VERINA: Benchmarking Verifiable Code Generation
Built a benchmark and evaluation pipeline for measuring whether generated code is actually correct and verifiable, targeting more reliable code agents.
Read arXivMachine Learning Systems • AI Research
Zhengxu (Jason) Yan
M.S. student at Stanford building feedback-driven learning systems, verifiable code benchmarks, and research pipelines for scientific discovery.
My recent work spans scalable model improvement, structured critique, diffusion transformer serving, and dataset intelligence for NLP. I’m interested in turning ambitious ML ideas into robust systems and clear empirical results.
Stanford, CA
Research: LLMs, ML systems, RL
Current Focus
- Scaling feedback descent and iterative refinement
- Verifiable code generation and benchmark design
- Dataset metadata and claim extraction for NLP papers
Selected Work
Research highlights
Representative papers and projects that best capture the direction of my recent research.
DiT-Serve
Designed serving improvements for diffusion transformers using attention optimizations and continuous batching for efficient high-throughput generation.
NoveltyRank
Explores how to estimate conceptual novelty in AI papers by extracting and modeling structured signals from scientific content.
Read arXivCurrent Roles
What I’m working on now
Current research appointments and the systems questions they’re anchored around.
Graduate Research Assistant
IRIS Lab, Stanford AI Lab
Investigating feedback descent, structured critique models, and multi-round refinement pipelines for scalable model improvement.
Graduate Research Assistant
SALT Lab, Stanford AI Lab
Building large-scale pipelines to extract dataset claims and structured metadata from NLP papers, with a focus on novelty and adoption signals.
Research Assistant
Sky Computing Lab, UC Berkeley
Led attention-system integration and batching strategy work for diffusion transformer serving, pushing toward scalable video generation systems.
ML Engineer Lead
UC Berkeley School of Public Health
Built automation and LLM-assisted review pipelines for infectious disease preprints in collaboration with Berkeley and UCSF researchers.
Selected Projects
Systems + applied ML
A few projects outside formal publications that reflect how I build.
Multi-Agent LLM Trading System
2024Integrated specialized agents, multimodal inputs, layered memory, and retrieval to push beyond standard market prediction baselines.
Pintos Operating System
2023Extended a teaching OS across memory, scheduling, file systems, networking, and security, with a focus on systems fundamentals.
Secure Client Application
2023Built a secure Go client with cryptographic support for authentication, file sharing, and revocation.
Education
Training
Stanford University
Master of Science in Computer Science
University of California, Berkeley
B.A. in Computer Science and Data Science
Selected coursework: optimization, probability for data science, security, operating systems, algorithms, computer vision, AI, machine learning, and LLM agents.