LEV: Thinking Fast and Deep with Looped Decision Models
AuthorsDingjie Song, Zhiling Yan, Yuxuan Zhang, Lichao Sun
AffiliationsLehigh University · University of British Columbia · Vector Institute
LEV lets a model think longer by looping its layers, improving answers to hard questions while adapting compute to the task.
Key results
Percentage-point accuracy improvement from depth one to depth four, rising from 63.2% to 76.2%.
The reported savings range from 15% to 24%, with under one accuracy point lost.
The reported 5.4 to 8.8 percentage-point advantage over dense models and cascades below 50 ms.
Records used to train the main model set.
What the paper found
LEV gives multiple-choice decision models a way to spend more computation without generating reasoning text: it reruns the same layers of a natively looped language model, scores the answer options after every loop, and jointly trains all depths in one checkpoint. A lightweight stopper then decides whether to continue, using unlabelled requests to learn when the current answer agrees with the deepest one. Across Ouro and Nanbeige4.2 backbones, LEV-2.6B improves from 63.2% to 76.2% accuracy on JevBench as depth increases from one to four, a gain of 13.0 percentage points, concentrated on harder questions. On-demand stopping saves 15% to 24% of computation while costing less than one accuracy point. On locked tests, the 1.4B and 3B models outperform dense models and cascades by 5.4 to 8.8 percentage points at comparable latency below 50 ms. The models are trained on 31,329 records; comparisons include dense Qwen3.5 baselines. The key result is adaptive latent depth: one decision model can trade computation for accuracy question by question, without paying the token-generation cost of chain-of-thought.
Original abstract
Decision models answer typed questions about an input with one probability per option, computed in a single forward pass without generating text. Like the fast System 1 of dual-process theory, they answer at once, but their depth is fixed at training time, so 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. We present LEV, a decision model that thinks deeper by rerunning its layers in a loop instead of generating tokens. LEV builds on a language model pretrained as a loop, reads the options after every loop step and trains all steps jointly, so one checkpoint can answer at any depth. On three looped backbones, accuracy rises with depth on the same weights, mostly on hard questions, and a 2.6B LEV goes from 63.2% at depth one to 76.2% at depth four on JevBench. On-demand depth, set per question by a small stopper with a threshold fitted on unlabelled requests, saves 15% to 24% of the computation at a cost of under one accuracy point, level with or slightly above a fixed mix of depths. On a locked test set, the 1.4B and 3B models are 5.4 to 8.8 points more accurate than dense (non-looped) models and cascades of the same latency below 50 ms.
Read the original paperMore in AI Reasoning
Browse all 45 papers →Inducing Process Supervision from Outcome-Only Reinforcement Learning
Shengda Fan, Xin Cong, Zhong Zhang, Haotian Chen, Yankai Lin
TIPS trains models to judge their own reasoning steps using only outcome feedback, avoiding the need for costly step-by-step annotations.
LeanPRM: Lean-Grounded Process Supervision for Natural Language Proof Verification
Rodney Lafuente-Mercado, Salman Rahman, Jingyan Shen, James Shiffer, Saadia Gabriel, Pavel Izmailov
LeanPRM uses Lean-checked formal proofs to generate scalable training labels that help AI models verify each step of mathematical proofs.
Self-Conditioning Enables Deep Sequential Reasoning in Flow Matching
Zeyang Zhang, Linxiang Gao, Meiqi Gu, Xingyan Chen, Minrui Luo, Kaifeng Lyu, Tianxing He
Self-conditioning lets flow-matching models use more sampling steps for deeper visual reasoning.