Su Pan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Su Pan has a total of 94 co-authors. Here are the top co-authors by number of shared publications:
Zibo Li
School of Communication and Information Engineering Nanjing University of Posts and Telecommunications Nanjing China
6 shared publicationsWeiwei Zhou
Key Lab of Broadband Wireless Communication and Sensor Network Technology, Ministry of Education Nanjing University of Posts and Telecommunications Nanjing China
3 shared publicationsXixia Sun
College of Internet of Things, Nanjing University of Posts and Telecommunications, China
3 shared publications- Xixia Sun
School of Internet of Things, Nanjing University of Posts and Telecommunications, Nanjing, China
3 shared publications Yuqing Chen
Key Lab of Broadband Wireless Communication and Sensor Network Technology, Jiangsu Engineering Research Center of Communication and Network Technology, Nanjing University of Posts and Telecommunications, Ministry of Education
3 shared publications- Yuxi Qin
School of Communication and Information Engineering Nanjing University of Posts and Telecommunications Nanjing Jiangsu China
3 shared publications
Latest Publications
- 31 July 2026
- Published by Elsevier in Neural Networks
- Vol. 205 (Pt B) , 109437
Distilling large language models for code generation via ranking supervision
- 31 July 2026
- Published by Elsevier in Neural Networks
- Vol. 205 (Pt B) , 109437
Distilling large language models for code generation via ranking supervision
- 21 July 2026
- Published by Springer Nature in Scientometrics
- Vol. 131 (9) , 6469-6496
CRM-RMS: context-aware citation recommender model using recurrent reasoning memory network over SPECTER
- 21 July 2026
- Published by Springer Nature in Scientometrics
- Vol. 131 (9) , 6469-6496
CRM-RMS: context-aware citation recommender model using recurrent reasoning memory network over SPECTER
- Open Access
- 1 March 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (5) , 841
A Two-Stage Algorithm for Time Series Compression: ARIMA-Based Pre-Compression and Reinforcement Learning Optimized Chunking
The explosive growth of time series gives rise to a large amount of data, which emphasizes the importance of data compression. The data compression not only reduces storage costs but also enhances data transmission efficiency and processing speed. However, traditional compression algorithms usually suffer an insufficient compression ratio and an excessive computational cost. To address these problems above, in this paper, we propose a two-stage compression algorithm for the large-scale time series data. In the first stage, we transform the time series data into low-volatility residual data by using Autoregressive Integrated Moving Average (ARIMA) modeling and apply adaptive precision quantization to improve compressibility. In the second stage, we implement a reinforcement learning-based compression strategy, which utilizes the Q-learning to select the number of blocks to divide the quantized data segment and achieves compression by storing the same content between the divided data blocks only once and storing the different content separately; and we incorporate the Upper Confidence Bound (UCB) to balance exploration and exploitation in order to track changes in data patterns and improve compression performance. Experimental results demonstrate that our algorithm achieves a higher compression ratio while maintaining a low computational complexity compared with traditional compression algorithms.
- 1 March 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (5) , 841
Open AccessA Two-Stage Algorithm for Time Series Compression: ARIMA-Based Pre-Compression and Reinforcement Learning Optimized Chunking
The explosive growth of time series gives rise to a large amount of data, which emphasizes the importance of data compression. The data compression not only reduces storage costs but also enhances data transmission efficiency and processing speed. However, traditional compression algorithms usually suffer an insufficient compression ratio and an excessive computational cost. To address these problems above, in this paper, we propose a two-stage compression algorithm for the large-scale time series data. In the first stage, we transform the time series data into low-volatility residual data by using Autoregressive Integrated Moving Average (ARIMA) modeling and apply adaptive precision quantization to improve compressibility. In the second stage, we implement a reinforcement learning-based compression strategy, which utilizes the Q-learning to select the number of blocks to divide the quantized data segment and achieves compression by storing the same content between the divided data blocks only once and storing the different content separately; and we incorporate the Upper Confidence Bound (UCB) to balance exploration and exploitation in order to track changes in data patterns and improve compression performance. Experimental results demonstrate that our algorithm achieves a higher compression ratio while maintaining a low computational complexity compared with traditional compression algorithms.
- Open Access
Reinforcement Learning Enhanced Multi-hop Reasoning for Temporal Knowledge Question Answering
Temporal knowledge graph question answering (TKGQA) involves multi-hop reasoning over temporally constrained entity relationships in the knowledge graph to answer a given question. However, at each hop, large language models (LLMs) retrieve subgraphs with numerous temporally similar and semantically complex relations, increasing the risk of suboptimal decisions and error propagation. To address these challenges, we propose the multi-hop reasoning enhanced (MRE) framework, which enhances both forward and backward reasoning to improve the identification of globally optimal reasoning trajectories. Specifically, MRE begins with prompt engineering to guide the LLM in generating diverse reasoning trajectories for a given question. Valid reasoning trajectories are then selected for supervised fine-tuning, serving as a cold-start strategy. Finally, we introduce Tree-Group Relative Policy Optimization (T-GRPO), a recursive, tree-structured learning-by-exploration approach. At each hop, exploration establishes strong causal dependencies on the previous hop, while evaluation is informed by multi-path exploration feedback from subsequent hops. Experimental results on two TKGQA benchmarks indicate that the proposed MRE-based model consistently surpasses state-of-the-art (SOTA) approaches in handling complex multi-hop queries. Further analysis highlights improved interpretability and robustness to noisy temporal annotations.
