Table-R1: Region-based Reinforcement Learning for Table Understanding

TL;DR

This work tackles the challenge of tabular question answering by proposing Table-R1, a region-based reinforcement learning framework that integrates minimal table regions into reasoning. It combines Region-Enhanced Supervised Fine-Tuning (RE-SFT) to teach models to locate relevant regions with Table-Aware Group Relative Policy Optimization (TARPO) to jointly optimize region accuracy and answer quality via a decaying, mixed reward and consistency penalty. Across three benchmarks, Table-R1 delivers substantial gains (average ~14.36 points) and dramatically reduces token usage (up to 67.5%), outperforming larger baselines and even GPT-4o in some settings. The approach also introduces TableInstruct-RE for region-annotated CoT data, highlighting the potential of region-aware RL to enhance efficient, structured reasoning in LLMs.

Abstract

Tables present unique challenges for language models due to their structured row-column interactions, necessitating specialized approaches for effective comprehension. While large language models (LLMs) have demonstrated potential in table reasoning through prompting and techniques like chain-of-thought (CoT) and program-of-thought (PoT), optimizing their performance for table question answering remains underexplored. In this paper, we introduce region-based Table-R1, a novel reinforcement learning approach that enhances LLM table understanding by integrating region evidence into reasoning steps. Our method employs Region-Enhanced Supervised Fine-Tuning (RE-SFT) to guide models in identifying relevant table regions before generating answers, incorporating textual, symbolic, and program-based reasoning. Additionally, Table-Aware Group Relative Policy Optimization (TARPO) introduces a mixed reward system to dynamically balance region accuracy and answer correctness, with decaying region rewards and consistency penalties to align reasoning steps. Experiments show that Table-R1 achieves an average performance improvement of 14.36 points across multiple base models on three benchmark datasets, even outperforming baseline models with ten times the parameters, while TARPO reduces response token consumption by 67.5% compared to GRPO, significantly advancing LLM capabilities in efficient tabular reasoning.

Figures (9)

• Figure 1: In Table-R1, we adopt the Col & Row-based Table Region for its structured definition. Compared to the cell-based Table Region, it not only saves input tokens but also preserves the sub-table structure. The specific example demonstrates how the LLM can extract Table Region.
• Figure 2: The framework of Table-R1. In RE-SFT , we incorporate the minimum table region at the correct position in the CoT process, and add prompts in the instruction requiring the LLM to reason it. In TARPO, we introduce mixed reward to balance the intermediate table region result and the final answer result, and use $\alpha_\tau$ to dynamically adjust the weights during the learning process.
• Figure 3: Data statistics of reinforcement learning training on Qwen3-8B.
• Figure 4: A case study for PoT, comparing SFT and RE-SFT.
• Figure 5: Instruction for DP data in TableBench.