Chain-of-Thought (CoT) prompting enhances the reasoning abilities of Large Language Models (LLMs), enabling them to solve complex reasoning tasks step by step.
Challenge: These advanced reasoning capabilities only appear in models with billions of parameters, creating barriers for users limited to Small Language Models (SLMs).
Proposed solution: This paper introduces an Instruction-tuning-CoT method built around CoT demonstrations:
- Transfers reasoning abilities from large LLMs to smaller SLMs
- Enables multi-step, controlled reasoning answers via CoT
- Leverages outputs from robust models to guide smaller ones
Results: On question-answering and mathematical reasoning benchmarks, SLMs trained with this method outperform baselines in both in-domain and out-of-domain settings.
Reference: ACL Anthology Paper