Decision models stop answering in one pass: looping the same weights lifts JevBench hard questions, 63.2% to 76.2%
Synopsis
Decision models that answer fixed-option questions usually score every option in one forward pass, with a depth fixed at training time, so easy and hard questions cost the same. LEV swaps in a language-model backbone pretrained as a loop, reads the options after every step and trains all steps jointly, so one checkpoint answers at any depth; on JevBench the 2.6B model rises from 63.2% at depth one to 76.2% at depth four, mostly on hard questions. A small stopper sets depth per question with a threshold fitted on unlabelled requests, saving 15% to 24% of the computation at under one accuracy point.
Interpretation
A decision model can now spend more computation per question on the same weights: each extra loop step re-reads the options, so hard questions can take more steps. Earlier decision models returned one probability per option from a single forward pass with a depth fixed at training time, spending the same computation on easy and hard questions; models that reason in generated text can think longer but pay for every step in decoded tokens. On three looped backbones accuracy rises with depth, mostly on hard questions; the 2.6B model goes from 63.2% at depth one to 76.2% at depth four on JevBench, while the fourth step adds at most one point for the Ouro models.
On-demand depth lets each question decide where to stop, taking more steps on hard items and stopping early on easy ones. A fixed depth picks computation for a whole batch, and early exit by a confidence threshold set on labelled data does not beat a fixed mix of depths at the same cost for decision models. The stopper predicts whether the current answer already agrees with the deepest one, with its threshold fitted on unlabelled requests; on held-out panels it stays within one point of the deepest readout while computing 76% to 85% of its cost, and on the locked test it loses at most 1.0 point.
At low latency, looped decision models are more accurate than dense models and cascades of the same latency. No dense option answers between 22 and 79 ms, and a cascade reads uncertain questions again from the start and needs a second model to host. On the locked test set the 1.4B and 3B models are 5.4 to 8.8 points more accurate than dense models and cascades below 50 ms; counted in floating-point operations every LEV model is at or below the dense options, because each loop step runs the whole backbone.
Perspective
The study covers looped backbones of 1.4B to 3B parameters, the publicly available sizes, and JevBench results rest on its 231 public items. On-demand cost is counted in loop steps, and latency is measured one request at a time on one H200 GPU without fused kernels, graph capture or state caching.
The stopper predicts agreement with the deepest answer rather than correctness; for 12.6% of LevBench questions some shallower depth is right while the deepest is wrong, and such a stopper cannot recover them. Confidence separates the errors that more steps fix from the rest only weakly, which limits how far any stopper can beat the fixed-depth frontier. The depth curve on larger looped backbones has not been measured.