Lazy Knowledge Compilation for Discrete PPLs

Explore a 17-minute video presentation from the LAFI workshop (January 19, 2025) where researchers from MIT discuss lazy knowledge compilation for discrete probabilistic programming languages (PPLs). The talk revisits lazy evaluation techniques for inference in probabilistic programs, which can significantly reduce the execution space by automatically marginalizing random choices that don't affect program results. The presenters propose a new variant of knowledge compilation that leverages lazy evaluation to improve performance, with correctness proofs based on semantics for lazy, higher-order probabilistic programming. Early experimental results demonstrate significant performance gains, suggesting that combining insights from early PPLs with modern advances can be highly beneficial. This presentation by Maddy Bowers, Alexander K. Lew, Joshua B. Tenenbaum, Vikash K. Mansinghka, and Armando Solar-Lezama was sponsored by ACM SIGPLAN as part of the LAFI workshop at POPL25.