Table-R1: Region-based Reinforcement Learning for Table Understanding
Extracted source text · arXiv 2505.12415
\documentclass{article}
% if you need to pass options to natbib, use, e.g.:
% \PassOptionsToPackage{numbers, compress}{natbib}
% before loading neurips_2025
% ready for submission
\usepackage[preprint]{neurips_2025}
% to compile a preprint version, e.g., for submission to arXiv, add add the
% [preprint] option:
% \usepackage[preprint]{neurips_2025}
% to compile a camera-ready version, add the [final] option, e.g.:
% \usepackage[final]{neurips_2025}
% to avoid loading the natbib package, add option nonatbib:
% \usepackage[nonatbib]{neurips_2025}
\usepackage[utf8]{inputenc} % allow utf-8 input
\usepackage[T1]{fontenc} % use 8-bit T1 fonts
\usepackage{hyperref} % hyperlinks
\usepackage{url} % simple URL typesetting
\usepackage{booktabs} % professional-quality tables
\usepackage{amsfonts} % blackboard math symbols
\usepackage{subcaption}
\usepackage{nicefrac} % compact symbols for 1/2, etc.
\usepackage{microtype} % microtypography
\usepackage{enumitem}
\usepackage{times}
\usepackage{latexsym}
\usepackage{graphicx}
\usepackage{multicol}
\usepackage{multirow}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{bbding}
\usepackage{colortbl}
\usepackage{pifont}
\usepackage{makecell}
\usepackage{mathrsfs}
\usepackage{bm}
\usepackage{bbm}
\usepackage{dsfont}
\usepackage{pgf}
\usepackage{fontawesome}
\usepackage{wrapfig}
\usepackage{tcolorbox}
\usepackage{array}
\usepackage{longtable}
\usepackage{caption}
\usepackage{xcolor}
\usepackage{wrapfig}
\newcommand{\TinyEquation}{\fontsize{6.5pt}{\baselineskip}\selectfont}
\newcommand{\SmallEquation}{\fontsize{7.0pt}{\baselineskip}\selectfont}
\newcommand{\MiddleEquation}{\fontsize{9pt}{\baselineskip}\selectfont}
\newcommand{\BigEquation}{\fontsize{9.5pt}{\baselineskip}\selectfont}
\newcommand{\ourmethod}{Table-R1}
\title{\ourmethod{}: Region-based Reinforcement Learning for Table Understanding}
% The \author macro works with any number of authors. There are two commands
% used to separate the names and addresses of multiple authors: \And and \AND.
%
% Using \And between authors leaves it to LaTeX to determine where to break the
% lines. Using \AND forces a line break at that point. So, if LaTeX puts 3 of 4
% authors names on the first line, and the last on the second line, try using
% \AND instead of \And before the third author name.
%\author{%
%Zhenhe Wu\thanks{Use footnote for providing further information
%about author (webpage, alternative address)---\emph{not} for %acknowledging
%funding agencies.} \\
%Beihang University\\
%Cranberry-Lemon University\\
%\texttt{wuzhenhe@buaa.edu.cn} \\
% examples of more authors
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \AND
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
%}
\author{
Zhenhe Wu$^{1,2}$ \quad
Jian Yang$^{1}$\thanks{Corresponding author} \quad
Jiaheng Liu$^{3}$ \quad
Xianjie Wu$^1$ \quad
Changzai Pan$^2$ \quad
Jie Zhang$^{2}$ \\
\textbf{
Jiaxin Peng$^2$ \quad
Yu Zhao$^2$ \quad
Shuangyong Song$^2$ \quad
Yongxiang Li$^{2}$\footnotemark[1] \quad
Zhoujun Li$^1$}
\\
$^1$Beihang University \quad $^2$TeleAI, China Telecom Corp Ltd \quad $^3$Nanjing University\\
\{wuzhenhe, jiaya, wuxianjie, lizj\}@buaa.edu.cn, \ liujiaheng@nju.edu.cn\\
\{panzz, zhangj157, pengjx3, zhaoy11, songshy, liyx25\}@chinatelecom.cn \\
}
\begin{document}
\maketitle
\begin{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 \ourmethod{}, 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 \ourmethod{} 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.
\end{abstract}
\begin{figure*}[tp]
\centerline{\includegraphics[width=\columnwidth]{intro.pdf}}
\caption{In \ourmethod{}, 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.}
\label{table region}
\end{figure*}
\section{Introduction}
% TableLLM介绍
Tables are a widely used data format different from plain text, which poses unique challenges for language models due to their structured row-column interactions~\citep{chain_of_table,table_meet_llm}. Understanding tabular data is crucial for applications like fact verification and question answering, driving significant research interest. Unlike plain text, tables rely on complex row-column interactions, making them challenging for models to interpret. Previous researchers have developed specialized embedding layers, attention mechanisms, and pre-training objectives to enhance structural awareness.
The rise of large language models (LLMs), such as general LLMs GPT-4o/GPT-4.5~\citep{gpt4,gpt45,gpt4o}, has introduced new opportunities for table understanding, as their massive pre-training enables strong performance through prompting alone. Techniques like chain-of-thought (CoT) and program-of-thought (PoT) have further enhanced the reliability of LLMs by equipping responses with reasoning steps. The reasoning LLMs, such as o1/o2/o3~\citep{deepseek_r1,o1}, introduce the step-by-step reasoning trajectories to boost test-time accuracy, which is optimized by reinforcement learning (RL). Deepseek-R1~\citep{deepseek_r1} combines group relative policy optimization~\citep{grpo} (GRPO) and a rule-based reward system to effectively improve the performance of coding and tasks. Besides, a comprehensive and complex table question answering benchmark TableBench~\citep{tablebench} is proposed to evaluate table reasoning capabilities of LLMs.
However, \textit{how to more effectively optimize and scale the table question answering to improve the performance of LLMs in table understanding is still underexplored.}
To unleash the potential of LLMs in table understanding, we explore the approach of rule-based reinforcement learning to significantly strengthen table understanding capabilities by introducing the table structure information. We propose a region-based table question answering (\ourmethod{}) LLM by injecting the region evidence into the reasoning steps and then facilitating the LLM to infer the correct answer using reinforcement learning. Specifically, \ourmethod{} is first trained with the RE-SFT (Region-Enhanced Table SFT) and then enhanced by TARPO (Table-Aware Group Relative Policy Optimization). RE-SFT enhances supervised fine-tuning by guiding LLMs to first identify relevant table regions before generating answers, integrating this step into three reasoning processes, including textual chain-of-thought (TCoT), symbolic chain-of-thought (SCoT), and program-of-thoughts (PoT). TARPO extends reinforcement learning by introducing a mixed reward system that dynamically balances table region accuracy and answer correctness, with a consistency preference to ensure alignment between region identification and answer generation. By decaying the region reward weight over time and penalizing optimization inconsistencies, our approach effectively optimizes both region identification and answer quality, significantly improving table understanding performance.
\ourmethod{} gets the top-tier performance on different benchmarks by a large margin, which demonstrates the promising research of rule-based table reasoning. Our contributions and findings can be summarized as follows:
% \begin{enumerate}
% \item We present the first effective implementation of a rule-based table reinforcement learning approach. We introduce table regions into CoT reasoning and design reward functions based on the accuracy of both table regions and final answers in the responses.
% \item We introduce Table-R2, a unified framework combining RE-SFT and TARPO modules. It synergistically guides LLMs to perform table region selection and answer generation in both supervised fine-tuning and reinforcement learning phases.
% \item The experimental results demonstrate that our Table-R2 framework significantly improves performance and generalization on various benchmark. It achieves an average performance enhancement of 14.36 points while reducing token consumption by 67.5\%. Additionally, we contribute the TableInstruct-RE dataset by enriching CoT with explicit table region annotations on TableInstruct.
% \end{enumerate}
\begin{enumerate}
\item We introduce the first effective implementation of a rule-based table reinforcement learning approach, incorporating table regions into Chain-of-Thought (CoT) reasoning and designing reward functions that evaluate response accuracy based on both table region predictions and final answers.
\item We present \ourmethod{}, a unified framework that integrates RE-SFT (Region-Enhanced Supervised Fine-Tuning) and TARPO (Table-Aware Group Relative Optimization) to synergistically guide large language models (LLMs) in table region selection and answer generation across both supervised fine-tuning and reinforcement learning phases.
\item Our experimental results show that the \ourmethod{} framework delivers significant improvements in performance and generalization across various benchmarks, achieving an average performance gain of 14.36 points while reducing token consumption by 67.5\%. Furthermore, we introduce the TableInstruct-RE dataset, which enhances Chain-of-Thought (CoT) reasoning with explicit table region annotations derived from TableInstruct.
