SanSi Paper: Looping Layers Lifts 1.4B Decision Model by 13.5 Points
rohanpaul_ai · x · 2026-10-14
The arXiv paper SanSi: A Looped Typed Decision Model for System 1.5 Thinking (Gan, Lee, Kang; 43 pages) explores the middle ground between one forward pass and generated reasoning: turning a pretrained looped language model into a typed decision model whose layers are recursively applied several times before an option-probability readout.
- Every loop is trained to answer, so one model serves compute budgets from 1 to 8 loops in a single run.
- On 10,027 test decisions from 59 sources it reaches 72.0%: +13.5 points over a same-shape non-looped model, +5.3 over a newer same-size model, and just 1.8 below a model with 3x the parameters.
- On two depth-controlled tasks, loops extrapolate to depths never seen in training, where a larger single-pass model fails.
- As a judge for policy optimization without gold answers, it raises the generator's F1 by 7.7 points.
Related event: SanSi: Looped Layers Boost Small Decision Models(2 posts)→
More from Research
- Revisiting STaR: models self-improve by training only on rationales that reach correct answers — AndrewLampinen · 2026-10-14
- Google researcher: having models write new training data via ICL improves finetuning generalization — AndrewLampinen · 2026-10-14
- Study: model collapse doesn't happen when synthetic data accumulates, even without new real data — AndrewLampinen · 2026-10-14
- Synthetic data cuts loss on real distribution and yields 1.8x data efficiency in constrained pretraining — AndrewLampinen · 2026-10-14
- AlphaNIST robot training: Newton physics plus sparse reward did the heavy lifting — milesmacklin · 2026-10-14
- Modal Trains RL on Mixed Accelerators, Cutting Cost While Matching Homogeneous Stability — charles_irl · 2026-10-14