STEP: From Local Teacher Scaffolds to Student-Owned Reasoning in Small Language Models
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Improving the reasoning abilities of small language models requires successful reasoning experience that can be converted into independent problem-solving ability. Teacher distillation provides correct solutions, but teacher-generated reasoning paths are not always suitable for the student to learn from. Self-training on the student’s own successful traces, in contrast, rarely provides successful experience beyond the student’s current ability. We therefore ask how a small reasoning model can learn to reason beyond its current limits from self-generated experience with the help of temporary external scaffolds. Inspired by instructional scaffolding, we propose STEP (Scaffolded Training through Experience Production), in which a teacher provides a local scaffold at the reasoning obstacle exposed by a student’s failed attempt, allowing the student to complete the remaining reasoning and produce a verified successful trace. During training, STEP includes the scaffold explicitly in the input and concentrates supervision on the scaffold and the student continuation that follows it, turning assisted experience into a learning resource. We evaluate STEP after removing the scaffold. Across two student sizes, two teacher models, and five mathematical reasoning benchmarks, STEP raises the scaffold-free MATH500 mean@4 of Qwen2.5-3B-Instruct from 57.15% to 62.05% in the Qwen3-235B teacher setting, exceeding P-ALIGN by 4.80 percentage points. These results suggest that local teacher scaffolds can turn successful experience that the student could not otherwise generate independently into stronger independent reasoning ability.
- Primary area:
- foundation or frontier models, including LLMs