SanSi: A Looped Typed Decision Model for System 1.5 Thinking

SanSi. A Looped Typed Decision Model for System 1.5 Thinking. By Shuyu Gan and two others, from the University of Minnesota. Published on the sixth of October. Its code repository is public and real: about twenty-five source files.

A typed decision model answers a declared question without generating any text. A decision head returns a probability for each of the declared options, all in a single forward pass. That single pass is fast and intuitive, the kind of thinking usually called System 1. At the other extreme sits reasoning that is written out token by token. The authors ask what lies between the two.

Their answer is looping. The same layers are applied recursively, several times, before a single typed readout. Each loop lets the model revise its hidden state before it commits to an answer, and it does so without generating a token. The authors call this System one point five thinking. SanSi is their model for it: it turns a pre-trained looped language model into a typed decision model.

The training is what makes SanSi flexible. The option probabilities are read out after every loop, not only the last one, and every loop is trained with a proper scoring rule. As a result a single model serves every compute budget, from one loop up to eight, and all of them come out of one training run. Spending more loops is a choice made at inference time.

The test set has ten thousand and twenty-seven decisions drawn from fifty-nine sources. SanSi reaches seventy-two percent accuracy. That is thirteen point five points above a non-looped model of the same shape trained with the same recipe, and five point three points above a newer non-looped model of its size. It falls one point eight points short of a model with three times the parameters.

Two further results go beyond accuracy. On two depth-controlled tasks, loops extend the depth the model can solve past the depths seen in training, in places where the larger single-pass model fails. And when SanSi is used as the judge for policy optimization with reinforcement learning, with no gold answers available, it raises the generator's F1 by seven point seven points.

The claim is that there is a useful middle register between answering at once and thinking out loud. Looping gives a decision model room to reconsider internally, the per-loop training lets one model choose how much room to take, and the depth results suggest that the extra passes buy something a bigger single pass does not.