LEV Decision Model Increases Accuracy via Layer Looping

LEV enables decision models to increase reasoning depth by rerunning layers in a loop rather than generating additional text tokens. A 2.6B parameter LEV model improved accuracy from 63.2% to 76.2% on JevBench by increasing loop depth from one to four.