\end{enumerate}
% Rule-based RL 与表格推理 (TableBench)，现有工作的不足
% 我们的方法 area RL
% 主实验结果和分析性实验，列一下contribution 和 insight ?
\section{Method}
\label{sec:method}
\begin{figure*}[tp]
\centerline{\includegraphics[width=\columnwidth]{overall2.pdf}}
\caption{The framework of \ourmethod{}. 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. }
\label{overall}
\end{figure*}
\subsection{Problem Definition}
Table question answering (Table QA) is defined as: Given a semi-structured table $\mathcal{T}$ with $R$ rows and $C$ columns, the objective is to generate an answer $\mathcal{A}$ to a question $\mathcal{Q}$ by leveraging the information within $\mathcal{T}$. Here, $\mathcal{A}$ is a set of values or entities expressed as $\{a_1, a_2, \dots, a_k\}$ ($k \in \mathbb{N}^+$).
\subsection{Table Region Definition}
To concisely and intuitively describe a structured Table Region, we define it as: $T_{reg} = \{[c_1,c_2,...,c_i],[r_1,r_2,..,r_j]\}$. Figure~\ref{table region} illustrates why we use Col \& Row-based structure and present an example of how the LLM extract relevant Table Region. Compared to the cell-based structure, Col \& Row-based structure not only saves input tokens but also preserves the sub-table structure. Though it may inevitably include some irrelevant cells, it still significantly narrows the data-acquisition scope for the LLM when computing answers. In the specific example, by combining the column attributes and cell values of the table with the target question, LLM can select the minimal table region required to generate the final answer. This approach efficiently narrows down the relevant data, focusing on the "Single" column and the corresponding rows that match the query criteria.
\subsection{RE-SFT}
In this section, we present \textbf{RE-SFT}, a Region-Enhanced Supervised Fine-Tuning approach. It aims to steer LLMs toward first identifying the table region relevant to answering questions during the CoT process. Subsequently, by leveraging the data from these identified table regions, the LLMs can compute and arrive at the final answers. We formalize the ICL process as $\mathcal{M}(\mathcal{I}(\mathcal{T}, \mathcal{Q}),\mathcal{E})$, where $\mathcal{I}$ denotes the task-specific instruction, $\mathcal{E}$ represents a few output examples, and $\mathcal{M}$ refers to the LLM.
Tablebench~\citep{tablebench} has summarized four reasoning methods in TableQA: Direct Prompting (DP), Textual Chain-of-Thought (TCoT), Symbolic Chain-of-Thought (SCoT), and Program-of-Thoughts (PoT). We integrate the minimal table region $T_{reg}$ for answering questions into these four methods. For DP, the table region and answer are directly derived simultaneously. The other three methods are adjusted as follows:
%(1) DP: $\mathcal{M}(\mathcal{I}_{DP}(\mathcal{T}, \mathcal{Q}),\mathcal{E}) \rightarrow T_{reg}, \mathcal{A}$. (2) TCoT: $\mathcal{M}(\mathcal{I}_{TCoT}(\mathcal{T}, \mathcal{Q}),\mathcal{E}) \rightarrow \{r_1, r_2, \ldots,r_m,T_{reg}, r_{m+1},\ldots, r_k, \mathcal{A}\}$. (3) SCoT: $\mathcal{M}(\mathcal{I}_{SCoT}(\mathcal{T}, \mathcal{Q}),\mathcal{E}) \rightarrow \{(r_{a_1}, r_{p_1}, r_{s_1}), \ldots, T_{reg}, \ldots, (r_{a_k}, r_{p_k}, r_{s_k}), \mathcal{A}\}$. (4) PoT: $\mathcal{M}(\mathcal{I}_{PoT}(\mathcal{T}, \mathcal{Q}),\mathcal{E}) \rightarrow \{T_{reg},\mathcal{P}\} \rightarrow \mathcal{A}$.
\begin{equation}
TCoT: \mathcal{M}(\mathcal{I}_{TCoT}(\mathcal{T}, \mathcal{Q}),\mathcal{E}) \rightarrow \{r_1, r_2, \ldots,r_m,T_{reg}, r_{m+1},\ldots, r_k, \mathcal{A}\}
\end{equation}
\begin{equation}
SCoT: \mathcal{M}(\mathcal{I}_{SCoT}(\mathcal{T}, \mathcal{Q}),\mathcal{E}) \rightarrow \{(r_{a_1}, r_{p_1}, r_{s_1}), \ldots, T_{reg}, \ldots, (r_{a_k}, r_{p_k}, r_{s_k}), \mathcal{A}\}
\end{equation}
\begin{equation}
PoT: \mathcal{M}(\mathcal{I}_{PoT}(\mathcal{T}, \mathcal{Q}),\mathcal{E}) \rightarrow \{T_{reg},\mathcal{P}\} \rightarrow \mathcal{A}
\end{equation}
where $r_k$ is the $k$-th reasoning step in TCoT. $r_{a_k}$ is the analyzing step, $r_{p_k}$ is the program commands generating step, and $r_{s_k}$ is the result simulation step in SCoT. $\mathcal{P}$ is the generated code in PoT.
The training objective $\mathcal{L}_{\text{all}}$ of RE-SFT on datasets containing all four reasoning methods:
\begin{equation}
\mathcal{L}_{\text{all}} = -\sum_{n=1}^{N} \mathbb{E}_{q^{R_n}, a^{R_n} \sim \{D^{R_n}\}_{n=1}^{N}} \left[ \log P(a^{R_n} | q^{R_n}; \mathcal{M}) \right]
\end{equation}
where \( q^{R_n} \) and \( a^{R_n} \) denote the table-related question and answer from dataset \( D^{R_n} \) corresponding to reasoning method \( R_n \). \( N \) represents the total number of reasoning methods.
\subsection{TARPO}
Recently, reinforcement learning has shown remarkable performance on numerous tasks. In the specific tableQA task, we intend to offer additional rewards to the reasoning process that accurately identifies the relevant table region for better answer generation. Therefore, we have extended GRPO to develop the \textbf{TARPO} (Table-Aware Group Relative Policy Optimization) algorithm, which provides joint incentives for both table region identification and final answer generation in training.
\textbf{Mixed Reward.}\quad The reward $r_i$ for reasoning process $i$ is obtained by weighting the sum of the table region reward $r_i^{t}$ and the answer reward $r_i^{a}$:
\begin{equation}
r_i = \alpha_\tau r_i^{t} + (1-\alpha_\tau) r_i^{a}
\end{equation}
where $r_i^{a}$ is binary, taking a value of 1 when the inferred answer matches the correct answer, and 0 otherwise (Numerical answers match if equal, while string answers match if the Rouge-L score exceeds threshold $\zeta$). $r_i^{t}$ is calculated by computing the IoU (Intersection over Union) separately for the rows and columns between the table region obtained in the reasoning process and the ground truth. The results are then summed and averaged, with $r_i^{t}\in [0,1]$ :
\begin{equation}
r_i^{t} = \frac{IoU_{col}(T_{reg}, \hat{T_{reg}}) + IoU_{row}(T_{reg}, \hat{T_{reg}})}{2}
\end{equation}
where $IOU_{col}$ and $IOU_{row}$ respectively denote the IOU functions specific to rows and columns. $\hat{T_{reg}}$ stands for the predicted table region, while $T_{reg}$ represents the ground truth of the table region.
\textbf{Dynamic Weight of Reward.}\quad At the beginning of reinforcement learning, our goal is to capture table regions more accurately to assist with answer reasoning. However, as training progresses and the model acquires a certain level of table region selection capability, we wish to shift the focus of the reward more towards the accuracy of the answers. To achieve this, we have designed a dynamically changing weight, $\alpha_\tau$ , which decreases with increasing training step $\tau$, the formula is as follows:
\begin{equation}
\alpha_\tau = \gamma e^{-\rho \tau}
\end{equation}
where $\gamma \in [0,1]$ is the initial weight, $\rho > 0$ is the decay coefficient.
