About Me

I am a PhD candidate in the Brain & Cognitive Sciences (BCS) department at MIT, where I am advised by Roger Levy and Ev Fedorenko. I am also a member of the GenLM research consortium, where we are building an open-source ecosystem for language model probabilistic programming. This summer, I am working at Apple AIML as a student researcher. I am grateful for my many wonderful mentors and collaborators across these communities.

My research, which is funded by the NSF GRFP and an MIT Presidential Fellowship, draws from diverse disciplines spanning cognitive science, Bayesian machine learning, and NLP. Currently, I'm focused on developing train-time and test-time algorithms for reliably controlling language models. I am particularly interested in programming tasks involving long-horizon planning and sparse reward.

Prior to starting my PhD, I studied computational neuroscience at the University of Michigan, and worked for several years on machine learning applications to neurobiology, including publications in Nature, PNAS, and NeurIPS. I have also previously organized several interdisciplinary workshops including NLRSE @ ACL 2024 and NHLS @ The U.S. National Science Foundation.

Select Projects
‣ Sampling Algorithms & Programming Models for LLMs
    ‣ AWRS [arXiv]: Fast randomized algorithm for constrained decoding as posterior inference.

    ‣ GenLM [ICLR Oral]: Controlling LLM generation via programmable constraints and sequential Monte Carlo.

    ‣ Decoding [GitHub]: An open-source library for compositional language model programs.

‣ Reasoning, Pragmatics, & World Knowledge
    ‣ ProbSem [CogSci]: Pragmatic semantic parsing via LLM-mediated approximate inference.
    ‣ LINC [EMNLP Outstanding Paper]: Combining LLMs with SMT solvers to prove logic problems.
    ‣ EWoK [arXiv]: Benchmarking LLMs on core world knowledge.

‣ AI for Code & Mathematics

    ‣ BrainCode [NeurIPS]: An investigation of how LLMs encode computer programs.

    ‣ HumanMath [NeurIPS Math AI Workshop]: Opinion piece on the communicative role of mathematics.

Open Source
I care about open source and allocate a portion of my time towards community contributions. These have previously included the development and evaluation of code models with the Star Coder project by Hugging Face and Service Now, the first implementation of CFG-guided text generation for the Outlines library by dottxt-ai, and contributions to AI for mathematics with Project Numina.