Algorithm-System Co-Design for Efficient and IP-Protected LLMs: From Model Optimisation to Cluster Deployment

As large language models (LLMs) continue to grow in size, complexity, and commercial value, concerns around computational efficiency and intellectual property (IP) protection have become increasingly prominent. Dr He Xin’s project explores a novel algorithm-system co-design approach to address these issues. The research aims to improve the performance and deployability of LLMs while ensuring their originality and ownership can be reliably protected, even in resource-limited deployment environments. This work seeks to strengthen trust in LLM-based systems and support responsible, secure AI development across both academic and industrial settings.

More is Less: Leveraging Large Language Models to Compress Vision Models

As vision models become essential in daily life, their expanding use on resource-limited devices creates an urgent need for more efficient solutions. Dr He Yang proposes bridging large language models with small vision models, offering three strategies to enhance efficiency. His method uses language models to create more effective training datasets, implements a streamlined training process that modifies far fewer parameters, and develops smarter pruning decisions for faster model inference. This comprehensive approach supports Singapore's Smart Nation initiatives by enabling more efficient AI systems that require less storage, learn more efficiently, and respond more quickly across various sectors, including transportation, healthcare, and manufacturing.

Nonparametric Distributional Black-box Optimisation for Diffusion-model Target Generation

Guided diffusion-model generation presents a promising avenue for customising the generation process to satisfy the users’ requirements and preferences. However, both theoretical foundations and practical algorithms for query-efficient black-box target generation remain largely unexplored. Dr Lyu Yueming's project aims to develop a nonparametric distributional black-box optimisation framework to analyse and design diffusion-based target generation in a theoretically rigorous manner. This framework will establish a solid foundation for advancing both target generation and black-box optimisation, potentially benefiting a wide range of downstream tasks.

Multi-Modal Composite Material Design: Benchmark, Alignment, and Transfer

Conventional material design remains an extremely challenging and slow process, often taking a period of 20 to 30 years from starting of the design to final deployment. Artificial Intelligence (AI) methods have the potential to drastically shorten this cycle to a few years - or even months. Dr Qian Hangwei's proposal seeks to expedite the research on composite material design by developing a comprehensive multi-modal learning framework. The proposed research brings together interdisciplinary ideas from material science, machine learning, and explainable AI, and will establish a strong foundation for AI in Scientific Discovery.