The Hong Kong Polytechnic University (PolyU) has unveiled significant advances in Generative AI (GenAI) research, introducing a novel collaborative training approach called Co-GenAI. This innovative model aims to decentralize AI training, reducing costs and making high-level AI research more accessible to institutions worldwide. By lowering the resource requirements traditionally associated with training large AI models, PolyU is reshaping the landscape of AI research.
Traditionally, training foundation models has been prohibitively expensive, often requiring millions of hours of graphics processing unit (GPU) time. This has limited participation in advanced AI research to a select few organizations. The PAAI team has identified three major barriers: the high computational costs of model training, the siloing of data due to privacy and copyright issues, and the static nature of existing models that inhibit rapid iteration. To address these challenges, the team has developed a framework for ultra-low-resource training and decentralized model fusion.
Revolutionizing AI Training
PolyU is making strides as the first academic institution to open-source an end-to-end FP8 low-bit training solution, encompassing both continual pre-training (CPT) and post-training stages. This groundbreaking approach allows for over 20% faster training compared to BF16 precision, while also decreasing peak memory usage by more than 10%. Such advancements dramatically reduce training overheads without compromising performance.
The new framework integrates multiple training methodologies, including supervised fine-tuning (SFT) and reinforcement learning (RL), to achieve BF16 quality while minimizing training time and memory footprint. The team is now exploring even lower-cost FP4 precision training, with promising initial results documented in their research. In medical applications, the models trained using these pipelines have outperformed peer models in diagnosis and reasoning, showcasing their potential in critical fields.
The PolyU InfiFusion model fusion represents a significant achievement in AI research. By utilizing only hundreds of GPU hours, the team has successfully merged four state-of-the-art models that would typically require 1 to 2 million GPU hours to train from scratch. This breakthrough not only avoids substantial financial investments but also delivers fused models that surpass the original benchmarks.
Prof. YANG Hongxia, Executive Director of PAAI, remarked, “Ultra-low-resource foundation model training, combined with efficient model fusion, enables academic researchers worldwide to advance GenAI research through collaborative innovation.” The team’s work has been validated through rigorous mathematical derivation, leading to the introduction of the “Model Merging Scaling Law,” which suggests a new path toward achieving artificial general intelligence (AGI).
Collaborative Applications and Future Directions
PolyU’s PAAI is also collaborating with esteemed institutions such as Huashan Hospital affiliated with Fudan University and the Sun Yat-sen University Cancer Center to enhance medical foundations and cancer AI models. These collaborations aim to integrate high-quality, domain-specific data, allowing for personalized treatment options and AI-driven radiotherapy solutions.
Additionally, PAAI has launched a cutting-edge agentic AI application designed to assist in academic research. This tool serves as a graduate-level academic paper writer that supports a multimodal patent-search engine, streamlining the research and manuscript drafting process.
Prof. Christopher CHAO, Senior Vice President of Research and Innovation at PolyU, stated, “AI is a key driver in accelerating the development of new quality productive forces. The newly established PAAI is dedicated to expediting AI integration across key sectors and developing domain-specific models for diverse industries.”
The project, led by Prof. YANG, is supported by various funding initiatives, including the Theme-based Research Scheme 2025/26 under the Research Grants Council and the Artificial Intelligence Subsidy Scheme under Cyberport. This initiative marks a significant advancement for Hong Kong in the realm of global AI innovation, promoting the democratization and industrial implementation of AI technologies.