\textbf{Consistency Preference.}\quad Following GRPO, we can obtain the normalized group reward score ${A_i}$ by calculating the variance and standard deviation of the rewards. Combining equation (6), we can further decompose ${A}_i$ into two components $\Delta A_i^t$ and $\Delta A_i^a$:
\begin{equation}
\begin{split}
A_i &= \frac{r_i - \text{mean}(\{r_1, r_2, \ldots, r_G\})}{\text{std}(\{r_1, r_2, \ldots, r_G\})}\\&=\underbrace{\frac{\alpha_\tau[r_i^{t} - \text{mean}(\{r_1^{t}, r_2^{t}, \ldots, r_G^{t}\})]}{\text{std}(\{r_1, r_2, \ldots, r_G\})}}_{\Delta A_i^t}+\underbrace{\frac{(1-\alpha_\tau)[r_i^{a} - \text{mean}(\{r_1^{a}, r_2^{a}, \ldots, r_G^{a}\})]}{\text{std}(\{r_1, r_2, \ldots, r_G\})}}_{\Delta A_i^a}
\end{split}
\end{equation}
Since the correct table region can guide better answer generation, we prefer the model to optimize in the direction where both $\Delta A_i^t$ and $\Delta A_i^a$ are positive. If $\Delta A_i^t$ and $\Delta A_i^a$ are of opposite signs, we consider that the reasoning process might be accidental and violates the consistency preference. Therefore, we introduce a small penalty $\mathcal{P}_i$ to mitigate the tendency of optimizing towards strategies (or accelerate the optimization away from it) that do not satisfy consistency:
\begin{equation}
\mathcal{P}_i =
\begin{cases}
0 & \quad \text{if } \Delta r_i^t \cdot \Delta r_i^a >0\\
-\lambda\Delta A_i^t \cdot \Delta A_i^a & \quad\text{if } \Delta r_i^t \cdot \Delta r_i^a < 0
\end{cases}
\end{equation}
Incorporating the penalty term $\mathcal{P}_i$, the objective of TARPO is:
\begin{equation}
\begin{split}
J_{\text{TARPO}}(\theta) = \mathbb{E}_{q \sim P(Q), \{o_i\}_{i=1}^G \sim \pi_{\theta|q}} \Bigg[ \frac{1}{G} \sum_{i=1}^G \min \Bigg( \frac{\pi_{\theta}(o_i|q)}{\pi_{\theta_{\text{old}}}(o_i|q)} & (A_i - \mathcal{P}_i), \\
\text{clip} \left( \frac{\pi_{\theta}(o_i|q)}{\pi_{\theta_{\text{old}}}(o_i|q)}, 1 - \epsilon, 1 + \epsilon \right) (A_i - &\mathcal{P}_i) \Bigg) - \beta D_{\text{KL}}(\pi_{\theta} \|\pi_{\text{ref}}) \Bigg]
\end{split}
\end{equation}
\section{Experimental Setup}
\label{sec:settings}
\begin{wraptable}{r}{9cm}
\begin{center}
\resizebox{0.6 \textwidth}{!}{
\begin{tabular}{cccccc}
\bottomrule
\hline
\specialrule{0em}{1pt}{0pt}\raisebox{-2pt}[0pt][0pt]{
\multirow{2}{*} {Dataset}} & \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Tables}} & \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Samples}} & \multicolumn{3}{c}{Input Token Length} \\
\specialrule{0em}{1pt}{0pt}
\cline {4-6} \specialrule{0em}{1pt}{0pt} & & & \quad min \quad & max & median \\
\midrule
\specialrule{0em}{1pt}{0pt}
TableInstruct & 1.1K & 19,661 & 242 & 7,539 & 755\\
\hline
\specialrule{0em}{1pt}{0pt}
TableBench & 0.59K & 3,544 & 324 & 8,192 & 776\\
WikiTQ (test) & 0.4K & 4,344 & 266 & 2,120 & 691\\
WikiSQL (test) & 5K & 15,878 & 204 & 2,034 & 566\\
\bottomrule
\hline
\end{tabular}}
\end{center}
\caption{Statistics of datasets.}
\label{datasets}
\end{wraptable}
\textbf{Datasets.}\quad We use \textbf{TableInstruct}~\citep{tablebench} as the training set because it encompasses all four reasoning methods, includes four major and 18 sub-question categories with the total size of nearly 20,000. For testing purposes, we employed three benchmark datasets. \textbf{TableBench}~\citep{tablebench} is a comprehensive tabular benchmark designed to evaluate large language models across multiple table-related tasks. \textbf{WikiTQ}~\citep{wikitq} is a dataset for table question answering, featuring 22,033 question-answer pairs with 2,108 tables. \textbf{WikiSQL}~\citep{zhong2017seq2sql} is a dataset that annotates Wiki tables with SQL, which contains 81,000 questions and 24,000 tables. The dataset statistics are summarized in Table~\ref{datasets}.
\textbf{Evaluation Metrics.}\quad We utilize the official evaluation metrics and codes of each benchmark dataset. For TableBench, we apply a combination of \textbf{Exact Match (EM)}, \textbf{Rouge-L}~\citep{rouge} and \textbf{Pass@1}~\citep{pass@1} metrics, while for WikiTQ and WikiSQL, we evaluate \textbf{Accuracy}.
\textbf{Baselines.}\quad We select a range of competitive LLMs similar in size as our baselines. The first category is composed of the most advanced open-source general-purpose LLMs, including \textbf{DeepSeek-Coder-V2-Lite-16B}~\citep{deepseek-v2}, \textbf{Yi Coder-9B-Chat}~\citep{yicoder}, \textbf{Qwen2.5-Coder-7B-Instruct}~\citep{qwen25coder}, \textbf{Qwen2.5-7B-Instruct}~\citep{qwen25} and \textbf{QWQ-32B}~\citep{qwen3}. We also introduce \textbf{GPT-4o}~\citep{gpt4o} as a powerful baseline. The second category comprises LLMs specifically designed for table-related tasks. \textbf{TableLLM-13B}~\citep{tablellm} fine-tuned \textbf{CodeLlama-13B}~\citep{codellama} to handle a variety of table operations in spreadsheet and document settings, with an emphasis on real-world tabular data manipulation. \textbf{TableGPT2-7B}~\citep{tablegpt2} specifically designed a novel table encoder to capture schema-level and cell-level information. To better illustrate the improvement of \ourmethod{}, we introduce a third category of same-base LLMs that use the identical TableInstruct as the training set, including \textbf{TableLLMs}~\citep{tablebench} fine-tuned on \textbf{Qwen2-7B}~\citep{qwen2}, \textbf{CodeQwen-7B}~\citep{qwen}, \textbf{Deepseek-Coder-7B}~\citep{deepseek_coder}, \textbf{Llama3-8B}~\citep{llama3}, \textbf{Llama3.1-8B}~\citep{llama3}, \textbf{Qwen2.5-3B}~\citep{qwen25}, and \textbf{Qwen3-8B}~\citep{qwen3}.
