Zora Zhiruo Wang

Wang works in Carnegie Mellon’s Language Technologies Institute with Daniel Fried and Graham Neubig. Her research asks how an agent can adapt as it gains experience and how its behavior matches the work people actually do. She also collaborates with Diyi Yang at Stanford. Her doctoral studies are ongoing; a future completion date in her CV is a plan, not an awarded degree.[1][2]

Before her doctorate, Wang earned a mathematics degree at Beijing Normal University and a master’s in language technologies at Carnegie Mellon. Her industry work includes Microsoft Research Asia, an Amazon Alexa AI internship and consulting for GitHub Next. A visiting student researcher appointment at Stanford ran from May to August 2025. These roles span structured data, conversational systems and code-model evaluation.[2]

Remembering useful workflows

Agent Workflow Memory extracts reusable routines from an agent’s experience and supplies relevant routines during later tasks. Wang is the paper’s first author. The method can prepare workflows from examples or build them while an agent tackles new tasks. Its web-navigation experiments show why memory should be evaluated as an engineering choice: useful experience can shorten a future path, but the reported gains depend on the tasks and comparison systems.[3]

Tools that people can verify

TroVE asks a language model to create, reuse and periodically prune a toolbox of functions. Wang co-authored the method with Fried and Neubig. Instead of generating every solution from low-level operations, the model can assemble programs from useful reusable pieces. The paper evaluates both task accuracy and how readily people can check the resulting programs, making verification part of the design objective.[4]

Learning skills through practice

Wang also co-authored SkillWeaver, which lets a web agent explore a site, practice interactions and turn them into reusable programming interfaces. The resulting skills can be shared with other agents. The study evaluates the approach on WebArena and real websites; it provides evidence for that training setup rather than a guarantee that an agent can safely operate an unfamiliar service.[5][1]

TheAgentCompany places agents in a simulated software company with websites, documents, programs and coworkers. Wang is one of its co-authors. Tasks require agents to use several of those resources, so the benchmark tests a broader workflow than answering a question or editing one function. It is a controlled evaluation of work-related tasks, not evidence that an agent can replace an employee in an actual organization.[6]