zs·Zijun Song
Portrait of Zijun Song
ABOUT

Zijun Song

PhD student · School of Advanced Interdisciplinary Sciences, UCAS
Jointly trained with the Institute of Automation, CAS · Beijing

Email Scholar GitHub Blog

I am Zijun Song, a Ph.D. student at the School of Advanced Interdisciplinary Sciences, University of the Chinese Academy of Sciences, in a joint training program with the Institute of Automation, Chinese Academy of Sciences. I am advised by Prof. Dongbin Zhao.

My research focuses on post-training for search agents. At ModelBest, I contributed to the post-training of MiniCPM4 and AgentCPM-Explore. I currently work on post-training for MiSearchAgent at Xiaomi.

I received my B.Eng. in Network Engineering from the School of Software, Dalian University of Technology, in 2025. During my undergraduate studies, I conducted research on wireless sensing using Wi-Fi, ultra-wideband (UWB), and millimeter-wave radar.

LLMSEARCH AGENTPOST-TRAINING
EXPERIENCE

Internships

2024.07 – 2025.10

Research Intern, LLM Agent / Tool Learning, ModelBest

I worked on on-device LLM tool learning and agent post-training, contributing to MiniCPM4’s MCP tool-calling post-training and the development and core evaluation of AgentCPM-Explore. I led or contributed to multi-turn MCP tool-calling data synthesis, SFT and AgentRL training, and the AgentToLeaP evaluation platform. This work improved tool selection, argument generation, and multi-turn interaction, supporting a 4B on-device agent that achieved leading results in complex tool use and long-horizon exploration. Related publications include ToLeaP, Reflective Reinforcement Tool Learning, and the MiniCPM4 and AgentCPM-Explore technical reports.

2026.07 – Present

Search Agent Post-training Intern, Xiaomi

I work on post-training for search agents. On the product side, I contribute to post-training of the MiSearchAgent models, reducing latency without hurting performance. On the research side, I study how to internalize agent-system capabilities into the model and proposed HarnessOPD, which transfers capabilities from the harness into the model parameters.

BLOG

Selected posts

All posts ↗
METHOD NOTES2026.11.03 · 8 MIN READ

HarnessOPD: Keeping the Behavioral Gains of a Stronger Harness Inside the Model↗

A stronger search-agent harness can improve decisions, but the gains may vanish once those components are removed. HarnessOPD studies how to keep part of that advantage in the model parameters.

AgentPost-training
Research note
METHOD NOTES2026.11.03 · 8 MIN READ

RG-CTO: Letting Failure Experience Influence the Next Attempt Selectively↗

Failed traces should not be reused wholesale. RG-CTO estimates the reliability of negative experience before deciding how strongly it may shape generation.

AgentPost-training
Research note
PUBLICATIONS

Papers

Google Scholar ↗

Research Papers

Technical Reports

A work in progress.

Research is more than papers.
Some of the process belongs here too.

Get in touch ↗