Researcher @ BAAI
My current research focuses on autonomous agents dedicated to auto-research: AI systems that can actively formulate research questions, decompose complex tasks, search for evidence, reason over heterogeneous knowledge, and synthesize grounded findings.
This direction grows out of my work on memory-augmented retrieval and long-context understanding, such as MemoRAG, and extends toward agents that can conduct open-ended web investigation (WebThinker), maintain executive memory for reasoning (MemoBrain), and interact with scientific literature through agentic data interfaces (DeepXiv-SDK).
I am also enthusiastic about building systems that turn research into practical tools, including general agent platforms and domain-specific demos for scientific discovery, AI talent mining, and education planning. Please contact me if you would like invitation codes for demo access.
|
|
Agent
2025-Present
20
AI Agents represent the next evolution of LLMs, moving from passive conversation to active task execution.Publications
|
|
|
|
Retrieval-Augmentation Generation
2022-Present
12
Retrieval-Augmented Generation (RAG) is a method that first retrieves relevant information from an external knowledge source and then combines it with the model’s input to generate more accurate and informative responses.Publications
|
|
|
|
Conversational Search
2021-Present
9
Conversational search is an interactive search paradigm where users and systems engage in a dialogue, allowing queries, clarifications, and refinements across multiple turns to iteratively reach more accurate and context-aware results.Publications
|
|
|
|
Others
2020-Present
10
Dialogue System, QA System, Ranking, Retrieval, Theory, etc.Publications
|
Jun 2023 - Oct 2023, Beijing, China
Elensdata is a start-up company which offers high-calibre data science/AI solutions that help real businesses, in media, finance, etc.
I enjoy building systems that turn research ideas into practical tools. Demo access may require an invitation code; please contact me if you would like to try them.
This project focuses on exploring techniques to expand the knowledge scale and memory scale at the input stage. The goal is to overcome the limitations of current LLMs in complex knowledge reasoning, knowledge memorization, and global knowledge understanding. This will be achieved by constructing a hierarchical memory mechanism that enables the scaling, memorization, and dynamic, coordinated retrieval of multi-source, heterogeneous knowledge.
Supported by the Beijing Postdoctoral Research Foundation, this project studies adaptive information agents for deep search in complex scenarios. It explores self-evolution without parameter updates, reinforcement learning with rewards for correctness, information gain, and efficiency, and reasoning-tree-based data synthesis for verifiable agent training and evaluation.
MemoRAG is a next-generation retrieval-augmented generation system with long-term memory, enabling superior context-aware information retrieval and enhanced performance on complex tasks where traditional RAG systems struggle.
Informatica is a comprehensive collection of systematic research projects focused on deep research systems. Our mission is to provide open-source, scalable frameworks, datasets, data synthesis methods, models, and demonstrations.
DeepXiv-SDK is an agent-first interface for scientific papers, supporting paper search, progressive reading, and research workflows through CLI, Python SDK, MCP, and agent integrations. It is designed to help AI agents discover, inspect, and reason over scientific literature more effectively.
