interpretabilityfactbullishPRISM prototype language models can match or nearly match dense baseline performance while providing interpretable training data influence trackingMachine Learning (Statistics)27 Jul 2026http://arxiv.org/abs/2607.00510v1