Jiaqing Liang, who uses lsdefine and LJQ online, is an associate professor at Fudan University’s School of Data Science. His research combines knowledge graphs with deep learning and agent systems. He is a co-author of the GenericAgent technical report and leads open-source work on compact agent and reinforcement-learning implementations.[1][2][5][6]
Start with GenericAgent technical report →
Education
Links
Institutional profile Website
GitHub Associate Professor, School of Data Science · Fudan University
Career
Fudan’s current faculty directory identifies Liang as an associate professor and technical lead of the KnowledgeWorks knowledge-graph platform. It records his Fudan doctorate in 2020. His personal and lab pages list computer-science study from 2011 to 2015 and doctoral study from 2015 to 2020, followed by postdoctoral research from 2020 to 2022. The postdoctoral period is employment and research experience, not another degree.[1][2][3]
Contributions
Keep useful information inside the context budget
GenericAgent studies how an agent can retain decision-relevant information during long tasks. Its report describes a small tool set, hierarchical memory loaded on demand, reusable procedures distilled from verified task trajectories, and context compression. Those mechanisms address two losses: crowded conversations pushing out useful facts, and successful experience disappearing between episodes. Liang is the first listed author of a multi-author technical report; the repository and evaluation work are team contributions.[4][5]
Make learning algorithms inspectable
Simple GRPO presents a compact implementation for experimenting with group-relative policy optimization, a method that compares groups of generated answers when updating a model. Its README identifies Liang and Yanghua Xiao as project leads and names a broader development team. Together with GenericAgent, it offers code that exposes the main control and learning steps rather than presenting only an end-user interface.[6]
Tools, writing & talks
Writing, research & talks
Start here
GenericAgent technical reportarxiv.orgA multi-author account of context-information density and reusable experience in long-running agents.
Read the system components before the benchmark claims, then inspect the matching repository. Familiarity with tool-using language-model agents helps separate memory design from model capability.
First listed co-author[5]
Tools
- GenericAgentgithub.com
An agent framework combining a small tool interface, on-demand memory, reusable procedures, and context compression.
Repository owner and technical-report co-author[2][4][5]Original source - Simple GRPOgithub.com
A compact implementation for studying group-relative policy optimization in language models.
Co-lead with Yanghua Xiao[6]