A Multi-language Approach to Probabilistic Program Inference
summary
Claims a first probabilistic multilanguage that composes languages with different inference runtimes (example: Dice for discrete, Pyro for continuous) by adding explicit boundaries that lift Dice distributions into deltas and convert Pyro samples into weighted values. Implements a prototype that builds a partially collapsed importance sampler and runs on graph-reachability-with-counting examples, reporting measurable speedups on those classes while an in‑progress proof of correctness ties the boundary bookkeeping to preserved probabilistic semantics.
AI-written summary · about