Yichong Zhang

I am a senior undergraduate student at Tsinghua University, studying Mathematics and Physics alongside Electrical Engineering. I spent the 2025–2026 academic year as a visiting undergraduate student at Harvard College.

This past summer, I was fortunate to begin working with Prof. Kunle Olukotun and Prof. Olivia Hsu on testing and verification for coding agents, aiming to make agent-generated code more reliable and better aligned with intended behavior. During my visit to Harvard, I was grateful for the opportunity to begin working with Prof. Vijay Janapa Reddi and Prof. Zishen Wan on systems and architecture for embodied AI. Previously, I had the privilege of working with Prof. Yu Wang at Tsinghua on GPU systems and kernel optimization.

I am currently seeking a PhD position starting in Fall 2027, and am happy to discuss potential research fit and collaboration opportunities.

Research Interests

My research aims to enable agent-driven cross-stack optimization for building efficient and reliable AI systems. My current work spans two complementary directions:

Systems and architecture for emerging AI workloads. I study efficient execution across models, runtimes, and hardware, with a recent focus on embodied AI and its real-world deployment.
Reliable code generation and optimization with coding agents. I study how to make agent-generated code more reliable by improving how human intention is specified, evaluated, and fed back to agents.

Publications

Behavioral Test Diversity uses repair transfer to measure distinct code behaviors and guide test-suite pruning and expansion.

Behavioral Test Diversity: Quantifying Test Suite Diversity Through Agentic Code Repair

Yichong Zhang, Genghan Zhang, Zishen Wan, Kunle Olukotun, Olivia Hsu

Submitted to ICLR 2027 · Under review

More tests do not necessarily provide coding agents with more diverse behavioral specifications. BTD measures diversity by checking whether a repair for one test also satisfies others, and uses this signal to guide test-suite pruning and expansion.

Dyserve compiler and adaptive runtime architecture

Serving Agentic Workflows with a Physical-Plan Compiler and Adaptive Runtime

Jiayi Qian, Yichong Zhang, Hanchen Yang, Chun Tao, Souvik Kundu, Zishen Wan, Tushar Krishna

Submitted to ASPLOS 2027 · Under review

arXiv

Dyserve jointly optimizes model selection, verification policies, and backend assignment for agentic workflows. A profile-guided compiler and an adaptive runtime coordinate these decisions as serving load changes, balancing output quality, latency, and throughput.

FlashOverlap signaling-based overlap compared with other approaches

Efficient and Adaptable Overlapping for Computation and Communication via Signaling and Reordering

Ke Hong, Xiuhong Li, Minxu Liu, Qiuli Mao, Tianqi Wu, Zixiao Huang, Lufang Chen, Zhong Wang, Yichong Zhang, Zhenhua Zhu, Guohao Dai, Yu Wang

EuroSys, 2026

arXiv Code

FlashOverlap reduces communication overhead in multi-GPU workloads by allowing completed output tiles to be communicated while computation continues. Tile-level signaling and data reordering enable fine-grained overlap while reusing existing NCCL communication primitives.

PAROAttention pattern-aware token reordering and visual generation results

PAROAttention: Pattern-Aware ReOrdering for Efficient Sparse and Quantized Attention in Visual Generation Models

Tianchen Zhao, Ke Hong, Xinhao Yang, Xuefeng Xiao, Huixia Li, Feng Ling, Ruiqi Xie, Siqi Chen, Hongyu Zhu, Yichong Zhang, Yu Wang

NeurIPS, 2025

arXiv Project

PAROAttention reorganizes irregular visual attention patterns into hardware-friendly blocks through token reordering. This structure enables effective sparsification and low-bit quantization, accelerating image and video generation while maintaining generation quality.

Academic Service

Reviewer, ICLR 2027
Reviewer, NeurIPS 2026 Workshop Who Verifies the Agents? Toward Reliable Agent Development

Beyond Research

My Little Bilibili Corner

Jun 2023 – Present

Video creator

A GreatPerhaps

Mar 2024 – Jan 2025

Founder · Career interview project

Business Innovation Design (BID)

Sep 2024 – Dec 2024

Tsinghua University · Program member