AMLworld-Compact cuts transaction targets needed for model evaluation

Researchers report in a preprint that AMLworld-Compact reduces evaluation targets by 271x and 611x on AMLworld's HI-Small and LI-Small test splits. Evaluating language models on these datasets otherwise requires a separate graph prompt for each of over a million targets. The method retains every illicit test transaction, samples benign transactions by supervised-score strata, and weights results to estimate full-test confusion counts. The researchers released evaluation sets, construction code and audit tools. Recall is exact for fixed predictions, but precision and F1 are ratio estimates, and conservative uncertainty intervals remain broad.