Enhancing the Reasoning Capabilities of Small Language Models via Solution Guidance Fine-Tuning

TL;DR

This paper introduces Solution Guidance Fine-Tuning (SGFT), a novel approach that enhances small language models' reasoning abilities by training them to generate problem-solving guidance, improving performance on reasoning tasks with minimal data.

Contribution

The paper proposes a new reasoning strategy and a plug-and-play fine-tuning paradigm that significantly boosts small language models' reasoning capabilities using limited training data.

Findings

SGFT improves reasoning accuracy of small models

Method enables flexible, prompt-based problem solving

Significant performance gains on reasoning benchmarks

Abstract

Large language models (LLMs) have demonstrated remarkable performance across a wide range of tasks. Advances in prompt engineering and fine-tuning techniques have further enhanced their ability to address complex reasoning challenges. However, these advanced capabilities are often exclusive to models exceeding 100 billion parameters. Although Chain-of-Thought (CoT) fine-tuning methods have been explored for smaller models (under 10 billion parameters), they typically depend on extensive CoT training data, which can introduce inconsistencies and limit effectiveness in low-data settings. To overcome these limitations, this paper introduce a new reasoning strategy Solution Guidance (SG) and a plug-and-play training paradigm Solution-Guidance Fine-Tuning (SGFT) for enhancing the reasoning capabilities of small language models. SG focuses on problem understanding and decomposition at the semantic and logical levels, rather than specific computations, which can effectively improve the SLMs' generalization and reasoning abilities. With only a small amount of SG training data, SGFT can fine-tune a SLM to produce accurate problem-solving guidances, which can then be flexibly fed to any SLM as prompts, enabling it to generate correct answers directly. Experimental results demonstrate that our method significantly improves the performance of SLMs on various reasoning tasks, enhancing both their practicality and efficiency within resource-constrained environments.