Open AccessReinforcement Learning Enhanced Multi-hop Reasoning for Temporal Knowledge Question Answering
Temporal knowledge graph question answering (TKGQA) involves multi-hop reasoning over temporally constrained entity relationships in the knowledge graph to answer a given question. However, at each hop, large language models (LLMs) retrieve subgraphs with numerous temporally similar and semantically complex relations, increasing the risk of suboptimal decisions and error propagation. To address these challenges, we propose the multi-hop reasoning enhanced (MRE) framework, which enhances both forward and backward reasoning to improve the identification of globally optimal reasoning trajectories. Specifically, MRE begins with prompt engineering to guide the LLM in generating diverse reasoning trajectories for a given question. Valid reasoning trajectories are then selected for supervised fine-tuning, serving as a cold-start strategy. Finally, we introduce Tree-Group Relative Policy Optimization (T-GRPO), a recursive, tree-structured learning-by-exploration approach. At each hop, exploration establishes strong causal dependencies on the previous hop, while evaluation is informed by multi-path exploration feedback from subsequent hops. Experimental results on two TKGQA benchmarks indicate that the proposed MRE-based model consistently surpasses state-of-the-art (SOTA) approaches in handling complex multi-hop queries. Further analysis highlights improved interpretability and robustness to noisy temporal annotations.
- 5 November 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 299
Reward-guided direct preference optimization
- 5 November 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 299
Reward-guided direct preference optimization
- Open Access
- 31 August 2025
- Published by MDPI AG in Electronics
- Vol. 14 (17) , 3489
LR-SQL: A Supervised Fine-Tuning Method for Text2SQL Tasks Under Low-Resource Scenarios
In supervised fine-tuning (SFT) for Text2SQL tasks, particularly for databases with numerous tables, encoding schema features requires excessive tokens, escalating GPU resource requirements during fine-tuning. To bridge this gap, we propose LR-SQL, a general dual-model SFT framework comprising a schema linking model and an SQL generation model. At the core of our framework lies the schema linking model, which is trained on a novel downstream task termed slice-based related table filtering. This task dynamically partitions a database into adjustable slices of tables and sequentially evaluates the relevance of each slice to the input query, thereby reducing token consumption per iteration. However, slicing fragments destroys database information, impairing the model’s ability to comprehend the complete database. Thus, we integrate Chain of Thought (CoT) in training, enabling the model to reconstruct the full database context from discrete slices, thereby enhancing inference fidelity. Ultimately, the SQL generation model uses the result from the schema linking model to generate the final SQL. Extensive experiments demonstrate that our proposed LR-SQL reduces total GPU memory usage by 40% compared to baseline SFT methods, with only a 2% drop in table prediction accuracy for the schema linking task and a negligible 0.6% decrease in overall Text2SQL Execution Accuracy.
- 31 August 2025
- Published by MDPI AG in Electronics
- Vol. 14 (17) , 3489
Open AccessLR-SQL: A Supervised Fine-Tuning Method for Text2SQL Tasks Under Low-Resource Scenarios
In supervised fine-tuning (SFT) for Text2SQL tasks, particularly for databases with numerous tables, encoding schema features requires excessive tokens, escalating GPU resource requirements during fine-tuning. To bridge this gap, we propose LR-SQL, a general dual-model SFT framework comprising a schema linking model and an SQL generation model. At the core of our framework lies the schema linking model, which is trained on a novel downstream task termed slice-based related table filtering. This task dynamically partitions a database into adjustable slices of tables and sequentially evaluates the relevance of each slice to the input query, thereby reducing token consumption per iteration. However, slicing fragments destroys database information, impairing the model’s ability to comprehend the complete database. Thus, we integrate Chain of Thought (CoT) in training, enabling the model to reconstruct the full database context from discrete slices, thereby enhancing inference fidelity. Ultimately, the SQL generation model uses the result from the schema linking model to generate the final SQL. Extensive experiments demonstrate that our proposed LR-SQL reduces total GPU memory usage by 40% compared to baseline SFT methods, with only a 2% drop in table prediction accuracy for the schema linking task and a negligible 0.6% decrease in overall Text2SQL Execution Accuracy.
- Open Access
Schema-R1: A reasoning training approach for schema linking in Text-to-SQL Task
Schema linking is a critical step in Text-to-SQL task, aiming to accurately predict the table names and column names required for the SQL query based on the given question. However, current fine-tuning approaches for schema linking models employ a rote-learning paradigm, excessively optimizing for ground truth schema linking outcomes while compromising reasoning ability. This limitation arises because of the difficulty in acquiring a high-quality reasoning sample for downstream tasks. To address this, we propose Schema-R1, a reasoning schema linking model trained using reinforcement learning. Specifically, Schema-R1 consists of three key steps: constructing small batches of high-quality reasoning samples, supervised fine-tuning for cold-start initialization, and rule-based reinforcement learning training. The final results demonstrate that our method effectively enhances the reasoning ability of the schema linking model, achieving a 10\% improvement in filter accuracy compared to the existing method. Our code is available at https://github.com/hongWin/Schema-R1/.
