Weitong Qian

Weitong Qian钱韦潼

Undergraduate · Peking University · DAIR Lab

weitong.qian@stu.pku.edu.cn github google scholar

I am a junior undergraduate at Peking University, College of Engineering (COE), advised by Prof. Bin Cui and Asst. Prof. Xupeng Miao at the PKU-DAIR Lab.

I work on using AI to build AI: developing systems in which AI accelerates, automates, and extends the very process by which AI is created and deployed.

I am interested in turning AI inward on its own stack — the infrastructure it runs on, the algorithms that train it, the research process that produces it, and the physical embodiments through which it acts.

My research spans four layers of this stack:

  1. 01
    AI for AI Infrastructure Agents that write and optimize the code AI runs on.
  2. 02
    AI for AI Algorithms LLM-guided search for ML pipelines and hyperparameters.
  3. 03
    AI for AI Research Agent systems that compress the literature-to-publication loop.
  4. 04
    AI for the Physical World Embodied agents and world models that extend AI's reach beyond the screen.

Leadership & Service/ 01

Publications/ 02

Under Review

LB-MCTS: Synergizing Large Language Models and Bayesian Optimization for Efficient CASH

Beicheng Xu, Weitong Qian, Lingching Tung, Yupeng Lu, Bin Cui

Submitted to ICML 2026 · arXiv:2601.12355

Projects/ 03

01AI for AI Infrastructure

GPU Kernel Optimization Agent In progress

An open-source agent system that autonomously writes, profiles, and optimizes CUDA kernels — pushing AI further down its own software stack.

02AI for AI Algorithms

LB-MCTS ICML 2026 · under review

LLM-guided Monte Carlo Tree Search for the Combined Algorithm Selection and Hyperparameter optimization (CASH) problem. We use the LLM as a structural prior over the pipeline space, making Bayesian optimization dramatically more sample-efficient.

arxiv →

03AI for AI Research

OmegaWiki Open source ★382

Karpathy's LLM-Wiki vision, fully realized — a wiki-centric, full-lifecycle AI research platform powered by Claude Code. Twenty skills covering the loop from paper ingestion → idea generation → experiment design → paper writing → reviewer rebuttal. Knowledge compounds; failed experiments become anti-repetition memory.

github →
FrontierPilot 2nd Prize · Lobster Hackathon

An AI research-onboarding tool, built on the OpenClaw skill platform. Given a topic, it constructs a self-contained, growing knowledge base — field overview, foundational and frontier papers, peer reviews, knowledge graph — that a newcomer can converse with directly.

github → 3-min demo →

04AI for the Physical World

Embodied RL @ Galbot Feb – Jun 2025

Vision-based, RL-driven non-prehensile grasping; RL controllers for stable and adaptive locomotion on quadruped and biped robots over varied terrain. Advised by Prof. He Wang.

AI Robot Scientist Experiment System 3rd Prize · Beijing Challenge Cup

A system in which embodied AI agents autonomously plan and conduct scientific experiments — research automation extended into the physical world.

Experience/ 04

PKU-DAIR Lab, Peking University — Undergraduate Researcher Jul 2025 – Present

Working on automating the AI stack across algorithms, research, and infrastructure. Advised by Prof. Bin Cui and Asst. Prof. Xupeng Miao.

Galbot (北京银河通用机器人) — Algorithm Intern Feb 2025 – Jun 2025

Research on vision-based, RL-driven non-prehensile grasping; RL controllers for legged robots. Advised by Prof. He Wang.

Awards/ 05

Competitions & Recognition

Scholarships & Honors

News/ 06

  1. Apr 2026 Released OmegaWiki — a wiki-centric AI research lifecycle platform powered by Claude Code.
  2. Mar 2026 FrontierPilot wins 2nd Prize at the Zhongguancun OpenClaw Hackathon (Academic Track).
  3. Jan 2026 LB-MCTS submitted to ICML 2026 (under review).
  4. Jul 2025 Joined PKU-DAIR Lab.

Friends/ 07