MemT-1.0: Towards Memory-Native Evolution
ArXiv 2026. Team contribution (led the RL stage). 100K+ downloads within one week of release.
memtensor-team
๐คHaotian Chen
(he/him)
Research Assistant Professor ยท LLM Post-Training / AI4AI
Biography
I am a Research Assistant Professor at the School of Artificial Intelligence, Shanghai Jiao Tong University, where I work with Prof. Junchi Yan at RethinkLab. I study how to build AI systems that can automate long-horizon, effort-intensive, and creativity-demanding tasks such as research, engineering, and development. My current work focuses on LLM post-training for autonomous agents, agent memory and continual learning, and AI4Research / AI4AI. Before joining SJTU, I received my PhD in Data Science from Fudan University, advised by Prof. Xiangdong Zhou, and completed postdoctoral research at Tsinghua University (THUNLP), working with Prof. Zhiyuan Liu and Prof. Maosong Sun. I was also a research intern at the Machine Learning Research Group of Microsoft Research Asia, mentored by Xiao Yang, and at the Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University, working with Prof. Yang Yu.
Education
PhD in Data Science
2018-09-01
2024-06-30
Fudan University
BEng in EE
2014-09-01
2018-06-30
Dalian University of Technology
Interests
Autonomous Agents
LLM Post-Training & Agentic RL
Agent Memory & Continual Learning
AI4Research / AI4AI
Tool Learning
Experience
Research Assistant Professor
Shanghai Jiao Tong University
School of Artificial Intelligence, RethinkLab. I work with Prof. Junchi Yan on autonomous agents, LLM post-training, and AI4Research. PI of an NSFC Young Scientists Fund (Category C) project.
Postdoctoral Researcher
Tsinghua University
Department of Computer Science and Technology, THUNLP. I worked with Prof. Zhiyuan Liu and Prof. Maosong Sun on autonomous agents, LLMs, and agentic reinforcement learning for tool use. Honors: Tsinghua “Shuimu Scholar” Program; China Postdoctoral Science Foundation General Grant; National Postdoctoral Researcher Program (Category C).
Research Intern
Microsoft Research Asia
Machine Learning Research Group. I worked with Xiao Yang and Jiang Bian on autonomous agents and LLMs, and served as a founding core member of RD-Agent, featured in the Microsoft Build 2025 keynote. Honor: Microsoft “Star of Tomorrow”.
Research Intern
Tsinghua University (IIIS)
Institute for Interdisciplinary Information Sciences. I worked with Prof. Yang Yu on LLM safety auditing and AI decision-bias analysis.
Education
PhD in Data Science
Fudan University
School of Computer Science and Technology. Advisor: Prof. Xiangdong Zhou. I worked on NLP, data mining, and information extraction, and received honors including Outstanding Student, Outstanding Graduate, and Academic Excellence Scholarship.
BEng in EE
Dalian University of Technology
My undergraduate studies focused on underwater robotics, sensors, robot configurations, and mechanics, with research in image recognition. Advisors: Prof. Qianjin Yue and Prof. Mingfeng He. I ranked 1/74 in the program and graduated as a Liaoning Province Outstanding Graduate.
๐ฌ Research
I study how to build AI systems that can automate long-horizon work in research, engineering, and development. A central goal of my work is to make data-driven scientific research more scalable, reliable, and productive. I currently focus on four closely connected directions:
๐งช AI4Research / AI4AI โ I build AI research assistants for literature retrieval, experiment design, data analysis, and hypothesis generation and verification, and study how AI systems can improve the way AI itself is developed.
๐ค Autonomous Agents & LLM Post-Training โ I develop agent models and systems with environment perception, memory, planning, and tool-use capabilities, and post-train on-device LLMs (1Bโ8B) with agentic reinforcement learning so that small models solve long-horizon tasks that previously required closed-source frontier models.
๐ง Agent Memory & Continual Learning โ I work on memory-native foundation models and learnable dynamic memory operations that let agents accumulate experience, reuse it across tasks, and keep learning and evolving.
๐๏ธ Foundation Models โ I explore how to train and adapt general-purpose models for language understanding, knowledge representation, multimodal fusion, and downstream generalization.
๐ Featured Publications
โข
โข
1 min read
AgentCPM-Explore: Realizing Long-Horizon Deep Exploration for Edge-Scale Agents
ArXiv 2026. First author. Open-source 4B agent model achieving SOTA on GAIA & HLE, surpassing GPT-5 and Claude-4.5-Sonnet. 830+ GitHub Stars.
AgentCPM-GUI: Building Mobile-Use Agents with Reinforcement Fine-Tuning
EMNLP 2025 Demo. GUI agents with reinforcement fine-tuning. 1,400+ GitHub Stars.
zhong-zhang
MiniCPM4: Ultra-Efficient LLMs on End Devices
ArXiv 2025. Team contribution (led MCP agent capabilities). 10,700+ GitHub Stars.
minicpm-team
RD2Bench: Toward Data-Centric Automatic R&D
ICLR 2024 Workshop. First author. RD2Bench / RD-Agent for data-centric automatic R&D. Featured in the Microsoft Build 2025 keynote; 15,100+ GitHub Stars.
