Research

My research aims at self-improving AI, organized as a stack in which every layer has to hold:

• Data & Environments. Interactive playgrounds where agents run the real experiment loop, and automated pipelines that manufacture verifiable tasks at scale (MLE-Dojo, MLE-Smith, RLE-Bench).
• Algorithms. Post-training that stays dense, reliable, and on-policy over long horizons (Agent DAgger).
• Infrastructure. Modular systems that schedule and scale agentic reinforcement learning (STACX).
• Harnesses. Hierarchical orchestration that sustains long-horizon optimization and research (Matryoshka Agent), a robotics harness that extends coding agents to any robot (Kinex), and a self-improver built on it that solved all 365 RoboCasa tasks by evolving the robot’s skills and functions from its own trials (Physical RSI).

I wish to find the right way of scaling for Recursive Self-Improvement.

I strive for research and projects that are scalable and have real impact on frontier models.

I contribute to frontier models such as Gemini 3.x Flash and Gemini 4 models.

MLE-Dojo and MLE-Smith have been widely adopted for building MLE/RSI agents and systems, and have proven scalable (e.g., OpenRSI).

Publications & Preprints

• ICLR 2026 MLE-Smith: Scaling MLE Tasks with Automated Multi-Agent Pipeline Rushi Qiang, Yuchen Zhuang, Anikait Singh, Percy Liang, Chao Zhang, Sherry Yang, Bo Dai
• NeurIPS 2025 · Datasets & Benchmarks MLE-Dojo: Interactive Environments for Empowering LLM Agents in Machine Learning Engineering Rushi Qiang^*, Yuchen Zhuang^*, Yinghao Li, Dingu Sagar V K, Rongzhi Zhang, Changhao Li, Ian Shu-Hei Wong, Sherry Yang, Percy Liang, Chao Zhang, Bo Dai
• NeurIPS 2025 Matryoshka Pilot: Learning to Drive Black-Box LLMs with LLMs Changhao Li^*, Yuchen Zhuang^*, Rushi Qiang, Haotian Sun, Hanjun Dai, Chao Zhang, Bo Dai
• COLM 2025 Language Model Uncertainty Quantification with Attention Chain Yinghao Li, Rushi Qiang, Lama Moukheiber, Chao Zhang
• NeurIPS 2024 HYDRA: Model Factorization Framework for Black-Box LLM Personalization Yuchen Zhuang, Haotian Sun, Yue Yu, Rushi Qiang, Qifan Wang, Chao Zhang, Bo Dai
• NAACL 2024 AutoLoRA: Automatically Tuning Matrix Ranks in Low-Rank Adaptation Based on Meta Learning Ruiyi Zhang^*, Rushi Qiang^*, Sai Ashish Somayajula, Pengtao Xie
• CoRL 2026 · Agentic Robotics Workshop RLE-Bench: A Qualifying Exam for Coding Agents as Robot Learning Engineers Haitong Ma^*, Chenxiao Gao^*, Rushi Qiang^*, Bo Dai, Na Li
• NeurIPS 2026 · AgenticOS Workshop Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering Junlin Yang, Che Jiang, Yu Fu, …, Rushi Qiang, …, Bowen Zhou, Kaiyan Zhang (24 authors)
• arXiv · 2026 Matryoshka Agent: Unfolding Sub-Agents for Long-Horizon Machine Learning Engineering Rushi Qiang, Changhao Li, Haotian Sun, Yuchen Zhuang, Chao Zhang, Bo Dai
• arXiv · 2026 Revisiting DAgger in the Era of LLM-Agents Changhao Li, Rushi Qiang^†, Jiawei Huang^†, Chenxiao Gao^†, Chao Zhang, Niao He, Bo Dai
• arXiv · 2026 Forward-Free Diffusion Language Models Haotian Sun, Rushi Qiang, Yuqian Zheng, Bo Dai
• arXiv · 2026 TimeRSI: Towards Self-Improved Agentic Deployment for Time-Series Foundation Models Haoxin Liu, Deqing Fu, Lucheng Fu, Yiqiao Jin, Rushi Qiang, Harshavardhan Kamarthi, B. Aditya Prakash
• arXiv · 2025 Towards Better Instruction Following Retrieval Models Yuchen Zhuang^*, Aaron Trinh^*, Rushi Qiang^*, Haotian Sun, Chao Zhang, Hanjun Dai, Bo Dai

^* equal contribution  ·  ^† equal second authorship

Software

• Kinex An open-source robotics harness for coding agents. Three layers of devices, tools, and skills expose a robot as composable devices, so one agent adapts to arbitrary robots in simulation and on real hardware, including ROS 2. Also available as plugins for Pi, Claude Code, and Codex. Project website.
• STACX A modular infrastructure for end-to-end agentic reinforcement learning: a sandboxed task-execution engine coupled with a multi-algorithm training engine (GRPO, SFT, DAgger) for scheduling and scaling agent training.
• MLE-Dojo A Gym-style framework for training, evaluating, and improving LLM agents on iterative machine learning engineering workflows, with interactive sandbox environments built on 200+ real Kaggle competitions.
• RLE-Bench A benchmark for evaluating coding agents on robot learning engineering tasks through observation, experimentation, and iterative development.

Experience

• Google DeepMind, Mountain View Student Researcher · agentic MLE for Gemini Feb 2026 – present

Education

• Georgia Institute of Technology, Atlanta Ph.D. student in Machine Learning · co-advised by Bo Dai and Chao Zhang Aug 2024 – present
• Tsinghua University, Beijing B.Eng. in Automation Sep 2020 – Jun 2024

Academic Service

Reviewer for NeurIPS, ICLR, ICML, and ACL Rolling Review.