Abstract
Data modeling tools face trade-offs between accuracy, computational efficiency, data efficiency, and model flexibility. Physics-inspired, rigorous likelihood-based approaches, while offering high accuracy and data efficiency, remain limited in practice due to high computational cost, particularly when applied to larger-scale problems. This general limitation is further compounded by reliance on traditionally single-threaded iterative sampling or optimization procedures, which are difficult to scale. Although prior efforts have attempted to parallelize expensive likelihood-based approaches by partitioning data or running multiple sampling replicas in parallel, such strategies fail for algorithms requiring efficient communication between processes. Here, we introduce a fundamentally different strategy: we exploit the parallelism inherent in both likelihood evaluation and posterior sampling, operating on a single shared dataset. Our framework supports frequent yet lightweight inter-thread and inter-processor communication, making it well-suited for modern parallel architectures. Using diffraction-limited single-particle fluorescence tracking as a case study, this approach achieves up to a 50-fold speedup on a single mid-range GPU compared to conventional single-threaded CPU implementations, demonstrating a scalable and efficient solution for high-performance likelihood-based inference.