zj.

Senior Research Engineer · Meta

Zijian Jin.

You can call me Alex.

Coding agents &
asynchronous RL.

My research interests center on coding agents and asynchronous reinforcement learning. I’m interested in how language models learn from interaction and feedback to solve complex coding tasks.

More broadly, I’m interested in scalable post-training and reliable, long-horizon agent behavior. Previously, I was a Research Scientist at Microsoft and TikTok.

Explore my research GitHub
Email Google Scholar LinkedIn
Zijian Jin
Zijian Jin / Alex

01 / Research

Selected work

2021–2026
01Jul 2026
EMNLP 2026/Evaluation

ChainSWE

Benchmarking Coding Agents on Multi-Bug Software Maintenance

What happens when a coding agent has to fix the next bug, and the one after that? Evaluating software maintenance as a continuous process.

Qirui Jin, Lingching Tung, Kenan Li, et al. / Co-author: Zijian Jin

Paper
02Apr 2026
EMNLP 2026/Agent systems

SWE-Edit

Rethinking Code Editing for Efficient SWE-Agent

Separating code inspection from editing, with an editor trained to choose how to make each change.

Yikai Zhang, Jiaxin Pei, Kenan Li, et al. / Co-author: Zijian Jin

Paper Code
03Apr 2026
arXiv 2026/Agent analysis

ORACLE-SWE

Quantifying the Contribution of Oracle Information Signals on SWE Agents

Measuring which information helps coding agents succeed—from finding edit locations to understanding tests.

Kenan Li, Qirui Jin, Liao Zhu, et al. / Co-author: Zijian Jin

Paper
arXiv 2026RepoLaunch: Automating Build & Test Pipelines Across Languages and PlatformsarXiv 2025Lita: Light Agent Uncovers the Agentic Coding Capabilities of LLMsICPR 2024CoCoT: Contrastive Chain-of-Thought Prompting for Large Multimodal ModelsEMNLP 2021 · OralNeuralizing Regular Expressions for Slot FillingAll publications on Google Scholar

02 / Projects

From ideas to systems.

World models / Adaptive computation

RefineJEPA

How much thinking does each state need?

An adaptive refinement framework for latent world models. It learns when another refinement step is useful, allocating computation to the states that need it most.

Research implementation

SWE-Edit

A framework for more effective, more efficient code editing.

microsoft / SWE-Edit
More on GitHub

03 / Experience

Where I’ve been building.

Jan 2026 — Present

Meta

Senior Research Engineer

Sep 2024 — Jan 2026

Microsoft

Research Scientist

Jul 2022 — Sep 2024

TikTok

Research Scientist

M.S. Computer Engineering, New York University 2022
B.S. Electronic Science and Technology, Xidian University 2019