SanSi loops a language model for 'System 1.5' typed decisions, +13.5 pts over non-looped baseline
minnesotanlp · hf · 2026-10-09
- Idea: SanSi turns a pre-trained looped LM into a typed decision model that outputs option probabilities in a single pass (System 1); looping recursively reapplies the same layers so hidden states revise before committing—a "System 1.5" regime with zero generated tokens.
- Training: Every loop's readout is trained with a proper scoring rule, so one run serves budgets from 1 to 8 loops.
- Results: 72.0% accuracy on 10,027 test decisions from 59 sources—13.5 pts above a non-looped twin, 5.3 above a newer same-size model, 1.8 below a 3x-larger one; loops extrapolate beyond trained depths; as an RL judge without gold answers it lifts generator F1 by 7.7 pts.
More from Research
- Embedding arithmetic demo: cartoon − animation − comic = diagram — mayfer · 2026-10-23
- Domingos Cites New Result Claiming RL Is 'Basically Useless' vs Supervised Learning — pmddomingos · 2026-10-23
- Wolfram Opens Rule 30 Prizes to AIs: $30,000 for Three Open Problems — cephaloform · 2026-10-23
- Stanford and Databricks use AI agents to reproduce medical findings and test analytical robustness — ChengleiSi · 2026-10-23
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23