Large language models reason well mainly when scaled to hundreds of billions of parameters, leaving smaller models at a disadvantage. Knowledge distillation helps, but the authors of this work argue that even distilled small and mid-scale models stay brittle. They propose Ladders-of-Thought (LoT), a training method that generates a curriculum of reasoning traces that become progressively simpler over time.
Rather than copying outputs from a giant teacher, LoT lets the model itself evolve the curriculum. The idea is to give smaller models a structured path from more demanding reasoning steps to simpler ones, potentially strengthening their abilities where static distillation falls short. The abstract stops short of reporting results, so the effectiveness of the approach remains to be seen.