Weixian Xu

Student Researcher, Stanford SALT Lab

with Prof. Diyi Yang

ACM Honors Class, SJTU

Previously GAIR Lab

About Me

I am Weixian Xu, an undergraduate in the ACM Honors Class at Shanghai Jiao Tong University. I am currently conducting research at Stanford SALT Lab with Prof. Diyi Yang. Previously, I worked at GAIR Lab with Prof. Pengfei Liu, and I have also collaborated with Prof. Mengdi Wang and Shilong Liu at Princeton/AI2.

My long-term goal is to build AI systems that can recursively self-improve toward richer and more beneficial forms of intelligence. I am interested in the progression from analysis to reflection to improvement: systems that can examine their own behavior, understand why they succeed or fail, accumulate reusable knowledge, and use that knowledge to improve how they learn, solve problems, and collaborate with people.

I want AI to become stronger, but in the right direction: learning from analysis and reflection, and using its intelligence to help people in more meaningful ways.

Research Statement

My research asks two connected questions: how can AI keep improving itself, and what should that improvement mean? I study recursive self-improvement as a technical problem, but also as a question about the target of intelligence. The objective is to build agents that learn through analysis and reflection, improve over time, and use their intelligence to help people in meaningful ways.

This leads to two connected directions:

Continual Self-Improvement: Building agents that learn from their own research, tool use, experiments, and failures over long horizons, including systems that improve the process of AI research itself.
Human-Centered Intelligence: Defining intelligence by whether systems can analyze, reflect, learn, teach, collaborate, and help people in meaningful real-world settings.

Recent Work

ASI-Evolve: AI Accelerates AI

Closed-Loop Self-Improving Research Framework | Paper | Code

An agentic framework for recursive self-improvement in AI research that closes a long-horizon learn-design-experiment-analyze loop across data, architectures, and learning algorithms.

Core Idea: Combine accumulated research priors with an analyzer that turns experimental outcomes into reusable insights.
Scope: Studies closed-loop improvement across central components of the AI stack, emphasizing how agents can analyze results, learn from experiments, and carry knowledge into future research.
Research Direction: A step toward recursive self-improvement as a practical research process where better analysis and reflection lead to better future learning.

EEVEE: Test-Time Prompt Learning for Self-Improving Agents

Princeton/AI2 Collaboration | Paper | Project | Code

A multi-dataset test-time prompt learning framework for LLM agents operating under heterogeneous real-world task streams.

Core Idea: Use router-prompt co-evolution to assign inputs to specialized prompt slots and reduce cross-dataset interference.
Scope: Moves test-time prompt learning toward mixed domains, formats, and evaluation rules that better resemble real-world task streams.
Research Direction: Makes self-improving agents more practical by helping them specialize, adapt, and preserve useful behavior across changing contexts.

ASI-Arch: Autonomous Neural Architecture Discovery

Research Project (GAIR Lab) | Paper | Code (1.1k Stars)

An autonomous AI system designed to conduct neural architecture discovery and reveal new design patterns without human intervention.

Impact: The repository garnered 1.1k+ stars within a month of publication.
Contribution: Designed the autonomous discovery pipeline, orchestrated 1,773 large-scale experiments, and validated state-of-the-art architectures.

Engineering Foundation

ACore: Operating System Kernel

System Implementation | Code

A comprehensive OS kernel exploring core operating system concepts through independent development in Rust.

Implemented memory management with virtual memory and page allocation.
Developed process scheduling, thread management, and a VFS layer supporting multiple filesystems.
Engineered user/kernel space isolation for security.

Imxc: Advanced Compiler

System Implementation | Code

A full-featured compiler for a C-like language (Mx*), achieving performance comparable to Clang.

Implemented advanced optimizations including Mem2Reg, Function Inlining, SCCP, and ADCE.
Designed a high-performance Register Allocation module using Graph Coloring.
Supported incremental compilation for efficient multi-file builds.

RISC-V CPU (Tomasulo Architecture)

Hardware Implementation | Code

A sophisticated RISC-V processor core implementation in Verilog.

Features Tomasulo architecture for dynamic instruction scheduling and out-of-order execution.
Supports the complete RV32I instruction set with FPGA deployment capabilities.

Technical Background

Research: Python, PyTorch, Transformers, CUDA, LLM Agents, Self-Improving AI, Test-Time Learning, AI Research Agents, Neural Architecture Discovery.
Systems: Rust, C++, C, Verilog, OS Development, Compiler Design.
Infrastructure: Linux, Docker, Git, CI/CD.

Personal Interests

Beyond research and coding, music is a big part of my life. I play both acoustic and electric guitar, and I hold a Grade 9 certification in clarinet. I also enjoy badminton and video games, especially for their storytelling and mechanics. Back in high school, I was a competitive physics student (CPhO Provincial Second Prize), which shaped my analytical thinking.