\begin{table*}[tp]
\begin{center}
\resizebox{1 \textwidth}{!}{
\begin{tabular}{cccccccccc}
\bottomrule
\hline
\specialrule{0em}{1pt}{0pt}
\raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Model}}&
\raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Base Model}}& \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Size}} & \multicolumn{4}{c}{TableBench} & \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {WikiTQ}} & \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {WikiSQL}}& \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Overall}}\\
\cmidrule(r){4-7}
&&& DP & TCoT & SCoT & PoT \\
\hline
\specialrule{0em}{1pt}{0pt}
\rowcolor{gray!15}
\multicolumn{10}{c} {general-purpose LLMs} \\
\hline
\specialrule{0em}{1pt}{0pt}
DS-lite & DS-Coder & 16B & 29.60 & 30.93 & 22.61 & 10.90 & 47.65 & 38.30&30.00\\
Yi-Coder & Yi & 9B & 21.94 & 22.80 & 8.43 & 11.36 & 43.37 & 25.34&22.21\\
Qwen2.5-Coder & Qwen2.5 & 7B & {28.67} & {36.25} & {25.95} & 16.15 & {74.50} & 46.90&38.07\\
Qwen2.5-Instruct & Qwen2.5 & 7B & 25.18 & 29.77 & 24.35 & {22.58} & {68.55} & {47.42}&36.14 \\
QWQ & Qwen2.5 & 32B & 43.87 &43.48& 37.06& 31.58 & 70.50 & 47.67& 45.69\\
GPT-4o & - & - & 40.91&51.96&41.43&45.71 & 68.40 & 47.60&49.17\\
\hline
\specialrule{0em}{1pt}{0pt}
\rowcolor{gray!15}
\multicolumn{10}{c} {table-focused LLMs} \\
\hline
\specialrule{0em}{1pt}{0pt}
TableLLM & CodeLlama & 13B & 3.88 & 3.85 & 2.88 & 2.94 & 66.30 & 41.10&20.16 \\
TableGPT2 & Qwen2.5 & 7B & {27.95} & {41.05} & {31.4} & {38.67} & 61.42 & {53.74}&42.20 \\
TableGPT2 & Qwen2.5 & 72B & {38.90} & {50.06} & {30.47} & {28.98} & 71.45 & {57.32}&46.20 \\
\hline
\specialrule{0em}{1pt}{0pt}
\rowcolor{gray!15}
\multicolumn{10}{c} {the same-base LLMs training on TableInstruct} \\
\hline
\specialrule{0em}{1pt}{0pt}
TableLLM & \multirow{3}{*} {Qwen2} & \multirow{3}{*} {7B} & 22.29 & 31.90 & 23.62 & 12.87 & \textbf{64.16} & 46.50&33.56\\
\quad\quad {w/ RE-SFT} & & & \underline{28.24} & \underline{37.33} & \underline{29.10} & \underline{41.05} & 57.55 & \underline{64.42}&\underline{42.95}\\
\textbf{\ourmethod{}} & & &\textbf{36.22} & \textbf{41.85} & \textbf{32.84} & \textbf{41.13} & \underline{61.69} & \textbf{64.80}&\textbf{46.42}\\
\hline
\specialrule{0em}{1pt}{0pt}
TableLLM & \multirow{3}{*} {CodeQwen} & \multirow{3}{*} {7B} & 20.15&24.81&20.55&15.14&36.05&37.20&25.65\\
\quad\quad {w/ RE-SFT} & & & \underline{21.47}&\underline{26.49}&\underline{24.34}&\textbf{42.52}& \underline{46.71}&\underline{57.35}&\underline{36.48}\\
\textbf{\ourmethod{}} & & &\textbf{26.63}&\textbf{28.42}&\textbf{26.10}&\underline{41.15}&\textbf{49.31}&\textbf{59.89}&\textbf{38.58}\\
\hline
\specialrule{0em}{1pt}{0pt}
TableLLM & \multirow{3}{*} {DS-Coder} & \multirow{3}{*} {7B} & 23.15&30.51&23.56&18.74& 36.05&36.14&25.99\\
\quad\quad {w/ RE-SFT} & & &\underline{25.63}&\underline{31.49}&\underline{25.88}&\underline{41.63}&\underline{47.88}&\underline{58.86}&\underline{36.13}\\
\textbf{\ourmethod{}} & & &\textbf{28.74}&\textbf{35.85}&\textbf{29.62}&\textbf{41.71}&\textbf{51.35}&\textbf{60.45}&\textbf{41.29}\\
\hline
\specialrule{0em}{1pt}{0pt}
TableLLM & \multirow{3}{*} {Llama3} & \multirow{3}{*} {8B} & 20.78&29.80&20.78&14.75&34.95&39.26&26.72\\
\quad\quad {w/ RE-SFT} & & &\underline{26.98}&\underline{36.41}&\underline{28.51}&\underline{40.95}&\underline{60.01}&\underline{62.68}&\underline{42.59}\\
\textbf{\ourmethod{}} & & &\textbf{36.06}&\textbf{38.34}&\textbf{36.91}&\textbf{41.56}& \textbf{61.84}& \textbf{63.92}&\textbf{46.44}\\
\hline
\specialrule{0em}{1pt}{0pt}
TableLLM & \multirow{3}{*} {Llama3.1} & \multirow{3}{*} {8B} & 22.30&30.77&21.92&27.17&38.84&39.00&30.00\\
\quad\quad {w/ RE-SFT} & & &\underline{28.71}&\underline{38.36}&\underline{30.46}&\underline{42.14}&\underline{61.03}&\underline{62.92}&\underline{43.94}\\
\textbf{\ourmethod{}} & & &\textbf{38.5}&\textbf{40.21}&\textbf{38.51}&\textbf{42.78}&\textbf{63.51}&\textbf{64.30}&\textbf{47.97}\\
\hline
\specialrule{0em}{1pt}{0pt}
TableLLM & \multirow{3}{*} {Qwen2.5} & \multirow{3}{*} {3B} & 20.08&30.02&22.11&27.55&\underline{55.50}&61.46&36.12\\
\quad\quad {w/ RE-SFT} & & &\underline{22.38}&\underline{32.60}&\underline{25.57}&\underline{41.64}&{51.57}&\underline{61.55}&\underline{39.22}\\
\textbf{\ourmethod{}} & & &\textbf{35.71}&\textbf{39.54}&\textbf{36.78}&\textbf{41.93}&\textbf{56.51}&\textbf{63.98}&\textbf{45.74}\\
\hline
\specialrule{0em}{1pt}{0pt}
TableLLM & \multirow{3}{*} {Qwen3} & \multirow{3}{*} {8B} & 22.42&35.88&27.58&29.15&\underline{73.02}&\underline{71.14}&43.20\\
\quad\quad {w/ RE-SFT} & & &\underline{31.53}&\underline{44.67}&\underline{34.70}&\underline{44.36}&69.06&69.57&\underline{48.98}\\
\textbf{\ourmethod{}} & & &\textbf{44.54}&\textbf{49.30}&\textbf{47.21}&\textbf{44.54}&\textbf{73.87}&\textbf{72.50}&\textbf{55.33}\\
\bottomrule
\hline
\end{tabular}}
\end{center}
\caption{The experimental results on three datasets. We compare \ourmethod{} LLMs with general-purpose LLMs, table-focused LLMs, and same-base LLMs. TableLLM-RE indicates the model built on TableLLM with only RE-SFT training, while \ourmethod{} represents our full framework, including both RE-SFT and TARPO. The experimental results demonstrate the effectiveness of RE-SFT and TARPO and and shows that \ourmethod{} achieves better performance with different base models. }
\label{exp}
\end{table*}
\textbf{Implementation Details.}\quad \textbf{(1) For Datasets:} We employ DeepSeek-R1 to generate the minimal table regions required for solving problems at the critical reasoning points in the original TableInstruct dataset, thereby obtaining the \textbf{TableInstruct-RE} dataset. We also add instructions related to table region generation in the prompts of both the training and test sets. As WikiSQL and WikiTQ lack ICL prompts, we use the modified TCoT prompt template of TableBench for them. \textbf{(2) For RE-SFT:} We achieve supervised fine-tuning on TableInstruct-RE dataset. Following the same setting of TableBench\citep{tablebench}, we use the entire dataset for training purposes and do not partition a validation set. We utilize a cosine annealing scheduler, which sets the initial learning rate at $2e^{-5}$, and conduct training over 3 epochs. Optimization is performed using the Adam optimizer, with a batch size of 512 and a maximum sequence length of 4096. We use 8*A100 GPUs for training, which takes 2 to 2.5 hours for each training. \textbf{(3) For TARPO:} We randomly shuffled the TableInstruct-RE dataset and split it into training and validation sets in a ratio of 9:1. We set the batch size to 32, with the max prompt length and max response length at 8192 and 2048, respectively. The learning rate is $7e^{-7}$, and the group number $G$ for each question is 16. The 7B/8B-sized LLMs are trained on 4*80GB-H100 GPUs, and the 3B-sized LLM on 8*40GB-A100 GPUs. The training lasts for 3 epochs and around 72 hours. For the hyper parameters, we set \{$\zeta$, $\gamma$, $\rho$, $\lambda$\} to \{0.6,
0.3, $9e^{-4}$, 0.1\}.
