Corbin Rosset, also published as Corby Rosset, is a researcher at Microsoft Research. He works on evaluating computer-use agents and on search systems that help people ask and answer useful questions.[1][2][3]
Start with Your LLM Judge Is a Confident Liar: Building Better Verifiers →
X Career
Rosset identifies his current role as researcher at Microsoft Research, following work with Microsoft’s Turing team and You.com. He publishes under the name Corby Rosset. His research spans conversational search and the evaluation of agents that act through a browser, connecting the usefulness of an answer with evidence that the surrounding task actually succeeded.[1][2][3]
Contributions
Check the environment rather than the agent’s confidence
Rosset is first author of the Universal Verifier paper, a collaboration between Microsoft Research and Browserbase. The system generates a rubric, selects screenshots relevant to each criterion and scores the evidence. It distinguishes the quality of the process from whether the requested outcome was achieved, and separates failures an agent can control from external obstacles such as a CAPTCHA. The team also introduces human-labeled trajectories for evaluating the verifier itself.[5]
Suggest questions that move a search forward
Rosset coauthored research on generating useful conversational follow-up questions for search. The work studies how questions can help a person explore an information need rather than simply return another ranked list of documents. It develops and evaluates methods for selecting questions that fit the current search context.[3]
Ideas
In the joint talk with Miguel González Fernández, Rosset explains how a poor rubric can penalize a task for requirements the user never specified or count one mistake repeatedly. The verifier compares the agent’s claims with the visible environment instead of treating a declaration of completion as proof. Miguel presents Browserbase’s infrastructure perspective; Rosset contributes the Microsoft research and verifier-design account.[4]
Tools, writing & talks
Writing, research & talks
Start here
A joint presentation about rubric design and evidence-based evaluation of browser agents.
Start with the examples of an agent claiming success when the screen shows otherwise. The talk introduces the verification problem without requiring implementation knowledge; the paper gives the full experimental design.
Co-presenter with Miguel González Fernández[4]Browserbase and Microsoft Research present a Universal Verifier for computer-use agents. Rubrics, screenshot evidence ranking and error isolation cut false positives from about half to near zero, and an autoresearch loop tries rebuilding it.
Catalog workOriginal sourceThe Universal Verifier research and human-labeled evaluation benchmark.
First author[5]Research on follow-up questions that support useful conversational search.
Coauthor[3]