Guiding Reasoning in Small Language Models with LLM Assistance

Yujin Kim, Euiin Yi, Minu Kim, Se-Young Yun, Taehyeon Kim

Abstract: Small language models (SLMs) typically falter on tasks requiring deep, multi-step reasoning. This paper introduces SMART (Small Reasons, Large Hints), a framework where large language models (LLMs) provide targeted, selective guidance to augment SLM reasoning. Drawing from cognitive scaffolding, SMART uses a score-based mechanism to identify uncertain SLM reasoning steps, triggering LLM correction only when essential. This approach, framing structured reasoning as an optimal policy search, steers SLMs towards correct solutions without exhaustive sampling. On mathematical reasoning datasets, SMART enables SLMs to achieve up to 98.9% of LLM-level performance while reducing LLM token usage by up to 90.0%. Our work paves the way for collaborative use of both SLM and LL to tackle complex reasoning tasks that are currently unsolvable by SLMs alone.