Researchers from multiple universities and research institutions have proposed a recurrent decision model called Lev. Built on a native recurrent language model, Lev runs the same set of decoding layers repeatedly for multiple rounds, outputting option probabilities at each round. Through joint training across rounds, a single set of weights can support answering at any depth. A lightweight stopper then determines the number of loop rounds based on the problem, creating a "System 1.5" mode that balances low latency with greater computational investment for hard problems. Experiments show that with the same Lev weights, accuracy gradually improves as loop depth increases, with the extra computation mainly benefiting hard problems. A stopper trained on the objective of "whether the current answer matches the deepest-round answer" can reduce computation by 15–24% with less than a 1 percentage point loss in accuracy. In the low-latency range of under 50 milliseconds, the smaller Lev achieves higher accuracy than ordinary dense models and cascaded schemes at the same latency. The related paper, code, models, and datasets have all been made public.