CHAIN-OF-TABLE is a framework that enhances table-based reasoning by integrating tabular data into the reasoning chain as a proxy for intermediate thoughts. It enables large language models (LLMs) to iteratively generate operations and update tables to represent a tabular reasoning chain. This process allows LLMs to dynamically plan the next operation based on previous results, forming a chain that captures structured intermediate results, leading to more accurate predictions. The framework is evaluated on three benchmarks: WikiTQ, FeTaQA, and TabFact, achieving state-of-the-art performance across multiple LLMs. The framework uses atomic table operations such as adding columns, selecting rows/columns, grouping, and sorting, which are common in SQL and DataFrame development. It dynamically plans operations based on the table and question, generating arguments for each operation and executing them to transform the table. The resulting intermediate tables guide the LLM to the correct answer. The framework outperforms generic and program-aided reasoning methods, particularly in handling complex tables. It also demonstrates efficiency by requiring fewer generated samples compared to other methods. The framework is adaptable to different tasks and can be generalized to new datasets with the same prompts. The results show that CHAIN-OF-TABLE significantly improves performance on table-based reasoning tasks, especially for complex tables, and is effective across various LLMs.CHAIN-OF-TABLE is a framework that enhances table-based reasoning by integrating tabular data into the reasoning chain as a proxy for intermediate thoughts. It enables large language models (LLMs) to iteratively generate operations and update tables to represent a tabular reasoning chain. This process allows LLMs to dynamically plan the next operation based on previous results, forming a chain that captures structured intermediate results, leading to more accurate predictions. The framework is evaluated on three benchmarks: WikiTQ, FeTaQA, and TabFact, achieving state-of-the-art performance across multiple LLMs. The framework uses atomic table operations such as adding columns, selecting rows/columns, grouping, and sorting, which are common in SQL and DataFrame development. It dynamically plans operations based on the table and question, generating arguments for each operation and executing them to transform the table. The resulting intermediate tables guide the LLM to the correct answer. The framework outperforms generic and program-aided reasoning methods, particularly in handling complex tables. It also demonstrates efficiency by requiring fewer generated samples compared to other methods. The framework is adaptable to different tasks and can be generalized to new datasets with the same prompts. The results show that CHAIN-OF-TABLE significantly improves performance on table-based reasoning tasks, especially for complex tables, and is effective across various LLMs.