Optimal growth under irreversible pollution: The Forster-Ramsey model
研究了在不可逆污染条件下，通过Forster-Ramsey模型寻求最优经济增长路径的问题，引入最优减排控制以评估其对不可逆污染的影响。
0 citationsRead paper
Selected work
Representative Papers
Recent Advances in Deep Learning-Based Drug-Target Binding Affinity Prediction
This study addresses the limitations in accuracy and evaluation of drug-target binding affinity prediction by systematically reviewing deep learning models and benchmark datasets. It identifies critical bottlenecks, including data bias, poor cold-start generalization, and the absence of standardized evaluation metrics. To overcome these challenges, this work proposes novel research paradigms encompassing standardized evaluation frameworks, improved dataset curation, and multimodal representation integration. By elucidating the core deficiencies of current methodologies, this paper establishes a future research roadmap prioritizing robustness and generalizability. Ultimately, this work provides essential theoretical foundations and methodological guidance to advance precision drug discovery, ensuring more reliable computational predictions in pharmaceutical development.
0 citationsRead paper
UA-ChatDev: Uncertainty-Aware Multi-Agent Collaboration for Reliable Software Development
This work addresses the issue of hallucination propagation in existing large language model–driven multi-agent software development frameworks, which often stems from neglecting the reliability of intermediate outputs and ultimately degrades software quality. To mitigate this, the authors propose an uncertainty-aware multi-agent collaboration framework that quantifies response uncertainty through lightweight token-level log-probability estimation. By integrating phase-adaptive threshold calibration, the framework selectively triggers retrieval-augmented verification at high-risk stages, enabling reliable and context-sensitive collaboration. Evaluated on the SRDD benchmark, the proposed approach significantly outperforms current single- and multi-agent methods across key metrics, including completeness, executability, consistency, and overall software quality.
0 citationsRead paper
Multi-task GINN-LP for Multi-target Symbolic Regression
Symbolic regression (SR) faces two key limitations in interpretable AI: (1) evaluation relies heavily on idealized scientific datasets, resulting in poor generalization; and (2) dominant approaches support only single-output modeling, hindering the capture of interdependencies among multiple objectives. This work pioneers multi-objective symbolic regression by proposing MTRGINN-LP—a novel architecture that employs a Graph Isomorphism Neural Network (GINN)-based power-term approximation module as a shared backbone, integrated with multi-task learning and an interpretable neural network design. Task-specific output heads and a divide-and-conquer loss function jointly optimize symbolic expression discovery and multi-output prediction. Evaluated on real-world multi-task benchmarks—including energy efficiency forecasting and sustainable agriculture—MTRGINN-LP achieves both high accuracy (12.7% lower MAE) and strong interpretability, while markedly improving cross-domain generalization.
0 citationsRead paper
Transmitter Identification and Protocol Categorization in Shared Spectrum via Multi-Task RF Classification at the Network Edge
In shared-spectrum environments, transmitter identification and communication protocol classification suffer from strong task coupling, severe feature overlap, and significant channel-induced interference. To address these challenges, this paper proposes a multi-channel input, multi-task convolutional neural network (CNN) framework for joint radio-frequency signal classification. The architecture employs a shared feature extraction backbone coupled with task-specific branches to simultaneously model transmitter-specific radio-frequency fingerprints and protocol-level semantic features, thereby enhancing model generalization and robustness. Evaluated on the real-world POWDER dataset, the framework achieves 90% accuracy for protocol classification, 100% for base station identification, and 92% for joint task classification—outperforming single-task baselines by an average of 7.3%. This work establishes a deployable intelligent sensing foundation for spectrum policy enforcement, dynamic spectrum sharing, and wireless network security monitoring.
0 citationsRead paper
Recent publications
Latest Papers
Optimal growth under irreversible pollution: The Forster-Ramsey model
研究了在不可逆污染条件下，通过Forster-Ramsey模型寻求最优经济增长路径的问题，引入最优减排控制以评估其对不可逆污染的影响。
0 citationsRead paper
Recent Advances in Deep Learning-Based Drug-Target Binding Affinity Prediction
This study addresses the limitations in accuracy and evaluation of drug-target binding affinity prediction by systematically reviewing deep learning models and benchmark datasets. It identifies critical bottlenecks, including data bias, poor cold-start generalization, and the absence of standardized evaluation metrics. To overcome these challenges, this work proposes novel research paradigms encompassing standardized evaluation frameworks, improved dataset curation, and multimodal representation integration. By elucidating the core deficiencies of current methodologies, this paper establishes a future research roadmap prioritizing robustness and generalizability. Ultimately, this work provides essential theoretical foundations and methodological guidance to advance precision drug discovery, ensuring more reliable computational predictions in pharmaceutical development.
0 citationsRead paper
UA-ChatDev: Uncertainty-Aware Multi-Agent Collaboration for Reliable Software Development
This work addresses the issue of hallucination propagation in existing large language model–driven multi-agent software development frameworks, which often stems from neglecting the reliability of intermediate outputs and ultimately degrades software quality. To mitigate this, the authors propose an uncertainty-aware multi-agent collaboration framework that quantifies response uncertainty through lightweight token-level log-probability estimation. By integrating phase-adaptive threshold calibration, the framework selectively triggers retrieval-augmented verification at high-risk stages, enabling reliable and context-sensitive collaboration. Evaluated on the SRDD benchmark, the proposed approach significantly outperforms current single- and multi-agent methods across key metrics, including completeness, executability, consistency, and overall software quality.
0 citationsRead paper
Multi-task GINN-LP for Multi-target Symbolic Regression
Symbolic regression (SR) faces two key limitations in interpretable AI: (1) evaluation relies heavily on idealized scientific datasets, resulting in poor generalization; and (2) dominant approaches support only single-output modeling, hindering the capture of interdependencies among multiple objectives. This work pioneers multi-objective symbolic regression by proposing MTRGINN-LP—a novel architecture that employs a Graph Isomorphism Neural Network (GINN)-based power-term approximation module as a shared backbone, integrated with multi-task learning and an interpretable neural network design. Task-specific output heads and a divide-and-conquer loss function jointly optimize symbolic expression discovery and multi-output prediction. Evaluated on real-world multi-task benchmarks—including energy efficiency forecasting and sustainable agriculture—MTRGINN-LP achieves both high accuracy (12.7% lower MAE) and strong interpretability, while markedly improving cross-domain generalization.
0 citationsRead paper
Transmitter Identification and Protocol Categorization in Shared Spectrum via Multi-Task RF Classification at the Network Edge
In shared-spectrum environments, transmitter identification and communication protocol classification suffer from strong task coupling, severe feature overlap, and significant channel-induced interference. To address these challenges, this paper proposes a multi-channel input, multi-task convolutional neural network (CNN) framework for joint radio-frequency signal classification. The architecture employs a shared feature extraction backbone coupled with task-specific branches to simultaneously model transmitter-specific radio-frequency fingerprints and protocol-level semantic features, thereby enhancing model generalization and robustness. Evaluated on the real-world POWDER dataset, the framework achieves 90% accuracy for protocol classification, 100% for base station identification, and 92% for joint task classification—outperforming single-task baselines by an average of 7.3%. This work establishes a deployable intelligent sensing foundation for spectrum policy enforcement, dynamic spectrum sharing, and wireless network security monitoring.
0 citationsRead paper