Open AccessSchema-R1: A reasoning training approach for schema linking in Text-to-SQL Task
Schema linking is a critical step in Text-to-SQL task, aiming to accurately predict the table names and column names required for the SQL query based on the given question. However, current fine-tuning approaches for schema linking models employ a rote-learning paradigm, excessively optimizing for ground truth schema linking outcomes while compromising reasoning ability. This limitation arises because of the difficulty in acquiring a high-quality reasoning sample for downstream tasks. To address this, we propose Schema-R1, a reasoning schema linking model trained using reinforcement learning. Specifically, Schema-R1 consists of three key steps: constructing small batches of high-quality reasoning samples, supervised fine-tuning for cold-start initialization, and rule-based reinforcement learning training. The final results demonstrate that our method effectively enhances the reasoning ability of the schema linking model, achieving a 10\% improvement in filter accuracy compared to the existing method. Our code is available at https://github.com/hongWin/Schema-R1/.
- Open Access
- 1 January 2025
- Published by Institution of Engineering and Technology (IET) in IET Communications
- Vol. 19 (1)
Cloud‐Fog Cooperative Computation Offloading and Resource Allocation in Heterogeneous Networks Based on Genetic Algorithm
In this paper, we investigate the computation offloading and resource allocation strategy of the coexistence and synergy between fog computing and cloud computing in heterogeneous networks. Consider that the reported schemes have prohibitive complexity when achieving the optimal computation offloading strategy in cloud‐fog cooperative heterogeneous networks, an improved genetic algorithm (IGA) is proposed in this paper, which can maintain a low computation complexity while obtaining the optimal solution. In the IGA algorithm, we propose to use a penalty function to express the constraint conditions of the optimisation problem and use a non‐uniform mutation operator to accelerate the convergence speed. Besides, an improved method of parameter self‐adaptation and a perturbation method of mutation probability based on population fitness standard deviation are proposed to optimise the genetic algorithm. The numerical results show that the proposed genetic algorithm can obtain a lower average cost of the system while keeping a smaller computational cost.
- 1 January 2025
- Published by Institution of Engineering and Technology (IET) in IET Communications
- Vol. 19 (1)
Open AccessCloud‐Fog Cooperative Computation Offloading and Resource Allocation in Heterogeneous Networks Based on Genetic Algorithm
In this paper, we investigate the computation offloading and resource allocation strategy of the coexistence and synergy between fog computing and cloud computing in heterogeneous networks. Consider that the reported schemes have prohibitive complexity when achieving the optimal computation offloading strategy in cloud‐fog cooperative heterogeneous networks, an improved genetic algorithm (IGA) is proposed in this paper, which can maintain a low computation complexity while obtaining the optimal solution. In the IGA algorithm, we propose to use a penalty function to express the constraint conditions of the optimisation problem and use a non‐uniform mutation operator to accelerate the convergence speed. Besides, an improved method of parameter self‐adaptation and a perturbation method of mutation probability based on population fitness standard deviation are proposed to optimise the genetic algorithm. The numerical results show that the proposed genetic algorithm can obtain a lower average cost of the system while keeping a smaller computational cost.
- 1 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Chinese Journal of Electronics
- Vol. 33 (6) , 1515-1525
Service Migration Algorithm Based on Markov Decision Process with Multiple Service Types and Multiple System Factors
This paper proposes a Markov decision process based service migration algorithm to satisfy quality of service (QoS) requirements when the terminals leave the original server. Services were divided into real-time services and non-real-time services, each type of them has different requirements on transmission bandwidth and latency, which were considered in the revenue function. Different values were assigned to the weight coefficients of QoS parameters for different service types in the revenue and cost functions so as to distinguish the differences between the two service types. The overall revenue was used for migration decisions, rather than fixed threshold or instant revenue. The Markov decision process was used to maximize the overall revenue of the system. Simulation results show that the proposed algorithm obtained more revenue compared with the existing works.
- 1 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Chinese Journal of Electronics
- Vol. 33 (6) , 1515-1525
Service Migration Algorithm Based on Markov Decision Process with Multiple Service Types and Multiple System Factors
This paper proposes a Markov decision process based service migration algorithm to satisfy quality of service (QoS) requirements when the terminals leave the original server. Services were divided into real-time services and non-real-time services, each type of them has different requirements on transmission bandwidth and latency, which were considered in the revenue function. Different values were assigned to the weight coefficients of QoS parameters for different service types in the revenue and cost functions so as to distinguish the differences between the two service types. The overall revenue was used for migration decisions, rather than fixed threshold or instant revenue. The Markov decision process was used to maximize the overall revenue of the system. Simulation results show that the proposed algorithm obtained more revenue compared with the existing works.