Toward Probabilistic Coarse-to-Fine Program Synthesis
Explore a 12-minute conference talk from ACM SIGPLAN's LAFI'24 that introduces a novel approach to program synthesis using probabilistic methods. Delve into the concept of coarse-to-fine program synthesis, where simple probabilistic programs are initially created and then iteratively refined. Learn how this method addresses the challenge of evaluating partial programs during the synthesis process by utilizing probabilistic semantics to assess the likelihood of input-output examples. Discover how this approach can be applied to synthesize both deterministic programs and probabilistic models of conditional output distributions. Examine preliminary evidence supporting the effectiveness of using coarse probabilistic program likelihoods to guide the synthesis process in various scenarios.