New AI method slashes simulation inference time, beats outliers with 100x speed boost

Researchers have introduced a novel, amortised simulation-based inference method that is provably robust to outliers in data—a common problem from faulty instruments or human error. Using a neural approximation of a weighted score-matching loss, it eliminates the need for slow Markov chain Monte Carlo sampling. The approach offers significant computational advantages, with complexity only a fraction of current state-of-the-art methods, enabling faster and more reliable inference across science and engineering.