Inventing and using symbolic tools with large language models

Large language models (LLMs) still struggle to perform many seemingly simple tasks, especially those that require consistent manipulation of numbers or symbols. Fortunately, LLMs have become very capable programmers, allowing them to overcome some of their shortcomings by using symbolic “tools” such as calculators, SAT solvers, and APIs. In this talk, I will discuss three recent projects that have pushed the frontier of this tool-usage paradigm. As a motivating example, I will begin by showcasing how LINC uses a first-order logic solver to help language models become more effective and trustworthy reasoners. I will then turn to the question of whether the models could go beyond using existing tools by also inventing new ones. Building on this idea, I will present LILO, a framework which integrates a symbolic tool-learning module into the LLM to equip it with an ever-growing toolbelt. Finally, I will dive deeper into Stitch, the algorithm used to learn new tools in LILO, highlighting key algorithmic insights which allow it to operate 100-1,000x faster than prior work.

About the Speaker:
Theo X. Olausson (he/him) is a PhD student at MIT, where he works with Armando Solar-Lezama on reliable and steerable foundation models through the lens of program synthesis and neurosymbolic AI. His work has recently appeared in venues such as POPL and EMNLP. Before coming to MIT, Theo graduated from the University of Edinburgh with a Master’s degree in Informatics. He was born and raised in Karlskrona, Sweden.

Location: Analysen and Zoom https://chalmers.zoom.us/j/64432575085 PW: mondays24