Photo of Arjun Radhakrishna

Arjun Radhakrishna

Principal Researcher, Microsoft

I am a researcher on the Excel Agent Science Team at Microsoft. My research interests lie in enabling programmers and end users to offload tasks to automated systems, while being assured of quality, reliability, and safety of the computer's work. Today, this means building reliable and traceable AI agents, evaluating them, and deploying them in products. I currently work on various aspects of Copilot in Excel including harness development, prompting, benchmarking and evaluation, self-improvement, and robustness. Prior to Excel, I worked on automating repetitive code edits in Visual Studio, unburdening programmers from routine and tedious editing tasks.

Before Microsoft, I was a postdoctoral researcher at the University of Pennsylvania with Rajeev Alur. I did my PhD at IST Austria with Thomas A. Henzinger, and my dissertation received the 2014 ACM SIGBED Paul Caspi Memorial Dissertation Award.

Selected Projects

An Excel sheet with a race-geography dashboard built by Copilot, next to the chat pane showing the request to analyze races by continent over the decades and Copilot's summary of the trends it found

Copilot in Excel

Spreadsheets are a demanding setting for an agent: the data is messy, requests are under-specified, and a partially correct result can be worse than none at all. High-stakes decisions, from budgets to financial reports, rest on spreadsheet analysis. This project tackles these challenges in Copilot in Excel, spanning its harness, prompts, skills, benchmarking, self-improvement, and robustness.

Overview of DUET: adaptive task selection, solver and grader agents, and an update step that alternates between revising the grader and the solver

Evaluation and Self-Improvement for Agents

How do we evaluate agents in complex settings? When the task involves back-and-forth with a human? When there is no single right answer? And once we can measure how good an agent is, can we automatically improve its prompts and harnesses? Our work ranges from rubric-based evaluation of human–AI conversations and interactive data-analysis benchmarks to methods that learn agent instructions, knowledge, and skills from feedback.

The Visual Studio lightbulb menu offering an IntelliCode suggestion based on recent edits, next to a preview that replaces .NotNull().NotEmpty() with NotNullOrEmpty

Blue Pencil and IntelliCode

Programmers shouldn't have to tell edit suggestion tools what they want automated; the tools should notice. In a paradigm we call modeless synthesis, the tool watches a programmer's edits and learns edit scripts that offer to finish any repetitive change in progress. Shipped in Visual Studio as IntelliCode suggestions, this work automated over a million edits a day at peak.

A loop in which annotations train an ML model, the model proposes candidate programs, these are combined into a disjunctive program, and a semi-automated annotator produces new annotations

Program Synthesis

Program synthesis is the task of automatically generating or repairing programs, starting from a specification. This project spans both new synthesis algorithms and applications in domains such as data extraction, concurrency, and systems programming. EUSolver, a tool that came out of this line of work, won the Syntax-Guided Synthesis competition in 2016 and 2017.

Recent Publications

    All publications →