About
I am a PhD candidate at the State Key Laboratory of Processors, Institute of Computing Technology, CAS, and University of Chinese Academy of Sciences, advised by Prof. Huimin Cui and Prof. Jiacheng Zhao.
My research and engineering focus on Kernel & Infrastructure Agents and Live System Reconfiguration. I build agents for low-level code and systems optimization, and study how systems can change their execution strategies while preserving state and continuing to run. I received my B.Eng. in Computer Science from UCAS in 2021.
Research Vision
I am interested in systems that stay running while their internal state, resource ownership, and execution strategies can be changed safely. SPLASH is my first step in this direction: it switches attention parallelism layouts while requests are running, reusing existing state and handing off at a batch boundary. My ongoing work explores this idea more broadly across systems infrastructure.
I see this direction and kernel/infrastructure agents as mutually reinforcing. Agents can help discover, implement, and validate optimizations close to the hardware; systems that support live reconfiguration can give those agents a way to apply and evaluate changes during execution. My goal is to connect agent-driven optimization with the runtime mechanisms that make those optimizations usable in a running system.
News
- Aug 2026 — First blog post: pi-think: Your Coding Agent’s ‘Private’ Thoughts Are Just Tool Arguments.
- Jul 2026 — New preprint Decode-Time Grammars: Constrained LLM Generation over a Refinement Order of Grammar Fragments on arXiv.
- Jul 2026 — MoonBright accepted at OSDI 2026.
- May 2026 — New preprint KLineage: Recovering the Missing When of Kernel Optimization by Deoptimizing Experts on arXiv.
- May 2026 — Two papers accepted at ICML 2026, including one spotlight (CONTINUUM).
- Apr 2026 — When Grammar Guides the Attack accepted at CCS 2026.
- Mar 2026 — Symbiotic MLLM Serving accepted at ISCA 2026.
- Mar 2026 — LEGO-Compiler accepted at CCF THPC.
- Feb 2026 — T2T receives the Distinguished Paper Award at CGO 2026.
- Sep 2025 — Two papers accepted at NeurIPS 2025 (posters).
Selected Publications
- [arXiv 2026] SPLASH: Switching Parallel Layouts of Attention with Seamless Handoff for LLM Serving — Chuan Liu, Shuoming Zhang*, Zhicheng Li, Qianqi Sun, Ruiyuan Xu, Qiuchu Yu, Xiyu Shi, Huimin Cui, Jiacheng Zhao. [PDF] [arXiv]
- [arXiv 2026] Decode-Time Grammars: Constrained LLM Generation over a Refinement Order of Grammar Fragments — Shuoming Zhang, Ruiyuan Xu, Haofeng Li, Qiuchu Yu, Yangyu Zhang, Chunwei Xia, Xiaobing Feng, Chenxi Wang, Huimin Cui, Jiacheng Zhao*
- [Architecture 2.0 @ ISCA 2026 workshop] From Skills to Tracelets: Dependency-Guided Transfer for LLM Kernel Agents — Shuoming Zhang, Ruiyuan Xu, Qiuchu Yu, Guangli Li, Huimin Cui, Jiacheng Zhao*
- [arXiv 2026] KLineage: Recovering the Missing When of Kernel Optimization by Deoptimizing Experts — Shuoming Zhang†, Qiuchu Yu†, Ruiyuan Xu, Chenjing Zhang, Junjie Peng, Zhicheng Xie, Yangyu Zhang, Xiyu Shi, Ying Liu, Guangli Li, Xiaobing Feng, Huimin Cui, Xingjun Zhang, Jiacheng Zhao*. [PDF] [arXiv]
- [ICML 2026 spotlight] CONTINUUM: Restoring the Contiguous Tensor Abstraction Efficiently for Dynamic AI Workloads via Hardware Virtualization — Yangyu Zhang†, Shuoming Zhang†, Chunwei Xia, Shuaijiang Li, Zhicheng Li, Ruiyuan Xu, Zheming Yang, Lei Chen, Yuan Wen, Guangli Li, Xiaobing Feng, Huimin Cui, Jiacheng Zhao*
- [ICML 2026] LEGO: An LLM-Enabled Hierarchical Optimizer for Tensor Computation Graphs with Structure-Aware Search and Compositional Synthesis — Ruiyuan Xu†, Shuoming Zhang†, Guangli Li, Qiuchu Yu, Rui Zhang, Yangyu Zhang, Hao Qian, Chunwei Xia, Jiacheng Zhao, Chenxi Wang, Xiaobing Feng, Jingling Xue, Huimin Cui
