Enhance Reasoning of Large Language Models via Macro-Micro Self-Training
Abstract: Decomposing complex problems into smaller stages has proven to be highly effective in enhancing the reasoning capabilities of Large Language Models (LLMs). However, as the reasoning process becomes more intricate, uncertainties and errors tend to accumulate, making it challenging to achieve precise final outcomes. Overcoming this challenge and addressing uncertainty in multi-step reasoning necessitates innovative approaches. In this regard, we propose a novel macro-micro self-training method. Our approach leverages self-evaluation and self-modification to enable LLMs to continuously refine their outputs. Through self-evaluation, LLMs assess the accuracy of their generated outputs, while the critical aspect of self-modification allows for iterative refinement of these outputs. To ensure comprehensive refinement, we combine macro evaluation and modification of the entire code structure with micro analysis, where each line of code is individually assessed and refined in line with the problem statement. This dual approach ensures coherent handling of both syntax and semantics. Empirically, our results demonstrate the effectiveness of our approach, as it outperforms existing methods across all settings. Our method enables LLMs to achieve new levels of reasoning capability, providing superior performance in various tasks.
Paper Type: Long
Research Area: Generation
Research Area Keywords: Reasoning, Self-Evaluation, Self-Training
Languages Studied: English
Submission Number: 5088
Loading