I'm a final-year PhD student in Berkeley EECS, where I am advised by Dan Klein and affiliated with Berkeley NLP and BAIR. I was previously funded by the NSF GRFP. Before coming to Berkeley, I was an undergrad at Brown University, where I majored in math and linguistics and was advised by Ellie Pavlick.

I am on the job market for academic and industry positions this upcoming year (2024-2025).

I'm broadly interested in natural language processing and machine learning. My  primary lines of research  involve building agents which use language to reason and communicate with humans. I have worked on:

• Language models as collaborative agents

  My work aims to enable language models to function as "agents" which can collaborate effectively and efficiently with humans. To benchmark performance in this area, we released a new type of task known as decision-oriented dialogue, in which human and computer agents must collaborate in order to solve challenging optimization problems. We also developed a state-of-the-art method for autonomously evaluating and improving language agents at both inference and finetuning time.

  ✧ Decision-Oriented Dialogue for Human-AI Collaboration (TACL 2024)
  ✧ Autonomous Evaluation and Refinement of Digital Agents (COLM 2024)

• Games and AI

  Many early AI breakthroughs, like Deep Blue and AlphaGo, occurred in game-playing domains. What can we learn from these approaches, and can similar techniques generalize to open-ended language tasks? My research has used natural language as a probe to interpret models like AlphaGo and investigated whether self-play can effectively be applied to train better language models. Right now, I'm excited about explaining the decisions made by superhuman game-playing models in a human-interpretable way.

  ✧ Understanding Game-Playing Agents with Natural Language Annotations (ACL 2022)
  ✧ Efficacy of Language Model Self-Play in Non-Zero-Sum Games (Preprint, 2024)

• Detecting LLM-generated text

  We built Ghostbuster, the world's best supervised detector for LLM-generated text. It works by passing documents through a series of weaker language models, running a structured search over combinations of features from these models, and then training a linear classifier on the extracted features. For a more accessible technical overview of our model, please check out our post on the Berkeley AI Research Blog.

  ✧ Ghostbuster: Detecting Text Ghostwritten by Large Language Models (NAACL 2024)

• Computer crossword solving

  We built the Berkeley Crossword Solver, a state-of-the-art system for solving American-style crosswords based on dense passage retrieval, belief propagation, and local search. In conjunction with Dr.Fill, our system was the first computer program to outscore all humans at the American Crossword Puzzle Tournament and was featured in Discover, Wired, New Scientist, Slate, and the BBC. I also co-wrote a pop article about the linguistics of crosswords for The Atlantic.

  ✧ Automated Crossword Solving (ACL 2022)
  ✧ The Unspoken Language of Crosswords (The Atlantic, 2023)

I have also worked on computational psycholinguistics, historical language reconstruction, and interfaces for language documentation. Please see here for a full list of publications, or visit my Google Scholar.