Dejun Ning

Large language models (LLMs) have demonstrated strong performance on complex reasoningtasks, yet their probabilistic generation mechanism still makes the reasoning process difficult tocontrol and prone to hallucination, reasoning drift, and logical inconsistency. Existing neurosymbolic methods improve reasoning reliability by combining LLMs with symbolic solvers,but they often provide limited modeling of deep semantic structure and implicit constraints,and they commonly rely on template-like symbolic code generation. These limitations makedynamic control of complex reasoning processes difficult. To address this problem, this paperproposes C²NS (Category-Constrained Neuro-Symbolic), a controllable neuro-symbolic framework. C²NS uses objects and morphisms from category-theoretic structure to construct structurepreserving semantic representations, combines implicit rule explicitization to complete latentconstraints, and thereby builds deep semantic models for complex reasoning contexts. It furtheruses collaborative intent routing and backtracking feedback to guide LLMs in dynamicallygenerating symbolic solving code, improving the generalization capability of symbolic programgeneration and enabling dynamic control over the reasoning process. Experiments on AR-LSAT,FOLIO, and ProofWriter show that C²NS consistently outperforms existing neuro-symbolicreasoning methods and chain-of-thought baselines, achieving stronger reasoning performance,symbolic solving stability, and overall reliability

Automated cardiac sound classification using phonocardiograms (PCGs) is pivotal for early cardiovascular disease screening, yet its reliability is frequently challenged by environmental noise and physiological interference such as lung sounds. To reconcile this robustness-fidelity trade-off, we propose D-IFGNet, a novel framework that integrates a D-pad based decomposition backbone with IMF Noise-aware Gating and Clean-adaptive Feature Fusion. Unlike methods relying on fixed basis functions, our architecture employs a learnable decomposition-reconstruction-decoupling mechanism to extract intrinsic mode-like components directly from raw signals. The IMF Noise-aware Gating utilizes frequency-band priors and learned corrections to dynamically reweight these components, effectively isolating diagnostic features from complex interference. Furthermore, a Clean-adaptive Feature Fusion module incorporates a soft switching strategy to adaptively balance noise-robust aggregation with the preservation of local temporal details in high-quality recordings. Extensive experiments on datasets contaminated with Additive White Gaussian Noise (AWGN) and real-world lung sounds demonstrate that D-IFGNet achieves consistent robustness across heterogeneous acoustic environments. In noisy conditions, our method outperforms the state-of-the-art baseline NRC-net, improving classification accuracy by an average of 1.55\%, while successfully maintaining average diagnostic precision on clean clinical signals.

Recent AI scientist systems demonstrate the promise of massive language models for autonomous scientific ideation; yet, many present approaches rely on rigid interaction protocols and coarse-grained novelty tests, which impede long-term creative exploration. In this paper, we proposed a multi-agent scientific ideation framework that blends a World Café-style collaborative structure and a multi-strategy turn-taking scheduler, allowing for organized yet flexible interactions. Agents are divided into rotating conversation groups, and speaking turns are distributed adaptively based on relevance, participation balance, and self-reported willingness, thereby reducing communication bottlenecks. To achieve a more detailed assessment of idea novelty, we incorporate the Relative Novelty Density (RND) metric into our novelty assessment system. Under regulated discussion cost conditions and with public academic knowledge graphs, our tests show improvements in total novelty compared to other single-agent and multi-agent baselines. Both the multi-strategy turn-taking scheduler and the RND metric enhance the systems creative performance.

With smart buildings being widely adopted in urban digital transformation, interactive semantic question answering (QA) systems serve as a crucial bridge between user intent and environmental response. However, they still face substantial challenges in semantic understanding and dynamic reasoning. Most existing systems rely on static frameworks built upon Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG), which suffer from rigid prompt design, breakdowns in multi-step reasoning, and inaccurate generation. To tackle these issues, we propose LoopRAG, a multi-agent RAG architecture that incorporates a Plan–Do–Check–Act (PDCA) closed-loop optimization mechanism. The architecture formulates a dynamic QA pipeline across four stages: task parsing, knowledge extraction, quality evaluation, and policy feedback, and further introduces a semantics-driven prompt reconfiguration algorithm and a heterogeneous knowledge fusion module. These components strengthen multi-source information handling and adaptive reasoning. Experiments on HotpotQA, MultiHop-RAG, and an in-house building QA dataset demonstrate that LoopRAG significantly outperforms conventional RAG systems in key metrics, including context recall of 90%, response relevance of 72%, and answer accuracy of 88%. The results indicate strong robustness and cross-task generalization. This work offers both theoretical foundations and an engineering pathway for constructing trustworthy and scalable semantic QA interaction systems in smart building settings.

