arXiv 2610.20951
LEV: Thinking Fast and Deep with Looped Decision Models
AI22 upvotes
Authors: Dingjie Song, Zhiling Yan, Yuxuan Zhang, Lichao Sun
Overview
The authors introduce LEV, a decision model that increases reasoning depth by rerunning its layers in a loop instead of generating extra tokens. LEV starts from a language model pretrained as a loop, reads the options after each loop step, and trains all steps jointly so one checkpoint can answer at any depth. A small stopper sets the depth per question using a threshold fitted on unlabelled requests.
Introduction
Decision models answer in a single forward pass, so their depth is fixed and they spend the same computation on easy and hard questions. Models that reason in generated text can think longer, but they pay for every extra step in decoded tokens.
Results
On the same weights, accuracy rises with loop depth, and a 2.6B LEV goes from 63.2% at depth one to 76.2% at depth four on JevBench. On-demand depth saves 15% to 24% of computation at a cost of under one accuracy point.
Conclusion
This lets practitioners trade computation for accuracy on each question without generating text, which matters for latency-sensitive systems.