Jiang Daxin is the chief executive of StepFun. His research career spans Microsoft’s language-technology work and foundation models, with publications on search, question answering and tools for building language-processing systems.[1][2]
Start with NeuronBlocks →
Career
Jiang’s published author biography records a computer-science doctorate from the University at Buffalo in 2005, an assistant professorship at Nanyang Technological University from 2005 to 2006 and a lead-researcher role at Microsoft Research Asia from 2007 to 2011. He later led natural-language research and development as Microsoft STCA’s partner chief scientist. StepFun’s leadership announcement identifies him as its CEO.[2][1]
Contributions
Build language models from reusable components
Jiang coauthored NeuronBlocks, a toolkit intended to reduce the coding needed to train language-processing models. Its Block Zoo contains reusable neural-network components, while its Model Zoo provides task configurations. A developer can reuse a model configuration or connect blocks to construct a different architecture. The project covers training and inference for tasks such as classification and question answering; it predates today’s general-purpose chat models.[3][4]
Separate a question’s topic from the reasoning it requires
In DisentangledQA, Jiang and coauthors distinguish retrieving material about a topic from finding the attribute needed to answer a question. Their example asks whether one geographic region could fit inside another: retrieving descriptions of both places is insufficient without their areas and a size comparison. The method separates these elements to guide evidence retrieval and answer inference. Its reported gains concern the paper’s evaluated datasets.[2]
Tools, writing & talks
Tools
Start here
NeuronBlocks
github.comA toolkit that assembles language-processing models from reusable blocks and task configurations.
Read the Block Zoo and Model Zoo examples before following the setup. Basic neural-network and Python knowledge helps; this is a historical toolkit with older dependencies, so inspect compatibility before running it.