TableMind introduces the study of autonomous table reasoning with large language models. We propose a two-stage fine-tuning framework that combines supervised trajectory learning with reinforcement optimization via RAPO, a rank-aware strategy for improving reasoning accuracy. Our agent performs multi-turn tool invocation, executes code in a secure sandbox for precise numerical analysis, and leverages planning and reflection for adaptive strategies.

This repository contains the official code for our paper:

TableMind: An Autonomous Programmatic Agent for Tool-Augmented Table Reasoning > Chuang Jiang, Mingyue Cheng, Xiaoyu Tao, Qingyang Mao, Jie Ouyang, Qi Liu

🚩 News (Sep. 2025): The final version of the paper was polished and successfully submitted to arXiv.

Large Language Models (LLMs) demonstrate impressive capabilities but often struggle with the complex numerical computations and fine-grained operations required for table reasoning. TableMind addresses this by introducing a novel two-stage fine-tuning paradigm, guided by a custom-designed multi-objective reward framework that explicitly shapes autonomous tool use and reflection. Our approach progressively develops:

1. (Stage 1: SFT for Warmup Adaption) Foundational skills through supervised fine-tuning , where the LLM learns to use tools and follow a plan-action-reflect cycle using high-quality reasoning trajectories distilled from expert models, ensuring proper structure and usage patterns.
2. (Stage 2: RFT for Strategic Autonomy) Advanced table reasoning via Reinforcement Fine-tuning, with rewards based on structural correctness, final answer accuracy, and strategic tool interaction. RAPO (Rank-Aware Policy Optimization) enhances reasoning paths by increasing the update weight of high-quality trajectories, guiding the model more consistently toward better and more accurate answers.

• Autonomous Agent with Plan-Action-Reflect Loop: Trains an LLM to perform deliberate, multi-step table reasoning by iteratively planning, generating and executing code, and reflecting on the outcomes to adapt its strategy.
• Two-Stage Training Paradigm: Combines warm-up Supervised Fine-tuning (SFT) to establish foundational tool-use patterns with Reinforcement Fine-tuning (RFT) for advanced, autonomous strategy optimization.
• RAPO: Rank-Aware Policy Optimization: A novel reinforcement learning algorithm that increases the update weight of high-quality reasoning trajectories.
• Multi-Objective Reward Design: Employs a fine-grained reward function composed of three key components—structural correctness, final answer accuracy, and strategic tool interaction—to effectively guide the agent's learning process.

conda env create -f environment.yml

1. Install Sandbox Fusion
2. Run Sandbox Fusion in a tmux session
   Please follow the official installation guide here: SandboxFusion

1. Extract the files under the csv_files directory.
2. Update the CSV file path in the agent_r1/tool/tools/python_tool.py file.

# Run training
bash run.sh