LILO: Learning Interpretable Libraries by Compressing and Documenting Code

Watch a research seminar presentation from Harvard CMSA's New Technologies in Mathematics series where MIT CSAIL researcher Gabe Grand introduces LILO, a neurosymbolic framework that revolutionizes code library development through automated synthesis, compression and documentation. Learn how LILO combines large language model-guided program synthesis with symbolic compression to identify optimal lambda abstractions and generate human-readable documentation. Explore how this innovative system outperforms existing methods in solving complex tasks across string editing, scene reasoning, and graphics composition while creating linguistically-grounded, interpretable code libraries. Discover the technical details of LILO's auto-documentation procedure and its impact on both human readability and synthesis performance through practical examples and benchmark evaluations.