Aijia Sun
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Aijia Sun has a total of 5 co-authors. Here are the top co-authors by number of shared publications:
- Chengrui Zhou
Fu Foundation School of Engineering and Applied Science Columbia University New York NY USA
1 shared publication - Dannier Li
School of Computing University of Nebraska -Lincoln Lincoln NE USA
1 shared publication - Yinuo Yang
Independent Researcher, China
1 shared publication - Yushang Zhao1 shared publication
- Zhongheng Yang
College of Computer Sciences Northeastern University Jersey City NJ USA
1 shared publication
Latest Publications
- Open Access
- 15 October 2025
- Published by MDPI AG in Preprints.org
Safety-Constrained Real-Time Decision Making for Autonomous Vehicles via NEURAL-QWEN
This paper introduces NEURAL-QWEN (Neuro-Enhanced Unified Reasoning and Adaptive Learning for Qwen-based Driving), a framework for real-time decision-making and safety constraint enforcement in autonomous driving. The framework uses a Mixture of Experts (MoE) to cut reasoning latency and cost, Adaptive Low-Rank Decomposition (AdaLoRA) to avoid catastrophic forgetting, and a Time Fusion Transformer (TFT) for trajectory prediction. It also uses federated multi-agent coordination with differential privacy for secure knowledge sharing, and Constitutional AI safety constraints to enforce driving rules. The framework uses reinforcement learning with human feedback (RLHF) to align simulated and real-world driving, and a semantic attention mechanism (SAM) for cross-modal reasoning. NEURAL-QWEN combines fast performance with safety standards and gives a solution for safe and interpretable decision-making.
- 15 October 2025
- Published by MDPI AG in Preprints.org
Open AccessSafety-Constrained Real-Time Decision Making for Autonomous Vehicles via NEURAL-QWEN
This paper introduces NEURAL-QWEN (Neuro-Enhanced Unified Reasoning and Adaptive Learning for Qwen-based Driving), a framework for real-time decision-making and safety constraint enforcement in autonomous driving. The framework uses a Mixture of Experts (MoE) to cut reasoning latency and cost, Adaptive Low-Rank Decomposition (AdaLoRA) to avoid catastrophic forgetting, and a Time Fusion Transformer (TFT) for trajectory prediction. It also uses federated multi-agent coordination with differential privacy for secure knowledge sharing, and Constitutional AI safety constraints to enforce driving rules. The framework uses reinforcement learning with human feedback (RLHF) to align simulated and real-world driving, and a semantic attention mechanism (SAM) for cross-modal reasoning. NEURAL-QWEN combines fast performance with safety standards and gives a solution for safe and interpretable decision-making.
- Open Access
- 26 September 2025
- Published by MDPI AG in Preprints.org
Real-Time Delivery Prediction Framework with Spatio-Temporal Fusion and LLM Semantic Enhancement
Food delivery services grew fast during the COVID-19 pandemic. This made it harder to predict rider behavior and delivery time. Many models do not handle the changing and complex nature of this task well. This paper shows STELLAR, a new framework that uses large language models and spatio-temporal learning. STELLAR uses a Qwen-14B-based semantic module to get context from rider data, orders, and regions. A graph attention network captures spatio-temporal patterns. The framework also uses multi-task learning to predict actions and delivery time together. Experiments show STELLAR works better than other models in this setting.
- 26 September 2025
- Published by MDPI AG in Preprints.org
Open AccessReal-Time Delivery Prediction Framework with Spatio-Temporal Fusion and LLM Semantic Enhancement
Food delivery services grew fast during the COVID-19 pandemic. This made it harder to predict rider behavior and delivery time. Many models do not handle the changing and complex nature of this task well. This paper shows STELLAR, a new framework that uses large language models and spatio-temporal learning. STELLAR uses a Qwen-14B-based semantic module to get context from rider data, orders, and regions. A graph attention network captures spatio-temporal patterns. The framework also uses multi-task learning to predict actions and delivery time together. Experiments show STELLAR works better than other models in this setting.
