LEAP preprint reports faster AI agents with a small action predictor

Researchers report in a preprint that LEAP makes AI agents up to 60% faster in end-to-end wall clock time, with no systematic change in task success. The method trains a 0.6B model to predict the actions a target model will choose, allowing proposed actions to be checked by the target. Its latency framework weighs prediction accuracy and remaining task steps against drafting, verification and tool execution costs. The researchers also report that online training without prior trace collection matches offline training performance. The speed gains depend on those operational costs and how often the predictor agrees with the target.