Researchers have proposed SanSi, a looped typed decision model that answers a declared question without generating any text. Instead of producing a reasoning chain, a decision head returns a probability for each declared option in a single forward pass — what the authors describe as fast, intuitive "System 1" thinking.
The paper explores the space between a single pass and full generated reasoning through looping: the same layers are applied recursively several times before a single typed readout. Each loop lets the model revise its hidden state before committing to an answer, without emitting a token. The authors call this middle ground System 1.5 thinking.
SanSi converts a pre-trained looped language model into a typed decision model. Option probabilities are read after every loop, and each loop is trained with a proper scoring rule, so a single model serves every compute budget from one loop to eight in one run.
On 10,027 test decisions drawn from 59 sources, SanSi reaches 72.0% accuracy. That is 13.5 points above a non-looped model of the same shape trained with the same recipe, 5.3 points above a newer non-looped model of its size, and 1.8 points below a model with three times the parameters. On two depth-controlled tasks, the loops extend the solvable depth beyond the depths seen during training, where the larger single-pass model fails.
Used as the judge for policy optimization with reinforcement learning — and without gold answers — SanSi raises the generator's F1 by 7.7 points. According to the paper's TL;DR, the 1.4B SanSi comes within 1.8 points of Qwen3.5-4B, a model three times its size, and unlike it solves reasoning chains longer than any seen in training.