LEV Uses Looped Decision Models for Deeper, More Efficient Reasoning

Summary

Decision models answer typed questions by assigning one probability to each option in a single forward pass, but their computation depth is fixed during training. LEV introduces a model that deepens its computation by rerunning its layers in a loop instead of generating reasoning tokens. It builds on a language model pretrained as a loop, reads the answer options after every loop step, and trains all steps jointly, allowing one checkpoint to answer at different depths. Across three looped backbones, accuracy increases with depth, especially on harder questions. A 2.6B-parameter LEV rises from 63.2% accuracy at depth one to 76.2% at depth four on JevBench. A small stopper can choose the depth for each question using a threshold fitted on unlabeled requests, reducing computation by 15% to 24% while costing less than one accuracy point, matching or slightly exceeding a fixed mixture of depths. On a locked test set, the 1.4B and 3B models outperform dense, non-looped models and same-latency cascades by 5.4 to 8.8 percentage points at latencies below 50 milliseconds.