The paper "Decomposition for Enhancing Attention: Improving LLM-based Text-to-SQL through Workflow Paradigm" addresses the challenges of attention diffusion and inadequate performance in complex tasks like text-to-SQL using large language models (LLMs). It proposes a workflow paradigm method to enhance LLMs' attention and problem-solving scope through decomposition. The method includes five sub-modules: Information Determination, Classification & Hint, SQL Generation, Self-Correction, and Active Learning. Each module is designed to focus LLMs' attention, reduce irrelevant information, and improve their performance in text-to-SQL tasks. Extensive experiments on three datasets (Spider Dev, Spider-Realistic, and Bird Dev) demonstrate that the proposed approach outperforms existing methods by a significant margin, achieving 2-3 percentage point improvements and new state-of-the-art results on the Spider Test dataset. The code for the proposed method is available on GitHub. 1. **Information Determination**: Reduces interference information through a two-stage method to enhance LLMs' attention. 2. **Classification & Hint**: Categorizes problems into four types (easy, join, nested, join-nested) and provides different prompts for each type to improve accuracy. 3. **SQL Generation**: Uses question templates and few-shot learning to generate SQL queries, improving precision. 4. **Self-Correction**: Addresses common errors by providing specific prompts to correct mistakes. 5. **Active Learning**: Expands the model's capabilities by learning from error cases. - **Spider Dev**: Achieves 85.4% execution accuracy. - **Spider-Realistic**: Achieves 81.5% execution accuracy. - **Bird Dev**: Achieves significant improvements over existing methods. - **Cost Analysis**: The method consumes less inference time and token usage, making it efficient for real applications. - **Parameter Analysis**: The number of few-shots and the accuracy of the information filter significantly affect the final accuracy. - **Cost Issues**: The decomposition workflow may incur higher costs compared to zero-shot approaches. - **Instability**: LLMs' probabilistic nature can lead to inconsistent or unpredictable results. The paper introduces a comprehensive workflow paradigm for text-to-SQL tasks, demonstrating its effectiveness in enhancing LLMs' performance. The method's ability to focus attention and reduce irrelevant information makes it particularly useful for complex tasks.The paper "Decomposition for Enhancing Attention: Improving LLM-based Text-to-SQL through Workflow Paradigm" addresses the challenges of attention diffusion and inadequate performance in complex tasks like text-to-SQL using large language models (LLMs). It proposes a workflow paradigm method to enhance LLMs' attention and problem-solving scope through decomposition. The method includes five sub-modules: Information Determination, Classification & Hint, SQL Generation, Self-Correction, and Active Learning. Each module is designed to focus LLMs' attention, reduce irrelevant information, and improve their performance in text-to-SQL tasks. Extensive experiments on three datasets (Spider Dev, Spider-Realistic, and Bird Dev) demonstrate that the proposed approach outperforms existing methods by a significant margin, achieving 2-3 percentage point improvements and new state-of-the-art results on the Spider Test dataset. The code for the proposed method is available on GitHub. 1. **Information Determination**: Reduces interference information through a two-stage method to enhance LLMs' attention. 2. **Classification & Hint**: Categorizes problems into four types (easy, join, nested, join-nested) and provides different prompts for each type to improve accuracy. 3. **SQL Generation**: Uses question templates and few-shot learning to generate SQL queries, improving precision. 4. **Self-Correction**: Addresses common errors by providing specific prompts to correct mistakes. 5. **Active Learning**: Expands the model's capabilities by learning from error cases. - **Spider Dev**: Achieves 85.4% execution accuracy. - **Spider-Realistic**: Achieves 81.5% execution accuracy. - **Bird Dev**: Achieves significant improvements over existing methods. - **Cost Analysis**: The method consumes less inference time and token usage, making it efficient for real applications. - **Parameter Analysis**: The number of few-shots and the accuracy of the information filter significantly affect the final accuracy. - **Cost Issues**: The decomposition workflow may incur higher costs compared to zero-shot approaches. - **Instability**: LLMs' probabilistic nature can lead to inconsistent or unpredictable results. The paper introduces a comprehensive workflow paradigm for text-to-SQL tasks, demonstrating its effectiveness in enhancing LLMs' performance. The method's ability to focus attention and reduce irrelevant information makes it particularly useful for complex tasks.