Youwei Xiao

I am a Ph.D. candidate at the School of Integrated Circuits, Peking University, advised by Prof. Yun Liang, and a Large Language Model Algorithm Intern at ByteDance Seed. My current research centers on agents for LLM infrastructure on emerging AI chips. I contribute to native multi-agent coding systems, focusing on multi-agent orchestration strategies and workflow customization through programmable concurrency and human-in-the-loop workflows for long-horizon systems development.

I apply these approaches to LLM infrastructure on emerging AI chips. I evolve an abstraction-layered operator-optimization skill system and an agent-native compiler stack for operator implementation, chip-architecture-aware performance optimization, and automated toolchain development. This work improves model-serving efficiency on new chips, while real development data and trajectories become training signals for model-capability improvement.

My earlier research follows a compiler-driven approach to agile chip design: use DSLs and multi-level IRs to expose architecture choices, then build compilation and synthesis stacks that optimize software and hardware together. Hector provides multi-level IRs for hardware synthesis. Cement couples the cycle-deterministic CmtHDL DSL with the CmtC compiler for timing analysis and control synthesis. APS/Aquas builds an MLIR-based stack from architecture DSLs to hardware synthesis and software compilation, then extends it with domain-specific memory/synthesis directives and an e-graph retargetable compiler.

Across this compiler-driven stack, ISAMORE (ASPLOS 2026 Best Paper) and EggMind structure optimization search with formal and LLM-guided methods; Clay, Cayman, and SkyEgg automate microarchitecture and accelerator optimization; and IntelliC makes compiler artifacts inspectable for human-agent collaboration. PTO Runtime carries compiled task graphs onto Ascend chips and LingQu SuperPods, while Hive and Spine extend the story toward multi-agent LLM infrastructure and cross-layer agentic co-design. These systems are now the technical substrate on which my native multi-agent work operates.