In this series, the Center for AI Research (CAIR) introduces a selection of recent scientific publications by researchers at the Center. These works span multiple research directions in artificial intelligence (AI), from large language models, machine unlearning, recommender systems, and generative models to air traffic management, computational mechanics, and engineering materials.

Each publication offers a specific glimpse into the research activities at CAIR. When viewed together, these works show how CAIR research groups are developing specialized research directions that both address core challenges in modern AI and connect with the practical requirements of different fields.

See below each image to read more about the publication details, research projects and key significance of each work.

– Title: UniFLE: Uniform Fusion of Multiple LoRA Experts for Backdoor Defense in Large Language Models
– Publisher: Institute of Electrical and Electronics Engineers (IEEE)
– Journal name: IEEE Transactions on Dependable and Secure Computing
– Ranking: Q1
– Authors: Shuai Zhao; Qika Lin; Yanhao Jia; Xinyi Wu; Yuwen Li; Luu Anh Tuan (CAIR)
– Publication page: https://ieeexplore.ieee.org/document/11407462

This work addresses an important safety issue in the fine-tuning of large language models: backdoors may remain active even after the model is adapted using parameter-efficient methods such as LoRA. To mitigate this risk, the research team proposes UniFLE, a method that integrates multiple LoRA experts into the MLP blocks to expand the updatable feature space and decouple backdoor-related features from the model’s internal representations. The method also introduces a diversity regularization mechanism to encourage different experts to learn more diverse features. Experiments across multiple tasks, LLMs, and backdoor attack methods show that UniFLE can improve backdoor defense while preserving the model’s task performance.

– Title: Cross-domain fatigue assessment of air traffic controllers using multimodal monitoring and transfer learning
– Publisher: Elsevier BV
– Journal name: Aerospace Science and Technology
– Ranking: Q1
– Authors: Haoran Shen, Yanjun Wang, Yi Hu, Vu Duong (CAIR)
– Publication page: https://www.sciencedirect.com/…/pii/S1270963826004645…

This work studies fatigue assessment for air traffic controllers, a critical factor in aviation safety. Unlike many previous studies that treat fatigue as a general condition, this research considers the differences between operational positions, such as terminal/approach controllers and en-route controllers. The study proposes a position-independent fatigue assessment framework that integrates psychomotor vigilance tests, short-term memory tasks, and continuous facial–acoustic monitoring. A hybrid architecture combining multi-stream LSTM and Domain-Adversarial Neural Network is used to learn domain-invariant features across heterogeneous operational settings. The results show that the model can support more accurate and equitable fatigue estimation across controller positions, highlighting the importance of multimodal monitoring for air traffic safety management.

– Title: Distributional Surgery for Language Model Activations
– Publisher: Association for Computational Linguistics
– Conference name: EMNLP 2025 – 2025 Conference on Empirical Methods in Natural Language Processing
– Ranking: CORE A*
– Authors: Nguyen, Bao; Nguyen, Duy; Nguyen, Binh (CAIR); Nguyen, Viet Anh
– Publication page: https://www.scopus.com/pages/publications/105028983222…

This work proposes a method to detect and mitigate undesirable generations from language models, including harmful or toxic outputs. Instead of controlling only the final text output, the method intervenes in the model’s internal activations. The approach consists of two stages: first, an ensemble of layerwise classifiers is trained to detect undesirable content from activations; then, when risky content is detected, layerwise distributional steering policies are applied to transform attention heads. These policies are computed through semidefinite programming, aiming to minimally perturb the attention distribution while providing probabilistic guarantees about the effectiveness of the intervention. Experiments across multiple language models and datasets show that the method outperforms baseline approaches in reducing undesirable outputs, contributing to research on safer and more controllable LLMs.

– Title: A Survey of Machine Unlearning
– Publisher: Association for Computing Machinery (ACM)
– Journal name: ACM Transactions on Intelligent Systems and Technology
– Ranking: Q1
– Authors: Thanh Tam Nguyen, Thanh Trung Huynh (CAIR), Zhao Ren, Phi Le Nguyen, Alan Wee-Chung Liew, Hongzhi Yin, Quoc Viet Hung Nguyen
– Publication page: https://dl.acm.org/doi/full/10.1145/3749987

This survey systematizes an increasingly important research direction in AI: machine unlearning, which aims to make machine learning models “forget” specific data used during training. The problem is closely linked to privacy protection and the “right to be forgotten,” as deleting data from databases is not sufficient if trained models still retain information about that data. The paper reviews key concepts, formulations, design criteria, removal requests, algorithms, and applications of machine unlearning. It also highlights remaining gaps, including the lack of shared frameworks and resources for evaluation. This work serves as a valuable foundation for researchers and practitioners interested in privacy-preserving AI, trustworthy machine learning, and controllable data governance in AI systems.

