The Hong Kong Polytechnic University (PolyU) Academy for Artificial Intelligence (PAAI) has announced achieving several milestones in Generative AI (GenAI) research. The PAAI team is pushing the boundaries of AI with a novel collaborative GenAI paradigm known as Co-GenAI, which has the potential to transform frontier model training from a centralised, monolithic approach into a decentralised one. Significantly lowering training resource requirements, protecting data privacy and removing resource barriers such as graphics processing unit (GPU) monopolies paves the way for a more inclusive and accessible environment for global institutions to participate in AI research.
Advances in GenAI research are presently constrained by three major barriers: training foundation models being so computationally prohibitive that only a few organisations can afford it, effectively excluding global academia from frontier model development; domain knowledge and data remaining siloed due to privacy and copyright concerns, particularly for sensitive information in healthcare and finance; and foundation models being static and unable to evolve with emerging knowledge, while retraining each frontier model ab initio consumes an enormous amount of resources and makes rapid iteration impossible. To tackle these challenges, the PAAI team has developed a novel model training framework that enables ultra-low-resource training and decentralised model fusion. The framework is theoretically grounded and has been validated through extensive real-world applications.
PolyU is the first academic institution to open-source an end-to-end FP8 low-bit training solution that covers both continual pre-training (CPT) and post-training stages. This approach will set a new standard for training models with FP8 ultra-low resources while maintaining BF16 precision, in turn revolutionising the practice of model training and positioning PolyU among the few institutions worldwide to master this advanced training technique. Compared with BF16, FP8 delivers over 20% faster training, reduces peak memory by over 10% and dramatically lowers training overheads while maintaining performance. The pipeline integrates CPT, supervised fine-tuning (SFT) and reinforcement learning (RL) to achieve BF16 quality while shortening training time and reducing memory footprint. The team has begun exploring even lower-cost FP4 precision training, with initial results reported in academic publications1. In medical applications, the models trained by these pipelines outperform all peer models on diagnosis and reasoning across all key areas2. In research agent application, the models also demonstrate exceptional performance in complex task handling, generalisation and report quality3.
Until now, foundation model training has followed scaling laws: more parameters yield broader knowledge and stronger performance. However, centralised training typically requires millions of GPU hours—a resource available to only a few organisations. The PolyU InfiFusion model fusion achieves a key milestone in model fusion research: it uses only hundreds of GPU hours to fuse large models that would otherwise require 1–2 million GPU hours to train from scratch. The team has merged four state-of-the-art models in 160 GPU hours4-5, avoiding million-scale training budgets while delivering fused models that significantly outperform the originals across multiple key benchmarks.
The team has published the first theoretical validation of model fusion—a concept championed by Thinking Machines Lab. Through rigorous mathematical derivation, they proposed the "Model Merging Scaling Law," suggesting there is another viable pathway to artificial general intelligence (AGI)6. Prof. YANG Hongxia, Executive Director of PolyU PAAI, Associate Dean (Global Engagement) of the Faculty of Computer and Mathematical Sciences, and Professor of the Department of Computing, stated, "Ultra-low-resource foundation model training, combined with efficient model fusion, enables academic researchers worldwide to advance GenAI research through collaborative innovation."
The team has also demonstrated the potential of its training pipelines through applications across specific domains, including state-of-the-art medical foundation and cancer AI models that achieve best-in-class performance. With the integration of high-quality domain-specific data, these models can adapt to medical devices for different scenarios, including personalised treatment and AI-based radiotherapy for oncology. In this context, the team is now collaborating with Huashan Hospital affiliated to Fudan University, Sun Yat-sen University Cancer Center, Shandong Cancer Hospital and Queen Elizabeth Hospital in Hong Kong. PAAI has also introduced a leading agentic AI application in deep search and academic paper assistance—a graduate-level academic paper writer with agentic capability that supports a multimodal patent-search engine for end-to-end research and manuscript drafting.
Prof. Christopher CHAO, Senior Vice President (Research and Innovation) of 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. These initiatives will not only solidify the leading position of PolyU in related fields, but also help position Hong Kong as a global hub for GenAI."
The research project led by Prof. Yang Hongxia is supported and funded by the Theme-based Research Scheme 2025/26 under the Research Grants Council, the Research, Academic and Industry Sectors One-plus Scheme under the Innovation and Technology Commission of the HKSAR Government, and the Artificial Intelligence Subsidy Scheme under Cyberport. It marks a significant step forward for Hong Kong in global AI innovation and accelerating the democratisation and industrial implementation of AI technology.
1InfiR2: A Comprehensive FP8 Training Recipe for Reasoning-Enhanced Language Models, https://arxiv.org/html/2509.22536v3
2InfiMed: Low-Resource Medical MLLMs with Advancing Understanding and Reasoning, https://arxiv.org/html/2505.23867
3InfiAgent: Self-Evolving Pyramid Agent Framework for Infinite Scenarios, https://arxiv.org/html/2509.22502
4InfiGFusion: Graph-on-Logits Distillation via Efficient Gromov-Wasserstein for Model Fusion, https://arxiv.org/html/2505.13893
5InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models, https://arxiv.org/abs/2505.13878
6Model Merging Scaling Laws in Large Language Models, https://arxiv.org/html/2509.24244