\begin{table*}[tp]
\begin{center}
\resizebox{1 \textwidth}{!}{
\begin{tabular}{lccccccccc}
\bottomrule
\hline
\specialrule{0em}{1pt}{0pt}
\raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Model}} & \multicolumn{5}{c}{TableBench} & \multicolumn{2}{c}{WikiTQ} & \multicolumn{2}{c}{WikiSQL}\\
\cmidrule(r){2-6}\cmidrule(r){7-8}\cmidrule(r){9-10}
& DP & TCoT & SCoT & PoT & Avg tokens & Acc & Avg tokens & Acc & Avg tokens \\
\hline
\specialrule{0em}{2pt}{0pt}
TableLLM &22.42&35.88&27.58&29.15&\textbf{581}&73.02&\textbf{313}&71.14&\textbf{201}\\
\hline
\specialrule{0em}{2pt}{0pt}
\multirow{2}{*} {w/ RE-SFT} &31.53&44.67&34.70&44.36&1,330&69.06&1,028&69.57&296\\
&\textcolor{blue}{($\uparrow$ 9.81)}&\textcolor{blue}{($\uparrow$ 8.79)}&\textcolor{blue}{($\uparrow$ 7.12)}&\textcolor{blue}{($\uparrow$ 15.21)}&\textcolor{red}{($\uparrow$ 129\%)}&{($\downarrow$ 3.96)}&\textcolor{red}{($\uparrow$ 228\%)}&{($\downarrow$ 1.57)}&\textcolor{red}{($\uparrow$ 47.3\%)}\\
\hline
\specialrule{0em}{2pt}{0pt}
\multirow{2}{*} {\quad w/ GRPO}&42.41&48.64&45.62&42.44&1,549&\underline{73.15}&996&70.24&1,399\\
&\textcolor{blue}{($\uparrow$ 10.88)}&\textcolor{blue}{($\uparrow$ 3.97)}&\textcolor{blue}{($\uparrow$ 10.92)}&{($\downarrow$ 1.92)}&\textcolor{red}{($\uparrow$ 16.5\%)}&\textcolor{blue}{($\uparrow$ 4.09)}&{($\downarrow$ 0.03\%)}&{($\uparrow$ 0.67)}&\textcolor{red}{($\uparrow$ 373\%)}\\
\hline
\specialrule{0em}{2pt}{0pt}
\multirow{2}{*} {\quad w/ TARPO (w/o $\mathcal{P}$)}&\underline{44.43}&\underline{49.15}&\underline{46.88}&\underline{44.20}&630&73.09&416&\underline{70.83}&276\\
&{($\uparrow$ 2.02)}&{($\uparrow$ 0.51)}&{($\uparrow$ 1.26)}&{($\uparrow$ 1.76)}&\textcolor{blue}{($\downarrow$ 59.3\%)}&{($\downarrow$ 0.06)}&\textcolor{blue}{($\downarrow$ 58.2\%)}&{($\uparrow$ 0.59)}&\textcolor{blue}{($\downarrow$ 80.3\%)}\\
\hline
\specialrule{0em}{2pt}{0pt}
\multirow{2}{*} {\quad w/ TARPO (w/ $\mathcal{P}$)}&\textbf{44.54}&\textbf{49.30}&\textbf{47.21}&\textbf{44.54}&\underline{612}&\textbf{73.87}&\underline{397}&\textbf{72.50}&\underline{251}\\
&{($\uparrow$ 0.11)}&{($\uparrow$ 0.15)}&{($\uparrow$ 0.33)}&{($\uparrow$ 0.34)}&{($\downarrow$ 0.03\%)}&\textcolor{blue}{($\uparrow$ 0.78)}&{($\downarrow$ 0.05\%)}&\textcolor{blue}{($\uparrow$ 1.67)}&{($\downarrow$ 0.09\%)}\\
\bottomrule
\hline
\end{tabular}}
\end{center}
\caption{The ablation study of \ourmethod{} using Qwen3-8B as the base model (blue and red color stand for better and worse respectively).}
\label{exp2}
\end{table*}
\section{Results and Analysis}
\label{sec:results}
\subsection{Overall Results}
The overall results of all the baselines and our proposed \ourmethod{} on TableBench, WikiTQ and WikiSQL are shown in Table~\ref{exp}. To more intuitively evaluate the performance of LLMs, we calculate the average score across different test sets as the overall score. In general-purpose LLMs, the Qwen2.5 series models in small-sized models (below 16B) exhibit relatively good comprehensive performance. Specifically, the Qwen2.5-Coder achieves the overall score of 38.07. In contrast, larger models such as GPT-4o, achieve a higher overall score of 49.17 and consistently perform well across various datasets. Among table-focused LLMs, the TableGPT2-7B model stands out among smaller-sized models, achieving the overall score of 42.2, surpassing both general-purpose LLMs and TableLLMs of the same size. Upon increasing the size of the model, TableGPT2-72B attains an even higher overall score of 46.2.
We implement the \ourmethod{} method across different foundational LLM architectures and compare its performance with TableLLMs trained on the same TableInstruct training dataset. Each experimental group consisted of three model configurations: 1) the baseline TableLLM, 2) introducing the RE-SFT strategy through supervised fine-tuning, and 3) the model incorporating both the RE-SFT and TARPO training methodologies (\ourmethod{}). This hierarchical experimental design enables systematic evaluation of performance enhancements attributed to RE-SFT and TARPO. Experimental results demonstrate that RE-SFT integration induces comprehensive performance enhancements across three benchmark datasets, especially in PoT data of TableBench, with only occasional marginal declines ($\leqslant$ 3.96 in score) observed on cross-domain test sets (WikiTQ and WikiSQL). In summary, RE-SFT achieves an average score elevation of 21.27 in PoT and 9.86 in overall when implemented in seven different models. Then, the integration of TAPRO drives further performance improvement. Results in Table~\ref{exp} show the whole \ourmethod{} framework achieves significant improvement on DP, TCoT, and SCoT data of TableBench, WikiTQ and WikiSQL datasets, with a minor improvement on POT data. It achieves an average gain of 4.50 in overall score compared to only using RE-SFT.
In conclusion, \ourmethod{} has achieved state-of-the-art performance across LLMs with different base models. Especially, with the same base model Qwen2.5, 3B-sized \ourmethod{} surpasses the 7B-sized Qwen2.5-Coder, Qwen2.5-Instruct and TableGPT2 in overall score. In addition, \ourmethod{} on 8B-sized Qwen3 surpasses GPT-4o 6.16 in overall scores, only slightly underperforms by 0.76 on the TCoT data of TableBench.
\subsection{Ablation Study}
We employ Qwen3-8B as the base model for ablation study. Table~\ref{exp2} presents the experimental results.
Compared to the baseline TableLLM~\citep{tablebench}, the introduction of RE-SFT significantly enhances the model's performance on TableBench, with a remarkable improvement of 15.21 on the PoT data particularly. However, on the other two out-of-domain test sets WikiTQ and WikiSQL, the performance drops by 3.96 and 1.57 respectively. The incorporation of table regions in CoT also leads to an increase in the average number of response tokens across the three test sets, by 129\%, 228\%, and 47.3\% respectively. Subsequent to applying RE-SFT to the model, we utilize the GRPO algorithm for reinforcement learning. The experimental results show that the performance on DP, TCoT, and SCoT data of Tablebench, along with the WikiTQ data, are significantly enhanced. However, the performance on PoT data slightly decrease, while the effect on WikiSQL is slightly improved. Meanwhile, the problem of increased response tokens due to the introduction of table regions still persists, with an average increase of 16.5\% and 373\% on TableBench and WikiSQL. Then, we substitute GRPO with the TARPO algorithm (without penalty $\mathcal{P}$). The results demonstrate slight performance enhancements on TableBench and WikiSQL, while experiencing a minor decline on WikiTQ. Notably, TARPO achieve a significant reduction in average response tokens compared to GRPO, with decreases of 59.3\%, 58.2\%, and 80.3\% across three datasets. Furthermore, we add the penalty $\mathcal{P}$ into TARPO to achieve consistency preference. The results show a slight improvement on TableBench and more noticeable enhancements on WikiTQ and WikiSQL by 0.78\% and 1.67\%. This indicates that the penalty $\mathcal{P}$ effectively constrain the optimization consistency between table region and final answer in reinforcement training. It strengthens the rationality of the reasoning process by preventing the LLM from learning low-quality reasoning strategies that generate correct answers accidentally. Consequently, it improves the generalization of LLM on out-of-domain datasets.
In conclusion, the whole \ourmethod{} framework achieves optimal performance across all three datasets. Meanwhile, when incorporating table regions, TARPO merely increases the average response tokens by 5.3\%, 26.8\%, and 24.9\% respectively on the three datasets compared to TableLLM, but reduces them by an average of 60.5\%, 60.1\%, and 82.1\% compared to the GRPO, which effectively addresses the excessive token consumption caused by RE-SFT and GRPO.
\begin{figure}[tp]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{reward.png}
\caption{Mean Reward}
\label{fig:figure1}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{train_acc.png}
\caption{Train Accuracy}
\label{fig:figure2}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{val_acc.png}
\caption{Val Accuracy}
\label{fig:figure3}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{global_length.png}
\caption{Global Length (bz=32)}
\label{fig:figure4}
\end{subfigure}
\caption{Data statistics of reinforcement learning training on Qwen3-8B.}
\label{stastics}
\end{figure}
\subsection{Further Analysis}
\label{sec:further}
\textbf{Stastics Analysis.} We compare the data stastics of GRPO, fixed-weight TARPO ($\alpha_\tau=0.15$), and TARPO during reinforcement training. Figure~\ref{stastics} shows the mean reward, train accuracy, valid accuracy, and global token length as step increases. To better observe the overall trends, we apply EMA (Exponential Moving Average) to the three metrics except validation accuracy, with a decay factor of 0.05. It can be seen that TARPO's high initial $\alpha_\tau$ value keeps its mean reward high in the early training phase. However, as $\alpha_\tau$ decreases with step, the final reward of TARPO ends up higher than GRPO but lower than fixed-weight TARPO. This is because fixed-weight TARPO continues to benefit from the reward of the table region part due to the retained $\alpha_\tau$ value at the end of training. Figure~\ref{stastics} (b) and (c) illustrate the accuracy of the current batch and the validation set during training, respectively. TARPO demonstrates better performance compared to the other two methods, indicating both its advantage over GRPO and the effectiveness of dynamic weights. Figure~\ref{stastics} (d) shows the global token length per step. TARPO shows a superior ability in reducing output token length compared to GRPO and also exhibits a more significant reduction than fixed-weight TARPO.