– Title: PromptHG: Prompt-Enhanced Heterogeneous Graph for Personalized News Recommendation
– Publisher: Springer Nature Switzerland
– Conference name: 48th European Conference on Information Retrieval, ECIR 2026
– Ranking: CORE A
– Authors: Hai-Dang Kieu, Delvin Ce Zhang, Minh-Duc Nguyen, Qiang Wu, Min Xu & Dung D. Le (CAIR)
– Publication page: https://link.springer.com/…/10…/978-3-032-21289-4_35…

This work studies personalized news recommendation in the context of large language models, which offer new opportunities for understanding content and user preferences. However, many LLM-based methods focus mainly on semantic enrichment while underusing structural signals such as entity-level relations. To address this limitation, the authors propose PromptHG, a framework that combines LLM prompting with heterogeneous graph learning. The model uses LLMs to synthesize entities from news titles and links articles through shared entity nodes, uncovering relationships beyond conventional click-based signals. Experiments on two benchmark datasets show that PromptHG consistently improves recommendation performance over strong baselines. The work contributes to more context-aware and structurally informed personalized recommendation systems.

– Title: LayGenID: Layout-preserving multi-conditional diffusion transformer with LoRA fusion for interior design generation
– Publisher: Elsevier BV
– Journal name: Expert Systems with Applications
– Ranking: Q1
– Authors: Duong Q. Nguyen, Kim Q. Tran, H. Nguyen-Xuan (CAIR)
– Publication page: https://www.sciencedirect.com/…/pii/S0957417426012650…

This work addresses AI-based interior design generation while preserving fixed architectural elements such as walls, doors, windows, and spatial layouts. This is an important challenge because many generative models can produce visually appealing images but distort the original architectural structure, making the outputs less reliable for practical design use. The research team proposes LayGenID, a layout-aware generation framework based on a Diffusion Transformer. It incorporates semantic segmentation, depth maps, line maps, and text prompts as multiple conditions. These conditions are injected through LoRA modules and combined using a weighted fusion mechanism. The results suggest that LayGenID improves structural consistency under multi-condition control, offering a more controllable and reliable direction for generative AI in interior visualization.

– Title: Normalized energy-based physics-informed neural network: Theory and applications to solid mechanics problems
– Publisher: Elsevier BV
– Journal name: Finite Elements in Analysis and Design
– Ranking: Q1
– Authors: Thang Le-Duc, H. Nguyen-Xuan (CAIR), Jaehong Lee
– Publication page: https://www.sciencedirect.com/…/pii/S0168874X26000132…

This work develops an improved approach for energy-based physics-informed neural networks, a class of machine learning models used to solve partial differential equations in solid mechanics. The study identifies key limitations of traditional EPINNs, especially for non-trivial problems with small-intensity solutions and coupled PDE systems, where predictions may become biased toward larger-intensity responses. To address these issues, the authors propose normalized EPINN, which uses a scaling vector to calibrate network outputs and improve solution stability. The method is validated on a range of solid mechanics problems, including 1D beam bending, 2D and 3D linear elasticity, and a real-world 72-bar truss structure. The results show that nEPINN can improve prediction accuracy and stability, contributing to more reliable AI tools for scientific computing and engineering simulation.

– Title: Experimental and multi-scale modelling investigation of printable, low-cement engineered cementitious composites with different Polyether fiber contents
– Publisher: Elsevier BV
– Journal name: Construction and Building Materials
– Ranking: Q1
– Authors: Vuong Nguyen-Van, Cheah Chun Jie, Junying Lao, Zhao Huanyu, H. Nguyen-Xuan (CAIR), Hongjian Du, Shunzhi Qian
– Publication page: https://www.sciencedirect.com/…/pii/S0950061826005295…

This work investigates 3D-printable, low-cement engineered cementitious composites with different PE fiber contents, aiming to improve both mechanical performance and environmental sustainability. The study combines material experiments, multi-scale micromechanical modelling, and numerical simulations to assess printability, strength, ductility, and fracture behavior. The results show that appropriate fiber inclusion can improve buildability, tensile deformation capacity, and the multiple micro-cracking behavior characteristic of ECC materials. In particular, the 1.5% fiber mixture offers a strong balance between strength and ductility, while the 2.0% mixture achieves higher tensile strain and compressive strength. Life-cycle assessment also suggests that although PE fibers slightly increase global warming potential, strength-normalized environmental efficiency can be improved. The study provides an integrated experimental–numerical framework for designing more sustainable and high-performance 3D-printable structural materials.

– Title: Text2Traffic: Retrieval-Enhanced In-Context Learning for Complex Air Traffic Scenario Generation
– Publisher: American Institute of Aeronautics and Astronautics (AIAA)
– Journal name: Journal of Air Transportation
– Ranking: Q2
– Authors: Yash Guleria, Duc-Thinh Pham (CAIR), Ashton Low Kin Yun, Thaivalappil N. M. Nadirsha, Katherine Fennedy, Chunyao Ma and Sameer Alam
– Publication page: https://arc.aiaa.org/doi/10.2514/1.D0566

This paper introduces Text2Traffic, a method for generating complex air traffic scenarios from natural language descriptions. Instead of modifying historical data, creating randomized traffic, or manually designing scenarios, Text2Traffic combines large language models with retrieval-augmented generation to rapidly produce customizable simulation scenarios. Users can adjust traffic density, routing, separation, aircraft types, and operational constraints through natural language prompts. Experimental evaluations across multiple models show that the method can generate syntactically accurate and semantically meaningful traffic scenarios. The study offers a scalable and intuitive framework for air traffic management research, simulation, training, and future operational planning.