Decision models output a single probability per option in one forward pass, resembling the fast System 1 of dual‑process theory: they answer instantly, but their reasoning depth is fixed at training time, so easy and hard questions consume the same compute. Models that reason in generated text can think longer, yet each extra decoded token adds cost.
We introduce LEV (Looped Decision Model), which achieves deeper thinking by rerunning its layers in a loop instead of generating tokens. LEV builds on a language model pretrained as a loop, reads the answer options after every loop step, and trains all steps jointly, allowing a single checkpoint to answer at any depth.
Experiments on three looped backbones show that accuracy improves with depth on the same weights, especially on hard questions. For instance, a 2.6 B LEV rises from 63.2% at depth 1 to 76.2% at depth 4 on JevBench.
We also add an on‑demand depth controller: a small stopper with a threshold fitted on unlabeled requests selects the loop count per question. This saves 15%‑24% of computation while dropping accuracy by less than one point, matching or slightly surpassing a fixed‑depth mixture.
On a locked test set, the 1.4 B and 3 B models are 5.4‑8.8 points more accurate than dense (non‑looped) models and cascades with the same latency below 50 ms.
Review