- [CCS 2026] When Grammar Guides the Attack: Uncovering Control-Plane Vulnerabilities in LLMs with Structured Output — Shuoming Zhang, Jiacheng Zhao*, Hanyuan Dong, Ruiyuan Xu, Zhicheng Li, Yangyu Zhang, Shuaijiang Li, Yuan Wen, Chunwei Xia, Zheng Wang, Xiaobing Feng, Huimin Cui
- [ISCA 2026] Symbiotic MLLM Serving: Dynamically Balancing Parallelism Across GPUs and Resources Within GPUs — Zhicheng Li, Jiacheng Zhao*, Yangyu Zhang, Zhaolin Duan, Xinyu Liu, Siqi Li, Shuoming Zhang, Shuaijiang Li, Donglin Yu, Yuan Wen, Chunwei Xia, Xiyu Shi, Huimin Cui
- [OSDI 2026] MoonBright: A GPU Memory Allocator with Device-Side Page Table Materialization and Deferred TLB Coherence — Yangyu Zhang, Lei Chen, Chunwei Xia, Shuaijiang Li, Shuoming Zhang, Zhicheng Li, Qianqi Sun, Jiawei Xiao, Ruiyuan Xu, Ao Chen, Guangli Li, Xiaobing Feng, Huimin Cui, Chenxi Wang, Jiacheng Zhao*
- [CGO 2026 Distinguished Paper Award] From Threads to Tiles: T2T, a Compiler for CUDA-to-NPU Translation via 2D Vectorization — Shuaijiang Li, Jiacheng Zhao*, Ying Liu, Shuoming Zhang, Lei Chen, Yijin Li, Yangyu Zhang, Zhicheng Li, Runyu Zhou, Xiyu Shi, Chunwei Xia, Yuan Wen, Xiaobing Feng, Huimin Cui
See the full publication list.
Research Interests
- Kernel & Infrastructure Agents — generation, optimization, and verification of low-level kernels and systems code
- Live System Reconfiguration — changing runtime state, resource ownership, and execution strategies while preserving ongoing work
- LLM serving, runtime systems, and constrained decoding
Research Projects
Current
- Kernel & infrastructure agents — Building agents that synthesize, tune, and verify kernels and systems code, using execution feedback to guide optimization.
- Live system reconfiguration — Exploring how systems can retain state and adapt during execution, starting with SPLASH’s hot switching of attention parallelism layouts in LLM serving.
- Constrained and structured decoding — Decoding as LLM infrastructure: constrained and grammar-based decoding for controllable, reliable, and safe generation.
- Grammar-guided code generation — Decode-time grammars that steer LLMs with grammar constraints to produce syntactically and semantically valid code, eliminating invalid references by construction (arXiv).
Previous
- LLM-guided compilation workflows — Model-in-the-loop compilation pipelines for source-to-assembly translation and error recovery with LLM feedback.
- LLM-aware compiler construction — Reusable compiler components that leverage LLM reasoning for IR transformation, code generation, and verification.
- Heterogeneous model offloading with TVM (Intel collaboration) — Explored NPU/CPU co-execution and scheduling strategies within the TVM stack, prototyped a new TVM backend for simulator-based NPU.
- VLIW instruction scheduling (Huawei collaboration) — Developed instruction scheduling heuristics targeting domain-specific VLIW architectures.
Professional Service
- Reviewer — TMLR (Transactions on Machine Learning Research), Artificial Intelligence Review, ICML, NeurIPS, ICLR, AAAI 2027
Talks & Activities
- Jul 2026 — Gave two poster presentations at ICML 2026.
- Jul 2026 — Attended OSDI 2026.
- Jun 2026 — Attended ISCA 2026 and gave two talks: Symbiotic MLLM Serving (main conference) and Agentic Kernel Optimization (Architecture 2.0 workshop).
- May 2025 — Talk at Huawei (Beijing) on LLM for Code Generation.
- Nov 2024 — Talk at the Institute of Information Engineering (IIE), CAS on LLM for Compiler.
Education
- Ph.D. in Computer Architecture, ICT CAS & UCAS, 2021 – now
- B.Eng. in Computer Science, UCAS, 2017 – 2021