**TABLELLM: Enabling Tabular Data Manipulation by LLMs in Real Office Usage Scenarios** **Authors:** Xiaokang Zhang, Jing Zhang, Zeyao Ma, Yang Li, Bohan Zhang, Guanlin Li, Zijun Yao, Kangli Xu, Jinchang Zhou, Daniel Zhang-Li, Jifan Yu, Shu Zhao, Juanzi Li, Jie Tang **Institution:** Renmin University of China, Tsinghua University, Anhui University **Abstract:** TABLELLM is a large language model (LLM) designed to handle tabular data manipulation tasks, whether embedded within documents or spreadsheets, catering to real-world office scenarios. The model is trained using a distant supervision method that includes reasoning process extension and cross-way validation strategies to enhance the quality of automatically generated data. TABLELLM outperforms existing general-purpose and tabular data-focused LLMs in both document- and spreadsheet-based tasks. The model checkpoint, source code, benchmarks, and a web application are publicly available. **Keywords:** Large language model, Tabular data manipulation **Introduction:** Tabular data is widely used across various industries, but specific tasks can be laborious and error-prone. TABLELLM addresses these challenges by focusing on a wide range of table operations, including query, update, merge, and chart creation. The model is trained using a combination of existing benchmark data and automatically generated data through extended reasoning processes and cross-way validation. TABLELLM is evaluated using a comprehensive benchmark and evaluation pipeline, demonstrating its effectiveness in both document- and spreadsheet-embedded scenarios. **Related Work:** Previous research has focused on improving LLMs' reasoning capabilities for table question answering (tableQA) and other table-related tasks. However, existing methods often fall short in handling hybrid queries involving text and tabular data in real-world office scenarios. **User Study and Problem Definition:** An extensive user study was conducted to gather insights from office users, focusing on the characteristics of tables and the exploration of table-related tasks. The study revealed preferences for tasks such as tableQA, table revision, chart creation, and data cleaning. The problem definition includes tabular data manipulation in real office usage scenarios, focusing on query, update, merge, and chart operations. **TABLELLM:** TABLELLM is designed to handle a wide array of table operations in document- and spreadsheet-embedded scenarios. The model is trained using a combination of existing benchmark data and automatically generated data through extended reasoning processes and cross-way validation. The training process involves distinct prompts for document- and spreadsheet-embedded tabular data. **Evaluation:** TABLELLM is evaluated using a comprehensive benchmark and evaluation pipeline. The results show that TABLELLM generally surpasses other models in the spreadsheet-embedded scenario and is on par with GPT-3.5 in the document-embedded scenario. The model's performance is further enhanced by**TABLELLM: Enabling Tabular Data Manipulation by LLMs in Real Office Usage Scenarios** **Authors:** Xiaokang Zhang, Jing Zhang, Zeyao Ma, Yang Li, Bohan Zhang, Guanlin Li, Zijun Yao, Kangli Xu, Jinchang Zhou, Daniel Zhang-Li, Jifan Yu, Shu Zhao, Juanzi Li, Jie Tang **Institution:** Renmin University of China, Tsinghua University, Anhui University **Abstract:** TABLELLM is a large language model (LLM) designed to handle tabular data manipulation tasks, whether embedded within documents or spreadsheets, catering to real-world office scenarios. The model is trained using a distant supervision method that includes reasoning process extension and cross-way validation strategies to enhance the quality of automatically generated data. TABLELLM outperforms existing general-purpose and tabular data-focused LLMs in both document- and spreadsheet-based tasks. The model checkpoint, source code, benchmarks, and a web application are publicly available. **Keywords:** Large language model, Tabular data manipulation **Introduction:** Tabular data is widely used across various industries, but specific tasks can be laborious and error-prone. TABLELLM addresses these challenges by focusing on a wide range of table operations, including query, update, merge, and chart creation. The model is trained using a combination of existing benchmark data and automatically generated data through extended reasoning processes and cross-way validation. TABLELLM is evaluated using a comprehensive benchmark and evaluation pipeline, demonstrating its effectiveness in both document- and spreadsheet-embedded scenarios. **Related Work:** Previous research has focused on improving LLMs' reasoning capabilities for table question answering (tableQA) and other table-related tasks. However, existing methods often fall short in handling hybrid queries involving text and tabular data in real-world office scenarios. **User Study and Problem Definition:** An extensive user study was conducted to gather insights from office users, focusing on the characteristics of tables and the exploration of table-related tasks. The study revealed preferences for tasks such as tableQA, table revision, chart creation, and data cleaning. The problem definition includes tabular data manipulation in real office usage scenarios, focusing on query, update, merge, and chart operations. **TABLELLM:** TABLELLM is designed to handle a wide array of table operations in document- and spreadsheet-embedded scenarios. The model is trained using a combination of existing benchmark data and automatically generated data through extended reasoning processes and cross-way validation. The training process involves distinct prompts for document- and spreadsheet-embedded tabular data. **Evaluation:** TABLELLM is evaluated using a comprehensive benchmark and evaluation pipeline. The results show that TABLELLM generally surpasses other models in the spreadsheet-embedded scenario and is on par with GPT-3.5 in the document-embedded scenario. The model's performance is further enhanced by