Ben Lipkin


Ben Lipkin — CV

📧 E-Mail: lipkinb [at] mit [dot] edu
👾 GitHub: benlipkin
🐦 Twitter: ben_lipkin
🦋 BlueSky: benlipkin

About Me

I am a PhD student in the Brain & Cognitive Sciences (BCS) department at MIT, where I work with Roger LevyEv Fedorenko, and many other wonderful mentors and collaborators. I am grateful for funding support from the NSF GRFP and an MIT Presidential Fellowship. My research sits at the interface of Cognitive Science and AI, with a focus on modular resource-rational neurosymbolic programming. I am working on building robust, reliable, and pragmatic AI systems that can reason efficiently and effectively over natural language, programs, and mathematics, by jointly leveraging LLMs alongside tools from probabilistic and symbolic computation, including formal grammars, probabilistic programs, and theorem provers. Through this line of work, I also contribute to research in semantic parsing and algorithms for structured prediction and sequential inference. See ProbSem and LINC for early outputs of this research program, the latter of which won an outstanding paper award at EMNLP '23. I've also served on the organizing committee for several workshops including Natural Language Reasoning and Structured Explanations (NLRSE) at ACL '24. Prior to starting my PhD, I studied Computational Neuroscience and Complex Systems at the University of Michigan, and worked for several years on machine learning applications to neuroscience, including publications in Nature, PNAS, and NeurIPS.

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 and the first implementation of CFG-guided text generation for the Outlines library by .txt. Most recently, I have been contributing to AI for mathematics with Project Numina, who recently won the AIMO progress prize. I am also developing the Decoding library, a framework that makes it easy to design and implement custom LLM search and inference algorithms, built from a set of pure composable building blocks.

News

© Ben Lipkin

Designed with Mobirise - Get now