๐ All Publications
(2026).
MemT-1.0: Towards Memory-Native Evolution.
ArXiv 2026.
(2026).
AtomMem: Learnable Dynamic Agentic Memory with Atomic Memory Operation.
In EMNLP 2026.
(2026).
AgentProcessBench: Diagnosing Step-Level Process Quality in Tool-Using Agents.
In KDD 2026.
(2026).
Test-Time Deep Thinking to Explore Implicit Rules.
In KDD 2026.
(2026).
AgentCPM-Explore: Realizing Long-Horizon Deep Exploration for Edge-Scale Agents.
ArXiv 2026.
(2025).
AgentCPM-GUI: Building Mobile-Use Agents with Reinforcement Fine-Tuning.
In EMNLP 2025 Demo.
(2025).
MiniCPM4: Ultra-Efficient LLMs on End Devices.
ArXiv 2025.
(2025).
ToLeaP: Rethinking Development of Tool Learning with Large Language Models.
ArXiv 2025.
(2024).
Collaborative Evolving Strategy for Automatic Data-Centric Development.
ArXiv 2024.
(2024).
OODREB: Benchmarking State-of-the-Art Methods for Out-Of-Distribution Generalization on Relation Extraction.
In WWW 2024.
(2024).
RD2Bench: Toward Data-Centric Automatic R&D.
ICLR 2024 Workshop.
(2024).
Rethinking the Development of Large Language Models from the Causal Perspective: A Legal Text Prediction Case Study.
In AAAI 2024.
(2023).
SALAS: Supervised Aspect Learning Improves Abstractive Multi-Document Summarization through Aspect Information Loss.
In ECML-PKDD 2023.
๐ป Open-Source Projects
All my open-source contributions revolve around the central theme of LLM-powered agents, forming a cohesive ecosystem: Data (AgentCPM-GUI) โ Algorithm (AgentRL, AgentCPM-Explore, MiniCPM4-MCP, MiniCPM5) โ Memory (MemT-1.0, AtomMem) โ Execution (RD-Agent) โ Evaluation (ToLeaP). Together these projects have gathered 40K+ GitHub stars and 7M+ model downloads.
MemT-1.0
Technical Advisor ยท RL Lead โ memory-native foundation models (Nano 36B-A3B / Flash 288B-A17B); 100K+ downloads in the first week
MiniCPM5 (1B / 2B)
Agent Capability Lead โ open-source on-device LLMs (1B / 2B) whose agentic capabilities I built end to end; 6.9M+ downloads
RD-Agent
Core Founding Member & Primary Coder โ automated R&D agent for data science and finance; 15,100+ stars, Microsoft Build 2025 keynote
AgentCPM-Explore
Project Lead โ open-source 4B agent model for long-horizon deep exploration; tops GAIA and HLE among open models of its size
AgentRL
Project Lead โ fully asynchronous agent RL training infrastructure for the AgentCPM model family: 100+ tools, 20+ benchmarks, full-cycle visualization
MiniCPM4-MCP
Project Lead โ edge-scale (8B) agent LLM mastering MCP tools; surpasses GPT-4o and Qwen3-8B; 5,500+ downloads
AgentCPM-GUI
Training Data Lead โ multimodal LLM-based GUI agent for mobile and desktop; 1,400+ stars, trending on Hugging Face
๐ Academic Service
๐งโ๐ซ Student Supervision โ I co-supervise undergraduate, master’s, and PhD students on LLM agents and LLM post-training.
๐งญ Area Chair โ ACL.
๐ Conference Reviewing โ I regularly review for venues including NeurIPS, ICLR, EMNLP, KDD, WWW, and COLING.
๐ฐ Journal Reviewing โ I have also reviewed for journals such as TKDE, Science China, and AI Open.
๐๏ธ Membership โ Member of the CCF Technical Committee on Theoretical Computer Science (China Computer Federation).
๐ค Invited Talks โ Post-Training Methods for On-Device Agents on Long-Horizon Complex Tasks (RLChina 2026), Towards the Era of Agent 2.0 (SMP 2026), Autonomous Agents and Tool Learning with LLMs (RLChina 2025), and LLM-Driven Autonomous Agents (Huawei OpenHarmony AI Agent TSG).
๐ฐ Research Grants โ I serve as PI of an NSFC Young Scientists Fund (Category C) project, the China Postdoctoral Science Foundation General Grant, and the National Postdoctoral Researcher Program Category C.
๐ฌ Contact
Feel free to reach out if you would like to discuss research ideas, collaborations, or open-source projects.
โ๏ธ Email: haotian.chen@163.com / ht1ian.chen@gmail.com
๐ GitHub: github.com/Hytn
โ๏ธ X: @HytnChen
๐ Google Scholar: scholar profile