| Murong Ma | LLM Post-Training, Agentic RL, and Data Curation |
About Me
I am a Computer Science Ph.D. student at the National University of Singapore, supervised by Prof. Jin Song Dong and Prof. Yun Lin. I build post-training systems for long-horizon agents and am currently a research intern at Microsoft Research Asia (MSRA).
My research focuses on process supervision, agentic reinforcement learning, full-repository generation, and verified data curation. I have hands-on experience training 1.5B–32B models and running training and serving experiments on 8 to over 100 GPUs. My broader research spans explainable AI, multimodal learning, video understanding, and speech. I am currently open to research and industry opportunities.
Research Interests
- LLM Post-Training and Agentic RL: using privileged process supervision, trajectory synthesis, and verifiable intermediate rewards to make difficult, long-horizon tasks learnable.
- Complex Task Construction and Data Curation: building verified training tasks for issue resolution, full-repository generation, and terminal workflows; curating data to improve capability and reliability.
- Explainable AI (XAI): developing actionable explanations that help people understand model behavior and make informed decisions.
Education
National University of Singapore
National University of Singapore · Artificial Intelligence specialization
Beijing University of Posts and Telecommunications
Experience
Research Intern
Working with Dr. Yeyun Gong on post-training systems for long-horizon agents, from privileged supervision to complex task construction.
- Privileged supervision for agent post-training — project lead. Using reference artifacts to make hard tasks learnable through guided trajectory construction and verifiable intermediate rewards.
- P2T: process-supervised trajectory generation for SFT. Developed a multi-agent framework that uses reference patches to recover successful issue-resolution trajectories where unguided teacher rollouts fail. Fact-DAG distillation turns privileged information into actionable guidance, while a neural-symbolic critic preserves trajectory realism without exposing reference solutions to the agent. Served Qwen3-Coder-480B and GLM-5.1 teachers on 32 H200 GPUs and trained Qwen2.5-Coder 7B/14B/32B students on approximately 2,000 tasks with SWIFT-Megatron. Achieved up to 10.8 percentage points higher Pass@1 across SWE-bench Verified/Lite and DeepSWE, with approximately 15% lower inference cost per instance. Accepted at NeurIPS 2026.
- Milestone-based RL for long-horizon tasks. Extending privileged supervision to GRPO by decomposing reference solutions and terminal goals into intermediate, verifiable milestones, so partially successful rollouts provide learning signals. Training Qwen3.5-30B with slime on 64 MI300X GPUs over 4,000 SWE tasks, with 16 rollouts per task. Increased the fraction of rollouts receiving a learning signal from 18% to 82%.
- Complex task construction and teacher distillation — collaborator. Expanding agent training beyond issue fixing to full-repository generation and multi-step terminal workflows.
- Full-repository generation. Co-collected 22,000 repositories from PyPI and GitHub and constructed tasks through feature-to-test mapping and feature recombination, tying functional requirements to executable checks while increasing task diversity and difficulty. SFT on Qwen3.8-Next-Flash teacher trajectories improved Qwen3.5-27B from 27.3% to 38.0% on NL2Repo.
- Terminal-agent task synthesis and evolution. Co-developed a pipeline that grounds StackExchange Q&A pairs in real GitHub issues and repositories, constructs executable environments, and evolves tasks through algorithmic upgrades, fault diagnosis and repair, and toolchain integration. Each round introduces new capability requirements and broadens reasoning and tool use. SFT on GLM-5.2 teacher trajectories improved Qwen3.5-27B from 36.15% to 49.44% Pass@1 on Terminal-Bench 2.1.
- Repository-level mid-training — collaborator. Co-developed a large-scale pipeline to curate and verify pull requests as training signals for repository-level code editing. ICML 2026.
ShowLab · National University of Singapore
Research Assistant
- Completed the M.S. thesis Subject-Adaptive Speech-Driven 3D Facial Animation, focusing on 3D avatar generation and speech-driven animation.
- Built a high-resolution face video dataset—a 4K variant of VoxCeleb—for face generation, talking-head synthesis, and cross-modal research.
- Helped organize the Ego4D Challenge 2022 as a point of contact for the Talking to Me Challenge.
SMIIP Lab · Data Science Research Center · Duke Kunshan University
Research Intern
- Conducted research in speech recognition, bilingual query-by-example spoken-term detection, keyword spotting, and multimodal speech-driven facial animation.