\textbf{Broader impacts.} Our method handles substantial table data, potentially containing sensitive or confidential information. To prevent breaches that could compromise privacy or harm business interests, rigorous encryption, strict access controls, and comprehensive data policies are essential.
\textbf{Limitation.} Our experiments are conducted exclusively on models with 3B to 8B parameters and are limited to maximum input and output lengths of 8192 and 2048 tokens, respectively. The results may vary for larger models or longer token sequences. In addition, for simplicity, we only inserted the table region midway into CoT without re-refining the CoT process, the results may not be optimal.
\section{Related Work}
\label{sec:relatedwork}
\textbf{Table Understanding.} The evolution of Table Question Answering (Table QA) research~\citep{mueller2019answering,alm,jin2022survey,qwen3,execrepobench,qwen25coder} has been propelled by the creation of sophisticated evaluation resources that facilitate semantic parsing capabilities~\citep{DBLP:conf/coling/HuangMLHZ0Z24,soft_template,DBLP:journals/corr/abs-2311-11268, DBLP:journals/corr/abs-2402-11100,qwen,qwen2,codearena,xcoder,xcot}.
Foundational works including WTQ~\citep{pasupat2015compositional}, SQA~\citep{iyyer2017search}, and TabFact~\citep{chen2019tabfact} established initial evaluation paradigms through Wikipedia-derived HTML table QA pairs. While these resources provide structured testing grounds, their answer formulations predominantly depend on direct cell retrieval, limiting their capacity to emulate the multifaceted information needs observed in practical applications.
Recent initiatives address this limitation through innovative dataset designs. ToTTo~\citep{parikh2020totto} and FeTaQA~\citep{nan2022fetaqa} pioneer open-domain QA frameworks requiring generative responses that synthesize implicit table relationships.
Alternative approaches employ structured supervision signals, exemplified by WikiSQL~\citep{zhong2017seq2sql} and Spider~\citep{yu2018spider}, which utilize logical expressions to cultivate systematic reasoning skills.
\textbf{RL for LLMs.} Recent advancements in Online Reinforcement Learning (RL) frameworks have demonstrated significant improvements in augmenting the inferential capacities of large language models (LLMs), exemplified by architectures like DeepSeek‑R1~\citep{deepseek_r1} and OpenAI‑o1~\citep{o1}. Through cyclical switching between exploratory response generation and strategic knowledge utilization, these frameworks progressively refine output granularity while achieving sustained performance gains. State-of-the-art implementations predominantly leverage policy optimization techniques, including Proximal Policy Optimization (PPO)~\citep{ppo} and the parameter-efficient Group Relative Policy Optimization (GRPO)~\citep{grpo}. The latter methodology eliminates dependency on auxiliary value estimation networks, thereby achieving enhanced computational economy.
% Online Reinforcement Learning (RL) has proven effective in boosting the reasoning abilities of large language models (LLMs), notably DeepSeek‑R1~\cite{deepseek_r1} and OpenAI‑o1~\cite{o1}. By iteratively alternating between exploration and exploitation, these systems generate increasingly detailed responses and steadily enhance overall performance. Most successful pipelines employ policy gradient algorithms such as Proximal Policy Optimization (PPO)~\cite{ppo} and the value-free Group Relative Policy Optimization (GRPO)~\cite{grpo}. Because GRPO dispenses with a separate value network, it is more memory efficient.
% Nonetheless, PPO could be incorporated into a flow matching model through an analogous procedure.
\section{Conclusion}
In this paper, we introduce \ourmethod{}, a novel rule-based table reinforcement learning method. Through two-stage RE-SFT and TARPO training, \ourmethod{} effectively incorporates the extraction of the minimum table region into the reasoning CoT, thereby enhancing the guidance of answer generation. The experimental results show that \ourmethod{} delivers an average improvement of 14.36 points across multiple base models on three benchmarks, even surpassing GPT-4o with 8B-sized Qwen3 base model. Moreover, TARPO reduces response token consumption by an average of 67.5\% compared to GRPO. In future work, we hope \ourmethod{} will inspire more research on rule-based reinforcement learning methods for different domain-specific tasks. We also aim to encourage the exploration of signals beyond final answers in the CoT process that can be used for reward shaping to enhance the quality of reasoning process.
\bibliography{table_r1}
\bibliographystyle{natbib}
\clearpage
\appendix
\section{How RE-SFT enhances the performance of PoT}
\begin{table*}[ht]
\begin{center}
\resizebox{0.7 \textwidth}{!}{
\begin{tabular}{cccccccc}
\bottomrule
\hline
\specialrule{0em}{1pt}{0pt}
\raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Method}} & \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Base Model}}& \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Size}}& \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {FC}}& \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {NR}}& \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {DA}}& \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {VIZ}} & \raisebox{-2pt}[0pt][0pt]{\multirow{2}{*} {Overall}}\\\\
\hline
\specialrule{0em}{1pt}{0pt}
SFT &\multirow{2}{*} {CodeQwen}&\multirow{2}{*} {7B}& 11.46& 18.64& 14.28& 36.00& 15.14\\
RE-SFT &&&64.58&55.42&27.57&40.00&42.52\\
\hline
\specialrule{0em}{1pt}{0pt}
SFT &\multirow{2}{*} {DS-Coder}&\multirow{2}{*} {7B}& 10.42& 23.17& 18.58& 30.00& 18.72\\
RE-SFT &&&66.67& 55.16& 25.01& 36.00& 41.65\\
\hline
\specialrule{0em}{1pt}{0pt}
SFT &\multirow{2}{*} {Llama3}&\multirow{2}{*} {8B}& 12.50& 16.88& 15.02& 28.00& 14.75\\
RE-SFT &&&61.46& 53.40& 26.73& 28.00& 40.95\\
\hline
\specialrule{0em}{1pt}{0pt}
SFT &\multirow{2}{*} {Llama3.1}&\multirow{2}{*} {8B}& 27.08& 38.29& 18.16& 44.00& 27.14\\
RE-SFT &&&60.42& 55.92& 27.20& 32.00& 42.15\\
\hline
\specialrule{0em}{1pt}{0pt}
SFT &\multirow{2}{*} {Qwen2}&\multirow{2}{*} {7B}& 9.38& 13.35& 15.10& 26.00& 12.86\\
RE-SFT &&&61.46& 52.14& 18.08& 22.00& 41.07\\
\hline
\specialrule{0em}{1pt}{0pt}
SFT &\multirow{2}{*} {Qwen2.5}&\multirow{2}{*} {3B}& 38.54& 30.98& 24.48& 34.00& 27.55\\
RE-SFT &&&59.38& 52.90& 29.66& 32.00& 41.64\\
\hline
\specialrule{0em}{1pt}{0pt}
SFT &\multirow{2}{*} {Qwen3}&\multirow{2}{*} {8B}& 42.71& 33.50& 24.52& 42.00& 29.15\\
RE-SFT &&&67.71& 56.93& 29.70& 38.00& 44.36\\
\bottomrule
\hline
\end{tabular}}
\end{center}
\caption{Detailed results on PoT of TableBench. There are four types of questions: Fact Checking, Numerical Reasoning, Data Analysis, and Visualization.}
\label{pot}
\end{table*}
In the main experiment results, RE-SFT significantly enhances the performance of PoT. Consequently, we conduct a detailed classification of questions to investigate the specific improvements in each category. As shown in Figure 4, RE-SFT leads to substantial improvements in Fact Checking and Numerical Reasoning questions. It also results in partial enhancements in Data Analysis. However, it exerts no significant influence on Visualization.
\begin{figure*}[ht]
\centerline{\includegraphics[width=\columnwidth]{PoT_SFT.pdf}}
\caption{A case study for PoT, comparing SFT and RE-SFT.}
\label{POT_SFT}
\end{figure*}
We've conducted a case study on PoT data to investigate how model reasoning evolves after SFT and RE-SFT. Figure~\ref{POT_SFT} illustrates that RE-SFT model initially identifies a narrow table region, then generates code based on it, streamlining the code generation process. This approach reduces errors from ambiguous or incorrect reasoning compared to the SFT model.
