About
Scaling generative intelligence with no compromise.
My research spans multimodal generative models, low-rank methods, and AI algorithm–system co-design. I build efficient and interpretable techniques that make advanced models faster, leaner, and more deployable.
About Me
I am a Ph.D. student in Electrical and Computer Engineering at Duke University, advised by Prof. Yiran Chen in the Center for Computational Evolutionary Intelligence (CEI). My research spans MLSys, efficient AI, and multimodal AI, with a focus on making LLMs, MLLMs, diffusion models, and world models faster and more practical.
Research
I develop efficient systems and algorithms for large models, spanning low-rank LLM inference, mechanistic interpretability, diffusion acceleration, and scalable optimization.
Selected Projects
DecodeShare: Tracing the Shared Subspace of LLM Decode-Time Decisions
Duke CEI · Mechanistic interpretability and activation steering
DecodeShare identifies compact, task-general subspaces in KV-cached decode-time states and validates their causal influence on LLM decisions through matched decode-only interventions.
FlashSVD: Unified Runtime for Low-Rank Transformer Inference
Duke CEI · LLM systems and GPU kernels
FlashSVD is a unified runtime for SVD-compressed language models that uses rank-aware CUDA/Triton kernels and optimized decoding to reduce peak activation memory by 70.2% and transient memory by 75%.
ZEUS: Accelerating Diffusion Models with Only Second-Order Predictor
Duke CEI · Diffusion and generative systemsYixiao Wang*, Ting Jiang*, Zishan Shao* (* equal contribution)
ZEUS is a training-free second-order framework for ODE-based generative models that uses an interleaved schedule to avoid unstable back-to-back extrapolations while delivering up to 3.2× end-to-end speedup with minimal overhead and fewer than 20 lines of integration code.
Earlier Projects
SADA: Stability-guided Adaptive Diffusion Acceleration
Duke CEI · Diffusion and flow-model acceleration
SADA is a training-free framework that adapts step-wise and token-wise sparsity to each denoising trajectory, achieving at least 1.8× acceleration with minimal fidelity degradation.
ECCD: Enhanced Cyclic Coordinate Descent for Elastic-Net GLMs
Sparstitute, Wake Forest University · Scalable optimization
Scalable Dual Coordinate Descent for Kernel Methods
Sparstitute, Wake Forest University · High-performance computing
Research Experience
Duke University, Center for Computational Evolutionary Intelligence (CEI) · Durham, NC
Ph.D. Student and Researcher · Fall 2024 – Present
Advised by Prof. Yiran Chen. My research spans multimodal generative models, low-rank methods, and AI algorithm–system co-design, with a focus on efficient and interpretable techniques for large-scale deployment.
REASON Lab, Wake Forest University · Winston-Salem, NC
Research Assistant · Spring 2022 – Present
Worked on scalable kernel machine-learning algorithms and communication-avoiding optimization with Dr. Aditya Devarakonda.
Wake Forest IRSC Laboratory · Winston-Salem, NC
Research Assistant · Spring 2022 – Spring 2024
Conducted computer-vision research for rainforest image analysis under Prof. Victor Pauca.
Service & Recognition
- Conference Reviewer — AAAI (2026, 2027), ICLR (2026), and NeurIPS (2026; Main Track and Datasets & Benchmarks Track).
- Workshop Reviewer — AdaptFM @ ICML 2026.
- Research Mentorship — Mentored seven Duke students (one Ph.D., two M.S., and four B.S. students), contributing to an ICML Spotlight paper and a workshop submission. See the Mentorship page.
- Cloudexe GPU Catalyst Fellowship — Lab award (proposal co-author), 2026.
- Outstanding Paper Award — HPCAsia 2025.