SanSi: A Looped Typed Decision Model for System 1.5 Thinking

Shuyu Gan, Young-Jun Lee, Dongyeop Kang

arXiv 2610.07730 51 pages Read the paper

The short version

The problem: Many automated decisions need more than a quick first impression, such as applying several policy rules or following a chain of statements. A one-pass model must commit immediately, while asking it to write out reasoning can add time and text that a decision system may not need.
The idea: SanSi lets a model revisit its internal working state several times before returning a choice, without generating words. It produces a decision after every revisit, and training makes those early and later answers useful, so callers can choose how much computation to spend—like checking a calculation again before submitting it.
What it got: On 10,027 test decisions, SanSi got 72.0% right—13.5 percentage points more than a comparable model that decides in one pass. It also handled reasoning chains longer than those seen during training.

Module from this paper

SanSi: Looped Typed Decisions

A single-pass decision model can struggle when evidence must be followed step by step; SanSi repeatedly applies a shared language-model stack and reads trained option probabilities after every loop.

looped modelstyped decisionscalibrationadaptive computation