Inference Plans for Hybrid Particle Filtering

Watch a 19-minute conference talk from POPL 2025 that introduces inference plans, a programming interface enabling developers to control the partitioning of random variables during hybrid particle filtering in probabilistic programming languages. Learn how researchers from MIT, Binghamton University, Université Paris Cité, and IBM developed Siren, a new PPL with annotations for specifying inference plans that the system must implement. Discover how their abstract-interpretation-based static analysis determines inference plan satisfiability with proven soundness. The presentation demonstrates significant performance improvements: speed increases averaging 1.76x (up to 206x) to reach target accuracy, and accuracy improvements averaging 1.83x (up to 595x) with equal or less runtime compared to default heuristics. The talk includes evaluation results showing the static analysis correctly identified all satisfiable inference plans in 27 of 33 benchmark-algorithm settings. Supplementary materials and artifacts are available and have been evaluated as reusable.