\section{LLM Prompt Examples}
\begin{figure*}[ht]
\centerline{\includegraphics[width=\columnwidth]{DP.pdf}}
\caption{Instruction for DP data in TableBench.}
\label{DP}
\end{figure*}
\begin{figure*}[ht]
\centerline{\includegraphics[width=\columnwidth]{TCoT.pdf}}
\caption{Instruction for TCoT data in TableBench (along with WikiTQ and WikiSQL in our experiments).}
\label{TCoT}
\end{figure*}
\begin{figure*}[ht]
\centerline{\includegraphics[width=\columnwidth]{SCoT.pdf}}
\caption{Instruction for SCoT data in TableBench.}
\label{SCoT}
\end{figure*}
\begin{figure*}[ht]
\centerline{\includegraphics[width=\columnwidth]{PoT.pdf}}
\caption{Instruction for PoT data in TableBench.}
\end{figure*}
\begin{figure*}[ht]
\centerline{\includegraphics[width=\columnwidth]{deepseek.pdf}}
\caption{Instructions for DeepSeek-R1 inserting Table Regions in CoT.}
\end{figure*}
\end{document}
\begin{thebibliography}{43}
\expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
\bibitem[{01.AI(2024)}]{yicoder}
01.AI. 2024.
\newblock \href {https://github.com/01-ai/Yi-Coder} {Meet yi-coder: A small but mighty llm for code}.
\bibitem[{Bai et~al.(2023)Bai, Bai, Chu, Cui, Dang, Deng, Fan, Ge, Han, Huang et~al.}]{qwen}
Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu~Han, Fei Huang, et~al. 2023.
\newblock Qwen technical report.
\newblock \emph{arXiv preprint arXiv:2309.16609}.
\bibitem[{Chai et~al.(2024)Chai, Yang, Sun, Guo, Liu, Wang, Liang, Bai, Li, Peng, and Li}]{xcot}
Linzheng Chai, Jian Yang, Tao Sun, Hongcheng Guo, Jiaheng Liu, Bing Wang, Xinnian Liang, Jiaqi Bai, Tongliang Li, Qiyao Peng, and Zhoujun Li. 2024.
\newblock \href {https://doi.org/10.48550/ARXIV.2401.07037} {xcot: Cross-lingual instruction tuning for cross-lingual chain-of-thought reasoning}.
\newblock \emph{arXiv preprint arXiv:2401.07037}, abs/2401.07037.
\bibitem[{Chen et~al.(2021)Chen, Tworek, Jun, Yuan, de~Oliveira~Pinto, Kaplan, Edwards, Burda, Joseph, Brockman et~al.}]{pass@1}
Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique~Pond{\'{e}} de~Oliveira~Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et~al. 2021.
\newblock Evaluating large language models trained on code.
\newblock \emph{CoRR}, abs/2107.03374.
\bibitem[{Chen et~al.(2019)Chen, Wang, Chen, Zhang, Wang, Li, Zhou, and Wang}]{chen2019tabfact}
Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William~Yang Wang. 2019.
\newblock Tabfact: A large-scale dataset for table-based fact verification.
\newblock \emph{CoRR}.
\bibitem[{DeepSeek{-}AI et~al.(2024)DeepSeek{-}AI, Liu, Feng, Wang, Wang, Liu, Zhao, Deng, Ruan, Dai, Guo et~al.}]{deepseek-v2}
DeepSeek{-}AI, Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo~Liu, Chenggang Zhao, Chengqi Deng, Chong Ruan, Damai Dai, Daya Guo, et~al. 2024.
\newblock Deepseek-v2: {A} strong, economical, and efficient mixture-of-experts language model.
\newblock \emph{CoRR}, abs/2405.04434.
\bibitem[{Dubey et~al.(2024)Dubey, Jauhri, Pandey, Kadian, Al-Dahle, Letman, Mathur, Schelten, Yang, Fan et~al.}]{llama3}
Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et~al. 2024.
\newblock The llama 3 herd of models.
\newblock \emph{arXiv preprint arXiv:2407.21783}.
\bibitem[{Guo et~al.(2025)Guo, Yang, Zhang, Song, Zhang, Xu, Zhu, Ma, Wang, Bi et~al.}]{deepseek_r1}
Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et~al. 2025.
\newblock Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.
\newblock \emph{arXiv preprint arXiv:2501.12948}.
\bibitem[{Guo et~al.(2024)Guo, Zhu, Yang, Xie, Dong, Zhang, Chen, Bi, Wu, Li et~al.}]{deepseek_coder}
Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y~Wu, YK~Li, et~al. 2024.
\newblock \href {https://arxiv.org/abs/2401.14196} {Deepseek-coder: When the large language model meets programming--the rise of code intelligence}.
\newblock \emph{arXiv preprint arXiv:2401.14196}.
\bibitem[{Huang et~al.(2024)Huang, Ma, Li, Huang, Zou, Zhang, and Zheng}]{DBLP:conf/coling/HuangMLHZ0Z24}
Shulin Huang, Shirong Ma, Yinghui Li, Mengzuo Huang, Wuhe Zou, Weidong Zhang, and Haitao Zheng. 2024.
\newblock \href {https://aclanthology.org/2024.lrec-main.889} {Lateval: An interactive llms evaluation benchmark with incomplete information from lateral thinking puzzles}.
\newblock In \emph{Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, {LREC/COLING} 2024, 20-25 May, 2024, Torino, Italy}, pages 10186--10197. {ELRA} and {ICCL}.
\bibitem[{Hui et~al.(2024)Hui, Yang, Cui, Yang, Liu, Zhang, Liu, Zhang, Yu, Dang et~al.}]{qwen25coder}
Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Kai Dang, et~al. 2024.
\newblock Qwen2.5-coder technical report.
\newblock \emph{arXiv preprint arXiv:2409.12186}.
\bibitem[{Hurst et~al.(2024)Hurst, Lerer, Goucher, Perelman, Ramesh, Clark, Ostrow, Welihinda, Hayes, Radford et~al.}]{gpt4o}
Aaron Hurst, Adam Lerer, Adam~P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ~Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et~al. 2024.
\newblock Gpt-4o system card.
\newblock \emph{arXiv preprint arXiv:2410.21276}.
\bibitem[{Iyyer et~al.(2017)Iyyer, Yih, and Chang}]{iyyer2017search}
Mohit Iyyer, Wen-tau Yih, and Ming-Wei Chang. 2017.
\newblock Search-based neural structured learning for sequential question answering.
\newblock In \emph{ACL 2017}, pages 1821--1831.
\bibitem[{Jaech et~al.(2024)Jaech, Kalai, Lerer, Richardson, El-Kishky, Low, Helyar, Madry, Beutel, Carney et~al.}]{o1}
Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et~al. 2024.
\newblock Openai o1 system card.
\newblock \emph{arXiv preprint arXiv:2412.16720}.
\bibitem[{Jin et~al.(2022)Jin, Siebert, Li, and Chen}]{jin2022survey}
Nengzheng Jin, Joanna Siebert, Dongfang Li, and Qingcai Chen. 2022.
\newblock A survey on table question answering: recent advances.
\newblock In \emph{China Conference on Knowledge Graph and Semantic Computing}, pages 174--186. Springer.
\bibitem[{Li et~al.(2023)Li, Xu, Chen, Huang, Li, Jiang, Li, Zhou, Zheng, and Shen}]{DBLP:journals/corr/abs-2311-11268}
Yinghui Li, Zishan Xu, Shaoshen Chen, Haojing Huang, Yangning Li, Yong Jiang, Zhongli Li, Qingyu Zhou, Hai{-}Tao Zheng, and Ying Shen. 2023.
\newblock \href {https://doi.org/10.48550/ARXIV.2311.11268} {Towards real-world writing assistance: {A} chinese character checking benchmark with faked and misspelled characters}.
\newblock \emph{CoRR}, abs/2311.11268.
\bibitem[{Li et~al.(2024)Li, Zhou, Luo, Ma, Li, Zheng, Hu, and Yu}]{DBLP:journals/corr/abs-2402-11100}
Yinghui Li, Qingyu Zhou, Yuanzhen Luo, Shirong Ma, Yangning Li, Hai{-}Tao Zheng, Xuming Hu, and Philip~S. Yu. 2024.
\newblock \href {https://doi.org/10.48550/ARXIV.2402.11100} {When llms meet cunning questions: {A} fallacy understanding benchmark for large language models}.
\newblock \emph{CoRR}, abs/2402.11100.
\bibitem[{Lin(2004)}]{rouge}
Chin-Yew Lin. 2004.
\newblock {ROUGE}: A package for automatic evaluation of summaries.
\newblock In \emph{Text Summarization Branches Out}, Barcelona, Spain. Association for Computational Linguistics.
\bibitem[{Mueller et~al.(2019)Mueller, Piccinno, Shaw, Nicosia, and Altun}]{mueller2019answering}
Thomas Mueller, Francesco Piccinno, Peter Shaw, Massimo Nicosia, and Yasemin Altun. 2019.
\newblock Answering conversational questions on structured data without logical forms.
\newblock In \emph{EMNLP-IJCNLP 2019}, pages 5902--5910.
