Jiaqing Liang
Young Associate Researcher, School of Big Data, Fudan University
Dr. Jiaqing Liang, a young associate researcher at the School of Big Data, Fudan University, focuses on cognitive intelligence research in knowledge graph and big models. He has published more than 50 papers in top conferences and journals such as TKDE, AAAI, etc. The knowledge graph and big model application platform developed by him has been invoked more than 1.7 billion times, and he owns nearly 20 patents. He was awarded the first place in Information Extraction Competition of Language and Intelligent Technology Competition. The Chinese big model CuteGPT developed by CuteGPT has been applied in many companies. He has won many honors such as Gold Medal in ACM-ICPC Regional Competition, TopCoder Open Top 150 in the world, and Scientific and Technological Progress Award in Wu Wenjun Artificial Intelligence Award.
Topic
Domain-oriented big model thinking skills
In recent years, big models have made significant progress in the field of general intelligence, but the challenge of domain-specific thinking ability remains. Represented by o1 and r1 class models, their powerful reasoning ability provides new ideas for domain applications. However, multi-constraint adherence under complex instructions, accurate imitation of domain thinking, enhanced learning trial-and-error and tool utilization, and complex thinking processes supported by hybrid tools are still the core pain points in the application of large models. There is a need to focus on these bottlenecks and explore how to improve the performance of big models in domain tasks. For example, in complex instruction scenarios, models need to accurately understand and satisfy multiple constraints instead of simply generating answers; in domain thinking imitation, models are required to capture industry logic and expert experience in depth; to enhance the ability to solve dynamic problems through reinforcement learning trial and error, combined with external tools; and ultimately, to realize multistep reasoning and resource integration in complex thinking processes supported by hybrid tools. o1 Class models have potential in these potentials are promising, but also expose trade-offs between generality and specialization. We aim to push big models to better serve domain applications and unleash their thinking potential in complex scenarios by optimizing training paradigms and tool integration. Big Model: INTRODUCTION: Limitations of big models in general-purpose domains vs. reasoning potential of o1-class models. Core problem: Difficulty in following complex instructions, weak imitation of domain thinking, and rigid tool utilization. Research Focus: Assessment and enhancement of complex instruction following capabilities of reasoning models Agent reasoning framework incorporating domain thinking Trial-and-error based training for flexible tool application Experimental thinking process incorporating tools Outlook: future development of domain use cases and large models.
