About
Hello! My name is Weixian Xu. I am a senior majoring in Computer Science at Shanghai Jiao Tong University
, in the
Zhiyuan College ACM Honors Class. I am currently a research intern at the SALT Lab
at Stanford University, working with Prof. Diyi Yang. Previously, I was a remote research intern at the AI² Lab
at Princeton University with Prof. Mengdi Wang, and a research intern at the GAIR Lab
with Prof. Pengfei Liu.
I am always happy to talk about research and potential collaborations. Feel free to reach out.
Research Interests
I aim to develop increasingly intelligent systems to support practical applications and ultimately enhance human capabilities. A key strength of human intelligence lies in the ability to keep learning and growing. In contrast, as models face increasingly complex test-time environments that include intricate tasks and various human needs, their intelligence is still largely fixed after training.
My research therefore focuses on test-time learning, especially enabling AI to effectively identify what matters in increasingly heterogeneous feedback. I explore test-time learning across multiple directions.
Advancing AI itself
Draw insights from experiments to advance itself
Supporting practical applications
Meet diverse application requirements
Enhancing human capabilities
Infer users’ needs to better help humans
Publications
EEVEE: Towards Test-time Prompt Learning in the Real World for Self-Improving Agents
2nd Workshop on Lifelong Agents (LLA), COLM 2026
ASI-Evolve: Towards AI-Driven Discovery Across the AI Research Stack
Workshop on Language Models for Scientific Discovery (LM4Sci), COLM 2026
An AI System for Autonomous Algorithm Evolution in Drug Development
Preprint (bioRxiv)
Repositories
ASI-Evolve
871A unified AI scientist framework that runs a learn–design–experiment–analyze cycle across data curation, model architecture design, and algorithm discovery.
ASI-Arch
1194An AI scientist that autonomously designs linear attention architectures end to end, discovering 105 architectures that outperform Gated DeltaNet.
EEVEE
Test-time prompt learning that interleaves task routing and task solving to improve LLM agents on heterogeneous, real-world task streams.
Sherpa
Trains LLM teachers with multi-turn reinforcement learning to adapt their teaching to students with diverse learning preferences.
ACore
A comprehensive operating system kernel in Rust, supporting virtual memory and page allocation, process scheduling, thread management, user/kernel space isolation, and a virtual file system with multiple filesystems.
Imxc
A compiler for Mx*, a C-like language, written in Java. It reaches performance comparable to Clang through Mem2Reg and graph-coloring register allocation, with function inlining, SCCP, ADCE, and incremental multi-file builds.
RISC-V CPU
A RISC-V processor core in Verilog using Tomasulo’s algorithm, supporting the RV32I instruction set and deployment on FPGA hardware.
Background
Education

Research Experience

Honors
Teaching & Service