Photovoltaic power forecasting plays a crucial role in the integration of renewable energy into the power grid. However, existing methods suffer from issues such as cumulative multi-step prediction errors and the limitations of traditional evaluation metrics (e.g., MSE, MAE). To address these challenges, this study introduces DTCformer, a generative forecasting model based on Autoformer. The proposed model integrates a Temporal Convolution Feedforward Network module and a Variable Selection Embedding module, effectively capturing inter-variable dependencies and temporal periodicity. Furthermore, it incorporates the DILATE loss function, which significantly enhances both forecasting accuracy and robustness. Experimental results on publicly available datasets demonstrate that DTCformer surpasses mainstream models, improving overall performance metrics (DILATE values) by 5.0–42.3% in 24 h, 48 h, and 72 h forecasting tasks.

Knowledge distillation is a model compression method that transforms complex models into efficient ones. Traditional distillation, with two-stage training, is time-consuming and computationally expensive. Self-distillation methods can alleviate this deficiency by adopting a one-stage strategy. However, most of them lack a meticulously designed auxiliary network, resulting in the learned knowledge being simplistic and insufficient, which limits the ability of distillation. To address this challenge, we propose a training framework named deep-supervised Knowledge Enhancement Self-Distillation (KED) which organizes the teacher model hierarchically through stacking auxiliary classifiers after each shallow block to form the student models. Specifically, two modules, i.e., logit decouple distillation (LDD) and attention map generation (AT-Gen), are embedded in the framework to enhance the distillation knowledge. LDD divides the output of each classifier (logits) into target and non-target classes, making the knowledge from outputs more flexible and efficient. Moreover, AT-Gen extracts attention maps from the features of each block, emphasizing the knowledge from intermediate layers through attention maps' integration and interaction. Finally, knowledge from different sources works together to guide the training of the student models, compensating for the bias of the auxiliary networks. Experiments on public datasets demonstrate that our method outperforms other state-of-the-art methods.

In recent years, Large Language Models (LLMs) with massive parameters and complex structures have achieved significant breakthroughs in fields such as natural language processing and image generation, driving their widespread application in industrial sectors. Despite the enormous potential of industrial AI models in areas like design and development, monitoring and management, quality control, and maintenance, their actual construction and deployment still face a lot of challenges, including inherent model deficiencies and difficulties in aligning with industrial requirements. Future technological development trends include the generation of customized industrial datasets, the collaborative optimization of large and small models, the enhancement of adaptive capabilities, and the application of Retrieval-Augmented Generation (RAG) technology. These trends are expected to improve the effectiveness and scalability of AI models, better meeting the needs of the industrial domain. This paper systematically discusses the challenges, technological development trends, and practical applications and deployment of industrial AI models, providing valuable insights for future directions.

In a smart home with distributed energy resources, the home energy management system (HEMS) controls the photovoltaic (PV) storage system by executing the optimization algorithm to achieve the lowest power cost. The existing mixed integer linear programming (MILP) algorithm is not suitable for execution on the end-user side due to its high computational complexity. The HEMS algorithm based on a long short-term memory neural network (LSTM-HEMS) can effectively solve the problem of the high computational complexity of MILP, but its optimization outcome is not high due to the accumulation of prediction errors. In order to achieve a better balance between computational complexity and optimization outcome, this paper proposes a lightweight optimization algorithm called the FastInformer-HEMS, which introduces the E-Attn attention mechanism based on Informer and uses global average pooling to extract the attention characteristics. Meanwhile, the proposed method introduces the maximum self-consumption algorithm as a backup strategy to ensure the safe operation of the system. The simulated results show that the computational complexity of the proposed FastInformer-HEMS is the lowest among the existing algorithms. Compared with the existing LSTM-HEMS, the proposed algorithm reduces the power consumption cost by 12.3% and 6.6% in the two typical scenarios, while the execution time decreases by 13.6 times.

Despite the growing capabilities of the short-term prediction of photovoltaic power, we still face two challenges to longer time-range predictions: error accumulation and long-term time series feature extraction. In order to improve the longer time range prediction accuracy of photovoltaic power, this paper proposes a seq2seq prediction model TCNformer, which outperforms other state-of-the-art (SOTA) algorithms by introducing variable selection (VS), long- and short-term time series feature extraction (LSTFE), and one-step temporal convolutional network (TCN) decoding. A VS module employs correlation analysis and periodic analysis to separate the time series correlation information, LSTFE extracts multiple time series features from time series data, and one-step TCN decoding realizes generative predictions. We demonstrate here that TCNformer has the lowest mean squared error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) in contrast to the other algorithms in the field of the short-term prediction of photovoltaic power, and furthermore, the effectiveness of each module has been verified through ablation experiments.