Zhengxu (Jason) Yan

Supervised by Dr. Chelsea Finn

• Information-Theoretic Data Curation for Vision-Language-Action Models
• Co-led the design and development of observability-aware data curation methods for vision-language-action models, from research formulation and method design to implementation and evaluation.
• Co-designed the scoring framework and trajectory-level aggregation methods to identify demonstrations with insufficient task-relevant visual information.
• Implemented scoring models and trained and evaluated policies on curated demonstration subsets against standard baselines.
• Extensions to Feedback Descent
• Proposed and implemented variants of Feedback Descent that improved performance over the original method.

Supervised by Dr. Diyi Yang and Dr. Alexander Spangher

• Independently developed Dataset Contribution Units (DCUs), a typed representation of how a dataset extends prior work.
• Defined contribution dimensions spanning task, domain, annotation, scale, evaluation, and availability.
• Developed and evaluated type-routed retrieval and support-aware reranking to connect contribution claims with supporting or conflicting prior evidence.
• Built the paper bank, extraction pipeline, retrieval benchmark, and longitudinal-analysis tools for auditing NLP dataset contributions and adoption.

Supervised by Dr. Ion Stoica

• Led a five-person team developing DiT-Serve, an efficient serving system for diffusion transformers.
• Integrated Ring Attention and Brick Attention for distributed long-sequence video generation.
• Designed continuous batching and request scheduling for mixed-length, mixed-resolution workloads.
• Built shared model interfaces and multi-GPU benchmarks for memory, latency, and throughput.
• Contributed to DiT-Serve's system design, implementation, and experimental evaluation.

Supervised by Dr. Yehuda Kalay and Dr. Davide Schaumann

• Led a three-person team developing a multi-objective reinforcement-learning framework for adaptive building operations.
• Formulated cardiac-lab control as joint optimization over environmental and operational constraints.
• Built the simulation and implemented PPO and Q-learning controllers for multi-dimensional adaptation.
• Directed experiments and analysis for the resulting Journal of Building Engineering publication.

Supervised by Dr. Stefano Bertozzi

• Led ML engineering for Rapid Reviews: Infectious Diseases, an automated preprint-triage system.
• Developed SciBERT- and LLM-based models for scope classification and conflict-of-interest screening.
• Designed human-in-the-loop workflows for domain experts to inspect and correct model outputs.
• Directed teams building ingestion, deduplication, and metadata normalization across preprint repositories.
• Collaborated with Berkeley Public Health and UCSF researchers on deployment in review workflows.

Supervised by Dr. Dawn Song

• Co-developed VERINA, an ICLR 2026 benchmark for machine-verifiable Lean code, specifications, and proofs.
• Built compiler-guided generation and repair pipelines with executable evaluation for each output type.
• Extended the work to Vero, the first repository-level benchmark for joint implementation and proof synthesis.
• Contributed benchmark curation, dependency-aware evaluation, and audits for flawed specifications or reference code.

Supervised by Dr. Valerie Jones

• Developed a human-in-the-loop NLP method for analyzing older adults' voice-assistant interactions.
• Designed annotation and error-review workflows that let domain researchers refine model rules.
• Contributed to empirical analysis and two peer-reviewed journal publications.