridUML and QHP, and then find the association between them based on semantic consisteency, and finally define transformation rules using ATL (ATLAS Transformation Language) and template based code generation method to achieve transformation.

This paper proposes wireless sensor networks as a parallel and distributed computing platform for neurocomputing. The proposal entails leveraging the existing wireless sensor networks technology to serve as a hardware-software platform to implement and realize artificial neural network algorithms in fully parallel and distributed computation mode. The study describes the proposed parallel and distributed neurocomputing architecture, which is named as WSN-ANN, and its use as a hardware platform on a case study. A Hopfield neural network, which is configured to solve the minimum weakly connected dominating set problem, is embedded into a wireless sensor network. Simulation study results indicate that the proposed computing platform based on wireless sensor networks, WSN-ANN, is feasible and promising to serve as a parallel and distributed neurocomputer.