Acupuncture prescription generation is not merely a text generation problem. A clinically plausible prescription must connect case descriptions with acupoint combinations that respect meridian routes, affiliations, pairings, and cross-meridian relations. However, most large language models serialize such knowledge as text, making it difficult to preserve the structure that gives acupuncture prescriptions their domain logic. This paper presents TopoAcu, a meridian topology-aware graph-language model for structure-aware acupuncture prescription generation. TopoAcu separates acupoint entity semantics from meridian-relation topology alignment. A training-time meridian graph provides a fixed structural reference space, which is transferred to the language model through node-level acupoint alignment and prescription-level structural similarity constraints. A staged curriculum introduces these constraints after instruction tuning to stabilize graph-language fusion. Across a held-out standardized case set, TopoAcu improved prescription matching, meridian-level structural consistency, latent structural organization, and expert-rated plausibility relative to instruction-tuned, text-based knowledge-enhanced, and graph-control baselines. These results support topology-aware graph-language alignment as a practical strategy for preserving domain structure in knowledge-intensive decision-support tasks.
Low-light image enhancement (LLIE) aims to restore scene visibility while preserving color fidelity and structural integrity. Existing methods typically process luminance and chrominance jointly within the coupled RGB space, neglecting their asymmetric degradation; meanwhile, the complementarity between frequency and spatial domains remains under-exploited, and prevailing State Space Models (SSMs) rely on fixed scanning orders ill-suited to spatially non-uniform degradation. To address these limitations, we propose ICDMamba, a novel LLIE network driven by illumination-color decoupling and frequency-spatial collaborative modeling. ICDMamba first maps the input into the HSV space to explicitly separate the luminance prior from chrominance, alleviating color-shift distortions during brightening. A dual-branch architecture then models amplitude-phase complementarity in the frequency domain via a Cross-Guided Amplitude-Phase Attention (CGAPA) mechanism, while a Multiscale Pooling Mamba Block (MPMB) captures multi-scale spatial context. Crucially, a Degradation-Aware State Space Module (DASSM) dynamically reorders the token sequence guided by the illumination prior, enabling Mamba-based SSM processing to prioritize severely degraded regions. Extensive experiments on multiple benchmarks demonstrate that ICDMamba achieves state-of-the-art performance in both perceptual quality and cross-scene generalization, ranking first in comprehensive evaluation (RoR = 1).
In recent years, optimization problems in engineering, control, and machine learning have become increasingly nonconvex, constrained, distributed, and large-scale, making them more difficult for classical methods to solve reliably and efficiently. Collaborative neurodynamic optimization (CNO) has emerged as a promising framework that combines neurodynamic solvers with collaborative mechanisms among multiple agents or dynamical systems. Through local neurodynamic evolution and information exchange, CNO enhances search capability, supports distributed implementation, and improves adaptability in complex optimization environments. This survey aims to provide a systematic review and unified view of CNO as a general framework that integrates optimization modeling, neurodynamic solving, and collaborative enhancement for complex systems. Existing methods are classified from the perspectives of neurodynamic solvers and collaborative mechanisms, and recent advances and representative applications are discussed.
Digital twin modeling provides an effective approach for real-time condition monitoring of renewable energy generation systems, but achieving a balance between modeling accuracy and computational efficiency remains challenging. This paper proposes PhyOLNet, a physics-assisted optimal long short-term memory (LSTM) framework for lightweight digital twin modeling of renewable energy systems. Unlike conventional hybrid approaches that directly couple physical and data-driven models, the proposed method utilizes a reduced-order physical model as a structured feature extractor to generate physically meaningful state representations. A Physical-Temporal Gating Fusion mechanism is developed to adaptively integrate physics-derived features, historical system outputs, and environmental observations, while an optimized shallow LSTM network learns residual nonlinear relationships and temporal dependencies. By transforming physical knowledge into compact feature representations, the proposed framework improves model traceability while maintaining a lightweight architecture suitable for online deployment. Experimental evaluations on photovoltaic and wind power systems demonstrate that PhyOLNet achieves accurate and stable condition modeling across different operating scenarios. In particular, photovoltaic validation under six diverse scenarios shows that the proposed method consistently achieves the lowest rRMSE among comparison approaches while maintaining stable error characteristics. The results verify the effectiveness and generalization capability of PhyOLNet as a lightweight digital twin modeling framework for renewable energy applications.
Existing high-resolution remote sensing image segmentation (RSIS) methods universally face two types of semantic ambiguity problems when handling complex scenes: intra-class inconsistency and segmentation offsets. This is primarily attributed to traditional paradigms relying heavily on passive upsampling and static skip connections, which lack the capability to selectively extract the frequency aliasing information introduced during the encoder’s downsampling process. To address this, this paper proposes the Pyramid Frequency Query Guided Network (PFQGNet), redefining the decoding process as a cross-spatial dual information retrieval paradigm. The missing information during decoder upsampling objectively exists within the multi-level features of the encoder, and precisely characterizing this missing information is the core challenge for effective retrieval. First, we design Frequency Query Guided Upsampling (FQGU) as the primary retrieval mechanism, employing a spatial convolution-based frequency decoupling approach. The Low-Frequency Query Generator (LFQG) actively attenuates high-frequency components to retrieve smooth semantics, while the High-Frequency Query Generator (HFQG) extracts structural residuals; by synergizing these decoupled representations, it effectively mitigates the two aforementioned ambiguities. Second, Semantically-Enhanced Fusion Attention (SEFA) serves as the secondary retrieval mechanism, utilizing the frequency-repaired features as spatial query conditions to retrieve multi-level encoder features, thereby compensating for spatial details lost to aliasing. Finally, Multi-Dimensional Collaborative Attention (MDCA) dynamically rebalances the retrieved semantic and structural features. Experiments on multiple datasets demonstrate that the segmentation performance of PFQGNet outperforms existing methods. The code is located at https://github.com/CDRstn/PFQGNet.
