JT-DA: Enhancing Data Analysis with Tool-Integrated Table Reasoning Large Language Models

TL;DR

This work tackles the limited structural grounding and symbolic reasoning of LLMs in tabular data by introducing JT-DA-8B, a tool-integrated, decoder-only LLM trained on a data-synthesized, taxonomy-driven table reasoning corpus. It combines SFT and RL to internalize tool-augmented reasoning patterns, and enforces a four-stage workflow with table preprocessing, sensing, tool-assisted reasoning, and prompt engineering to improve accuracy and interpretability. A comprehensive data-collection and curation pipeline yields millions of high-quality training samples across 34 subtasks and six capabilities, including multi-step ADA tasks, using TCoT, PoT, and ICoT annotations and a sandboxed execution loop. Empirical results on the open TReB benchmark show JT-DA-8B achieving state-of-the-art performance for comparable-scale models, with notable gains in basic table operations and code-enabled computations, while RL further boosts multi-step reasoning capabilities. The work provides a practical, open-source framework for robust, interpretable table reasoning with potential impact on real-world data analysis tasks across business and research domains.

Abstract

In this work, we present JT-DA-8B (JiuTian Data Analyst 8B), a specialized large language model designed for complex table reasoning tasks across diverse real-world scenarios. To address the lack of high-quality supervision in tabular reasoning scenarios, we construct a comprehensive and diverse training corpus with 34 well-defined table reasoning tasks, by aggregating 29 public table QA datasets and 3 million tables. An automatic pipeline is proposed to generate realistic multi-step analytical tasks involving reasoning patterns. The model is trained upon open-source JT-Coder-8B model, an 8B-parameter decoder-only foundation model trained from scratch. In the training stage, we leverage LLM-based scoring and workflow-aligned filtering to distill high-quality, table-centric data. Both supervised fine-tuning (SFT) and Reinforcement learning (RL) are adopted to optimize our model. Afterwards, a four-stage table reasoning workflow is proposed, including table preprocessing, table sensing, tool-integrated reasoning, and prompt engineering, to improve model interpretability and execution accuracy. Experimental results show that JT-DA-8B achieves strong performance in various table reasoning tasks, demonstrating the effectiveness of data-centric generation and workflow-driven optimization.

Figures (8)

• Figure 2: Overall JT-DA-8B architecture.
• Figure 3: Data architecture overview.
• Figure 4: Data construction pipeline.
• Figure 5: Data cleaning pipeline.
• Figure 6: Complex table dataset synthesization.