- Helped organize the 2020 Personalized Voice Trigger Challenge; the team placed second in the Interspeech 2021 Auto-KWS Challenge.
Publications
Recent and Selected
From Patches to Trajectories: Privileged Process Supervision for Software-Engineering Agents
Murong Ma, Tianyu Chen, Yun Lin, Shuai Lu, Qinglin Zhu, Yeyun Gong, Zhiyong Huang, Peng Cheng, Yan Lu, and Jin Song Dong. In Neural Information Processing Systems (NeurIPS), 2026.Pull Requests as a Training Signal for Repo-Level Code Editing
Qinglin Zhu, Tianyu Chen, Shuai Lu, Lei Ji, Runcong Zhao, Murong Ma, Xiangxiang Dai, Yulan He, Lin Gui, Peng Cheng, and Yeyun Gong. In International Conference on Machine Learning (ICML), 2026.TrainRef: Curating Data with Label Distribution and Minimal Reference for Accurate Prediction and Reliable Confidence
Murong Ma, Ruofan Liu, Yun Lin, Zhiyong Huang, and Jin Song Dong. In The Fourteenth International Conference on Learning Representations (ICLR), 2026.Few-Shot Precise Event Spotting via Unified Multi-Entity Graph and Distillation
Zhaoyu Liu, Kan Jiang, Murong Ma, Zhe Hou, Yun Lin, and Jin Song Dong. Proceedings of the AAAI Conference on Artificial Intelligence, 40(9):7422–7430, 2026.F3Set: Towards Analyzing Fast, Frequent, and Fine-grained Events from Videos
Zhaoyu Liu, Kan Jiang, Murong Ma, Zhe Hou, Yun Lin, and Jin Song Dong. In International Conference on Learning Representations (ICLR), pp. 10566–10580, 2025.Revisiting the Conflict-Resolving Problem from a Semantic Perspective
Jinhao Dong, Jun Sun, Yun Lin, Yedi Zhang, Murong Ma, Jin Song Dong, and Dan Hao. In Proceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering (ASE), pp. 141–152, 2024.
Manuscripts under Review
- CausalCoder: Causality-Guided Data Curation for Robust Code Generation
Murong Ma, Yilu Cao, Ruofan Liu, Wenhao Zhang, Jin Song Dong, and Yun Lin. Submitted to FSE 2027; under review.
Earlier Work
Novel View Synthesis for High-fidelity Headshot Scenes
Satoshi Tsutsui, Weijia Mao, Sijing Lin, Yunyi Zhu, Murong Ma, and Mike Zheng Shou. arXiv preprint arXiv:2205.15595, 2022.Acoustic Word Embedding System for Code-Switching Query-by-Example Spoken Term Detection
Murong Ma, Haiwei Wu, Xuyang Wang, Lin Yang, Junjie Wang, and Ming Li. In 2021 12th International Symposium on Chinese Spoken Language Processing (ISCSLP), pp. 1–5, 2021.The DKU System Description for the Interspeech 2021 Auto-KWS Challenge
Yechen Wang, Yan Jia, Murong Ma, Zexin Cai, and Ming Li. arXiv preprint arXiv:2104.04993, 2021.Training Wake Word Detection with Synthesized Speech Data on Confusion Words
Yan Jia, Zexin Cai, Murong Ma, Zeqing Zhao, Xuyang Wang, Junjie Wang, and Ming Li. arXiv preprint arXiv:2011.01460, 2020.
Technical Expertise
Code-Agent Post-Training
Supervised fine-tuning (SFT), agentic RL (GRPO), privileged process supervision, milestone-based rewards, and trajectory synthesis and curation
Training & Inference Infrastructure
slime, SWIFT-Megatron, Megatron-LM, SGLang, vLLM, and distributed GPU training and serving
Evaluation & Verification
SWE-bench Verified/Lite, DeepSWE, NL2Repo, Terminal-Bench 2.1, repository-level task construction, feature-to-test mapping, and sandboxed verification
Programming & Systems
Python, C++, C, Linux, Git
Selected Honors
- NUS Research Scholarship
- Excellent Bachelor Thesis, Beijing University of Posts and Telecommunications (top 10 among 630 students)
- Second-Class Scholarship (top 30 among 320 students)