I build infrastructure for LLM agents — systems where agents are first-class computational entities rather than an application layer bolted onto a chat API.
I am a fifth-year Ph.D. candidate at UC San Diego, advised by Prof. Jishen Zhao, and since June 2026 a Deep Learning Compiler Intern at NVIDIA. Before UCSD I received my B.E. from the Department of Electronic Engineering at Tsinghua University.
My work sits between machine learning and computer systems, and the through-line is that I keep treating model-shaped problems as systems problems. What an agent needs is rarely a bigger prompt — it is a memory hierarchy, an index that survives edits, and a serving layer with measurable lifecycle costs. That premise drives most of what I build now: repository context serving, memory protocols for multi-agent systems, and benchmarks that measure software as it evolves rather than one patch at a time. A parallel thread puts agents to work on hardware, generating and verifying RTL.
Placed 1st / 4th / 4th across GPU MODE’s B200 linear-algebra challenges: Cholesky, QR v2, and Symmetric Eigh. I also published a visual tutorial connecting the mathematical derivations to GPU kernel mappings and retained benchmark evidence.
Jul 28, 2026
Released CodeNib, a multi-view data system for serving repository context to coding agents — paper, code, codenib.ai.
Jun 15, 2026
Started as a Deep Learning Compiler Intern at NVIDIA, working on compiler-agent co-design for GPU kernel evolution.
Mar 13, 2026
SWE-Milestone is out — a benchmark for continuous software evolution where frontier agents fall from >80% on isolated tasks to 38%. To appear at ICML 2026.
CAKE co-designs GPU kernel agents with a compiler around CAKE IR, a typed, hardware-explicit schedule representation that exposes warp roles, memory movement, synchronization, and pipelines. Verification, cost modeling, and localized diagnostics turn recurring failures into reusable compiler and agent improvements. Across Flash-KMeans, Kimi Delta Attention, KNN, and KMeans, CAKE produces kernels competitive with or faster than tuned baselines across NVIDIA GPUs from Ampere through Blackwell.
@article{ye2026cake,title={{CAKE}: Compiler-Agent Co-Design for Frontier Kernel Evolution},author={Ye, Zihao and Huang, Yingyi and Jin, Hongyi and Hou, Bohan and Shao, Junru and Yu, Zhongming and Chen, Jinqi and Cowan, Meghan and Cao, Shiyi and Xing, Shanli and Chen, Hanfeng and Grover, Vinod and Chen, Tianqi and Ceze, Luis},journal={arXiv preprint arXiv:2608.12629},year={2026},}
Coding agents repeatedly search, navigate, and retain context from evolving repositories, but disconnected indexes, language servers, and task-local histories force repeated discovery and obscure lifecycle costs. CodeNib builds reusable lexical, dense, and structural views per repository commit, maps outputs to repository-relative source ranges, maintains selected views across edits, and serves ranked search, symbol navigation, and bounded context through one runtime. Across 100 snapshots, graph and vector updates are 8.7x and 25.4x faster when outputs match independent rebuilds, static navigation requests that match live-server locations see a 4.7x median latency improvement, and context selection policies reduce trajectory tokens by 50-87% relative to grep/read baselines.
title = {CodeNib: A Multi-View Data System for Serving Repository Context to Coding Agents},author = {Yu, Zhongming and Yu, Hengjia and Yuan, Boqin and Zhao, Shuting and Chen, Yizhao and Dokania, Aryan and Jagtap, Mihir and Chang, Jiayu and Ma, Yitong and Jayswal, Yash and Ni, Wentao and Zhang, Hejia and Chen, Zhaoling and Deng, Gangda and Zhao, Jishen},journal = {arXiv preprint arXiv:2607.25431},year = {2026},}
Real software is maintained, not written once. We introduce SWE-Milestone, a benchmark that evaluates agents on streams of milestone-level tasks reconstructed from commit logs by DeepCommit, requiring sustained system integrity rather than isolated patches. Evaluating 12 frontier models across 4 agent frameworks, overall performance drops from >80% on isolated tasks to 38.03% in continuous settings, exposing error propagation and technical-debt accumulation that current benchmarks do not measure.
@inproceedings{deng2026swemilestone,title={SWE-Milestone: Evaluating AI Agents on Continuous Software Evolution},author={Deng, Gangda and Chen, Zhaoling and Yu, Zhongming and Fan, Haoyang and Liu, Yuhong and Yang, Yuxin and Parikh, Dhruv and Kannan, Rajgopal and Cong, Le and Wang, Mengdi and Zhang, Qian and Prasanna, Viktor and Tang, Xiangru and Wang, Xingyao},booktitle={Proceedings of the 43rd International Conference on Machine Learning},year={2026},}
As LLM agents evolve into collaborative multi-agent systems, their memory requirements grow rapidly in complexity. This position paper frames multi-agent memory as a computer architecture problem: we distinguish shared from distributed memory paradigms, propose a three-layer memory hierarchy (I/O, cache, and memory), and identify two protocol gaps – cache sharing across agents and structured memory access control. We argue that multi-agent memory consistency is the most pressing open challenge ahead.
@article{yu2026multiagentmemory,title={Multi-Agent Memory from a Computer Architecture Perspective: Visions and Challenges Ahead},author={Yu, Zhongming and Yu, Naicheng and Zhang, Hejia and Ni, Wentao and Yin, Mingrui and Yang, Jiaying and Zhao, Yujie and Zhao, Jishen},journal={arXiv preprint arXiv:2603.10062},year={2026},note={Position paper. Featured on ACM SIGARCH},}
Recent developments in LLM agent systems have shown impressive capabilities for software issue resolution, yet accurately localizing the code to be modified remains a bottleneck. OrcaLoca improves accuracy through priority-based scheduling for LLM-guided action, action decomposition with relevance scoring, and distance-aware context pruning, setting a new state of the art in function-match rate on SWE-bench Lite.
@inproceedings{yu2025orcaloca,title={OrcaLoca: An LLM Agent Framework for Software Issue Localization},author={Yu, Zhongming and Zhang, Hejia and Zhao, Yujie and Huang, Hanxian and Yao, Matrix and Ding, Ke and Zhao, Jishen},booktitle={Proceedings of the 42nd International Conference on Machine Learning},year={2025},}
@inproceedings{tang2023torchsparsepp,title={TorchSparse++: Efficient Training and Inference Framework for Sparse Convolution on GPUs},author={Tang, Haotian and Yang, Shang and Liu, Zhijian and Hong, Ke and Yu, Zhongming and Li, Xiuyu and Dai, Guohao and Wang, Yu and Han, Song},booktitle={Proceedings of the 56th Annual IEEE/ACM International Symposium on Microarchitecture},year={2023},}
@article{zhang2022understanding,title={Understanding GNN Computational Graph: A Coordinated Computation, IO, and Memory Perspective},author={Zhang*, Hengrui and Yu*, Zhongming and Dai, Guohao and Huang, Guyue and Ding, Yufei and Xie, Yuan and Wang, Yu},journal={Proceedings of Machine Learning and Systems},note={* stands for equal contribution},volume={4},pages={467--484},year={2022},}
zhy025 [at] ucsd [dot] edu — happy to talk about agent infrastructure, evaluation, or anything sparse.