- Open Access
- 26 September 2025
- Published by MDPI AG in Preprints.org
A Scalable Multi-Agent Framework for Low-Resource E-Commerce Concept Extraction and Standardization
Understanding shopping concepts in e-commerce is hard because consumer behavior, product terms, and technical details vary a lot. This paper presents MALLM, a new multi-agent framework based on LLaMA-2 70B. It improves concept understanding by letting specialized agents work together. The model uses a layered agent system, domain-adaptive pretraining, retrieval-augmented generation (RAG), and cross-modal feature fusion to handle tasks clearly and well. It also applies knowledge distillation and multi-step training to adapt to small data. MALLM balances good results with practical deployment. It works well in real e-commerce situations.
- 26 September 2025
- Published by MDPI AG in Preprints.org
Open AccessA Scalable Multi-Agent Framework for Low-Resource E-Commerce Concept Extraction and Standardization
Understanding shopping concepts in e-commerce is hard because consumer behavior, product terms, and technical details vary a lot. This paper presents MALLM, a new multi-agent framework based on LLaMA-2 70B. It improves concept understanding by letting specialized agents work together. The model uses a layered agent system, domain-adaptive pretraining, retrieval-augmented generation (RAG), and cross-modal feature fusion to handle tasks clearly and well. It also applies knowledge distillation and multi-step training to adapt to small data. MALLM balances good results with practical deployment. It works well in real e-commerce situations.
- Open Access
RLHF Fine-Tuning of LLMs for Alignment with Implicit User Feedback in Conversational Recommenders
Conversational recommender systems (CRS) based on Large Language Models (LLMs) need to constantly be aligned to the user preferences to provide satisfying and context-relevant item recommendations. The traditional supervised fine-tuning cannot capture the implicit feedback signal, e.g., dwell time, sentiment polarity, or engagement patterns. In this paper, we share a fine-tuning solution using human feedback reinforcement learning (RLHF) to maximize implied user feedback (IUF) in a multi-turn recommendation context. We specify a reward model $R_φ$ learnt on weakly-labelled engagement information and maximize user-centric utility by optimizing the foundational LLM M_θ through a proximal policy optimization (PPO) approach. The architecture models conversational state transitions $s_t \to a_t \to s_{t +1}$, where the action $a_t$ is associated with LLM-generated item suggestions only on condition of conversation history in the past. The evaluation across synthetic and real-world datasets (e.g.REDIAL, OpenDialKG) demonstrates that our RLHF-fine-tuned models can perform better in terms of top-$k$ recommendation accuracy, coherence, and user satisfaction compared to (arrow-zero-cmwrquca-teja-falset ensuite 2Round group-deca States penalty give up This paper shows that implicit signal alignment can be efficient in achieving scalable and user-adaptive design of CRS.Open AccessRLHF Fine-Tuning of LLMs for Alignment with Implicit User Feedback in Conversational Recommenders
Conversational recommender systems (CRS) based on Large Language Models (LLMs) need to constantly be aligned to the user preferences to provide satisfying and context-relevant item recommendations. The traditional supervised fine-tuning cannot capture the implicit feedback signal, e.g., dwell time, sentiment polarity, or engagement patterns. In this paper, we share a fine-tuning solution using human feedback reinforcement learning (RLHF) to maximize implied user feedback (IUF) in a multi-turn recommendation context. We specify a reward model $R_φ$ learnt on weakly-labelled engagement information and maximize user-centric utility by optimizing the foundational LLM M_θ through a proximal policy optimization (PPO) approach. The architecture models conversational state transitions $s_t \to a_t \to s_{t +1}$, where the action $a_t$ is associated with LLM-generated item suggestions only on condition of conversation history in the past. The evaluation across synthetic and real-world datasets (e.g.REDIAL, OpenDialKG) demonstrates that our RLHF-fine-tuned models can perform better in terms of top-$k$ recommendation accuracy, coherence, and user satisfaction compared to (arrow-zero-cmwrquca-teja-falset ensuite 2Round group-deca States penalty give up This paper shows that implicit signal alignment can be efficient in achieving scalable and user-adaptive design of CRS.