Research note · November 3, 2026

Looping a Decision Model: One Extra Pass Helps, More Does Not

A 4B decision model whose middle layers run several times, trained with LoRA only: one extra pass adds 2–3 points on multi-step reasoning; more passes do not.