An arXiv preprint introduces LEV, a "looped decision model" that aims to combine fast and deep thinking. The abstract opens by situating standard decision models within dual-process theory, comparing their immediate, single-pass answers to the automatic responses of System 1.
Standard decision models, the abstract explains, answer typed questions about an input by computing one probability per option in a single forward pass, without generating any text. This makes them fast, but the abstract begins to note a limitation: "their depth…" — and the text cuts off there.
Based on the title and the available framing, LEV appears to address that limitation by looping the decision model, allowing it to think more deeply while retaining the speed of a non-text-generating approach. However, the truncated abstract does not describe the looping mechanism or the claimed trade-offs.
Because the source is only the abstract's opening lines, this digest can report the motivation but not the method. The full paper would presumably explain how LEV achieves depth and how it compares to other approaches to fast, deliberative inference.