On a more personal note, I am happily in a relationship and value the time spent exploring life’s journey with my partner.

Contact

I am always open to discussing self-improving AI, human-centered intelligence, systems engineering, or potential collaborations.

Email: hz.czar@sjtu.edu.cn
GitHub: @HZxCzar

selected publications

arXiv
ASI-Evolve: AI Accelerates AI

Weixian Xu, Tiantian Mi, Yixiu Liu, and 6 more authors

arXiv preprint arXiv:2603.29640, Mar 2026

Abs DOI arXiv Bib Code

Can AI accelerate the development of AI itself? ASI-Evolve is an agentic framework for self-improving AI research that closes a learn-design-experiment-analyze loop. The framework augments evolutionary agents with a cognition base that injects accumulated research priors and an analyzer that distills experimental outcomes into reusable insights. It demonstrates AI-driven discovery across data, architectures, and learning algorithms, offering evidence for closed-loop AI research.
@article{xu2026asievolve, title = {ASI-Evolve: AI Accelerates AI}, author = {Xu, Weixian and Mi, Tiantian and Liu, Yixiu and Nan, Yang and Zhou, Zhimeng and Ye, Lyumanshan and Zhang, Lin and Qiao, Yu and Liu, Pengfei}, journal = {arXiv preprint arXiv:2603.29640}, year = {2026}, month = mar, url = {https://arxiv.org/abs/2603.29640}, doi = {10.48550/arXiv.2603.29640}, }
arXiv
EEVEE: Towards Test-time Prompt Learning in the Real World for Self-Improving Agents

Weixian Xu, Shilong Liu, and Mengdi Wang

arXiv preprint arXiv:2606.11182, Jun 2026

Abs DOI arXiv Bib Code Website

EEVEE is a multi-dataset test-time prompt learning framework for LLM agents operating under real-world task streams. It uses a learned router to partition incoming inputs into task clusters and assigns them to specialized prompt configurations. Router-prompt co-evolution improves robustness under heterogeneous data streams while preserving useful learned behavior within each task family.
@article{xu2026eevee, title = {{EEVEE}: Towards Test-time Prompt Learning in the Real World for Self-Improving Agents}, author = {Xu, Weixian and Liu, Shilong and Wang, Mengdi}, journal = {arXiv preprint arXiv:2606.11182}, year = {2026}, month = jun, url = {https://arxiv.org/abs/2606.11182}, doi = {10.48550/arXiv.2606.11182}, }
arXiv
AlphaGo Moment for Model Architecture Discovery

Yixiu Liu, Yang Nan, Weixian Xu, and 4 more authors

arXiv preprint arXiv:2507.18074, Jul 2025

Abs DOI Bib

While AI systems demonstrate exponentially improving capabilities, the pace of AI research itself remains linearly bounded by human cognitive capacity, creating an increasingly severe development bottleneck. We present ASI-Arch, the first demonstration of Artificial Superintelligence for AI research (ASI4AI) in the critical domain of neural architecture discovery–a fully autonomous system that shatters this fundamental constraint by enabling AI to conduct its own architectural innovation. Moving beyond traditional Neural Architecture Search (NAS), which is fundamentally limited to exploring human-defined spaces, we introduce a paradigm shift from automated optimization to automated innovation. ASI-Arch can conduct end-to-end scientific research in the domain of architecture discovery, autonomously hypothesizing novel architectural concepts, implementing them as executable code, training and empirically validating their performance through rigorous experimentation and past experience. ASI-Arch conducted 1,773 autonomous experiments over 20,000 GPU hours, culminating in the discovery of 106 innovative, state-of-the-art (SOTA) linear attention architectures. Like AlphaGo’s Move 37 that revealed unexpected strategic insights invisible to human players, our AI-discovered architectures demonstrate emergent design principles that systematically surpass human-designed baselines and illuminate previously unknown pathways for architectural innovation. Crucially, we establish the first empirical scaling law for scientific discovery itself–demonstrating that architectural breakthroughs can be scaled computationally, transforming research progress from a human-limited to a computation-scalable process. We provide comprehensive analysis of the emergent design patterns and autonomous research capabilities that enabled these breakthroughs, establishing a blueprint for self-accelerating AI systems.
@article{liu2025alphago, title = {AlphaGo Moment for Model Architecture Discovery}, author = {Liu, Yixiu and Nan, Yang and Xu, Weixian and Hu, Xiangkun and Ye, Lyumanshan and Qin, Zhen and Liu, Pengfei}, journal = {arXiv preprint arXiv:2507.18074}, year = {2025}, month = jul, url = {https://arxiv.org/abs/2507.18074}, doi = {10.48550/arXiv.2507.18074}, }