SanSi: A Looped Typed Decision Model for System 1.5 Thinking
SanSi adds looped System 1.5 thinking to typed decision models, reaching 72% accuracy.
SanSi converts a pretrained looped language model into a typed decision model that returns option probabilities in one forward pass after recursive layer reuse, without generating tokens. Option probabilities are read after every loop and each loop is trained with a proper scoring rule, so one model covers budgets from one to eight loops. On 10,027 test decisions from 59 sources it reaches 72.0% accuracy, 13.5 points above a same-shape non-looped model and 1.8 points below a model with three times the parameters. Loops extend solvable depth beyond training depths, and as a reward judge without gold answers it raises a generator's F1 by 7.7 points.
- SanSi loops a pretrained model before a typed option readout, with no generated tokens.
- Every loop is scored so one model covers budgets from one to eight loops.
- 72.0% accuracy on 10,027 decisions from 59 sources, 13.5 points above a same-shape baseline.
- As an RL judge without gold labels, it raises generator F1 by 7.7 points.
From the original article
1 min read at huggingface.co
Typed decision models answer a declared question without generating text: a decision head returns a probability for each of the declared options in a single forward pass. A single pass is fast, intuitive System 1 thinking. We study what lies between one pass and generated reasoning: looping, in which the same layers are recursively applied several times before one typed readout.
By Shuyu Gan, Young-Jun Lee, Dongyeop Kang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.