\bibitem[{Nan et~al.(2022)Nan, Hsieh, Mao, Lin, Verma, Zhang, Kry{\'s}ci{\'n}ski, Schoelkopf, Kong, Tang et~al.}]{nan2022fetaqa}
Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi~Victoria Lin, Neha Verma, Rui Zhang, Wojciech Kry{\'s}ci{\'n}ski, Hailey Schoelkopf, Riley Kong, Xiangru Tang, et~al. 2022.
\newblock Fetaqa: Free-form table question answering.
\newblock \emph{TACL 2022}, 10:35--49.
\bibitem[{OpenAI(2023)}]{gpt4}
OpenAI. 2023.
\newblock \href {https://arxiv.org/abs/2303.08774} {Gpt-4 technical report}.
\newblock \emph{arXiv preprint arXiv:2303.08774}.
\bibitem[{OpenAI(2025)}]{gpt45}
OpenAI. 2025.
\newblock \href {https://openai.com/index/introducing-gpt-4-5/} {Introducing gpt-4.5}.
\bibitem[{Parikh et~al.(2020)Parikh, Wang, Gehrmann, Faruqui, Dhingra, Yang, and Das}]{parikh2020totto}
Ankur Parikh, Xuezhi Wang, Sebastian Gehrmann, Manaal Faruqui, Bhuwan Dhingra, Diyi Yang, and Dipanjan Das. 2020.
\newblock Totto: A controlled table-to-text generation dataset.
\newblock In \emph{EMNLP 2020}, pages 1173--1186.
\bibitem[{Pasupat and Liang(2015{\natexlab{a}})}]{wikitq}
Panupong Pasupat and Percy Liang. 2015{\natexlab{a}}.
\newblock Compositional semantic parsing on semi-structured tables.
\newblock In \emph{Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, {ACL} 2015, July 26-31, 2015, Beijing, China, Volume 1: Long Papers}.
\bibitem[{Pasupat and Liang(2015{\natexlab{b}})}]{pasupat2015compositional}
Panupong Pasupat and Percy Liang. 2015{\natexlab{b}}.
\newblock Compositional semantic parsing on semi-structured tables.
\newblock In \emph{ACL 2015}, pages 1470--1480.
\bibitem[{Roziere et~al.(2023)Roziere, Gehring, Gloeckle, Sootla, Gat, Tan, Adi, Liu, Remez, Rapin et~al.}]{codellama}
Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing~Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, J{\'e}r{\'e}my Rapin, et~al. 2023.
\newblock Code llama: Open foundation models for code.
\newblock \emph{CoRR}.
\bibitem[{Schulman et~al.(2017)Schulman, Wolski, Dhariwal, Radford, and Klimov}]{ppo}
John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017.
\newblock Proximal policy optimization algorithms.
\newblock \emph{arXiv preprint arXiv:1707.06347}.
\bibitem[{Shao et~al.(2024)Shao, Wang, Zhu, Xu, Song, Bi, Zhang, Zhang, Li, Wu et~al.}]{grpo}
Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK~Li, Y~Wu, et~al. 2024.
\newblock Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024.
\newblock \emph{URL https://arxiv. org/abs/2402.03300}.
\bibitem[{Su et~al.(2024)Su, Wang, Ye, Zhou, Zhang, Zhu, Wang, Xu, Chen, Li et~al.}]{tablegpt2}
Aofeng Su, Aowen Wang, Chao Ye, Chen Zhou, Ga~Zhang, Guangcheng Zhu, Haobo Wang, Haokai Xu, Hao Chen, Haoze Li, et~al. 2024.
\newblock Tablegpt2: A large multimodal model with tabular data integration.
\newblock \emph{arXiv preprint arXiv:2411.02059}.
\bibitem[{Sui et~al.(2024)Sui, Zhou, Zhou, Han, and Zhang}]{table_meet_llm}
Yuan Sui, Mengyu Zhou, Mingjie Zhou, Shi Han, and Dongmei Zhang. 2024.
\newblock Table meets llm: Can large language models understand structured table data? a benchmark and empirical study.
\newblock In \emph{Proceedings of the 17th ACM International Conference on Web Search and Data Mining}, pages 645--654.
\bibitem[{Wang et~al.(2024)Wang, Zhang, Li, Eisenschlos, Perot, Wang, Miculicich, Fujii, Shang, Lee et~al.}]{chain_of_table}
Zilong Wang, Hao Zhang, Chun-Liang Li, Julian~Martin Eisenschlos, Vincent Perot, Zifeng Wang, Lesly Miculicich, Yasuhisa Fujii, Jingbo Shang, Chen-Yu Lee, et~al. 2024.
\newblock Chain-of-table: Evolving tables in the reasoning chain for table understanding.
\newblock \emph{arXiv preprint arXiv:2401.04398}.
\bibitem[{Wu et~al.(2024)Wu, Yang, Chai, Zhang, Liu, Du, Liang, Shu, Cheng, Sun et~al.}]{tablebench}
Xianjie Wu, Jian Yang, Linzheng Chai, Ge~Zhang, Jiaheng Liu, Xinrun Du, Di~Liang, Daixin Shu, Xianfu Cheng, Tianzhen Sun, et~al. 2024.
\newblock Tablebench: A comprehensive and complex benchmark for table question answering.
\newblock \emph{arXiv preprint arXiv:2408.09174}.
\bibitem[{Yang et~al.(2025{\natexlab{a}})Yang, Li, Yang, Zhang, Hui, Zheng, Yu, Gao, Huang, Lv et~al.}]{qwen3}
An~Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo~Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et~al. 2025{\natexlab{a}}.
\newblock \href {http://arxiv.org/abs/2505.09388} {Qwen3 technical report}.
\bibitem[{Yang et~al.(2024{\natexlab{a}})Yang, Yang, Hui, Zheng, Yu, Zhou, Li, Li, Liu, Huang et~al.}]{qwen2}
An~Yang, Baosong Yang, Binyuan Hui, Bo~Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et~al. 2024{\natexlab{a}}.
\newblock Qwen2 technical report.
\newblock \emph{arXiv preprint arXiv:2407.10671}.
\bibitem[{Yang et~al.(2025{\natexlab{b}})Yang, Yang, Zhang, Hui, Zheng, Yu, Li, Liu, Huang, Wei et~al.}]{qwen25}
An~Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo~Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et~al. 2025{\natexlab{b}}.
\newblock Qwen2.5 technical report.
\newblock \emph{CoRR}, abs/2412.15115.
\bibitem[{Yang et~al.(2020{\natexlab{a}})Yang, Ma, Zhang, Li, and Zhou}]{soft_template}
Jian Yang, Shuming Ma, Dongdong Zhang, Zhoujun Li, and Ming Zhou. 2020{\natexlab{a}}.
\newblock \href {https://doi.org/10.18653/V1/2020.ACL-MAIN.531} {Improving neural machine translation with soft template prediction}.
\newblock In \emph{Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, {ACL} 2020, Online, July 5-10, 2020}, pages 5979--5989. Association for Computational Linguistics.
\bibitem[{Yang et~al.(2020{\natexlab{b}})Yang, Ma, Zhang, Wu, Li, and Zhou}]{alm}
Jian Yang, Shuming Ma, Dongdong Zhang, Shuangzhi Wu, Zhoujun Li, and Ming Zhou. 2020{\natexlab{b}}.
\newblock \href {https://doi.org/10.1609/AAAI.V34I05.6480} {Alternating language modeling for cross-lingual pre-training}.
\newblock In \emph{The Thirty-Fourth {AAAI} Conference on Artificial Intelligence, {AAAI} 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, {IAAI} 2020, The Tenth {AAAI} Symposium on Educational Advances in Artificial Intelligence, {EAAI} 2020, New York, NY, USA, February 7-12, 2020}, pages 9386--9393. {AAAI} Press.
\bibitem[{Yang et~al.(2024{\natexlab{b}})Yang, Yang, Jin, Miao, Zhang, Yang, Cui, Zhang, Hui, and Lin}]{codearena}
Jian Yang, Jiaxi Yang, Ke~Jin, Yibo Miao, Lei Zhang, Liqun Yang, Zeyu Cui, Yichang Zhang, Binyuan Hui, and Junyang Lin. 2024{\natexlab{b}}.
\newblock Evaluating and aligning codellms on human preference.
\newblock \emph{arXiv preprint arXiv:2412.05210}.
\bibitem[{Yang et~al.(2024{\natexlab{c}})Yang, Zhang, Yang, Jin, Zhang, Peng, Deng, Miao, Liu, Cui et~al.}]{execrepobench}
Jian Yang, Jiajun Zhang, Jiaxi Yang, Ke~Jin, Lei Zhang, Qiyao Peng, Ken Deng, Yibo Miao, Tianyu Liu, Zeyu Cui, et~al. 2024{\natexlab{c}}.
\newblock Execrepobench: Multi-level executable code completion evaluation.
\newblock \emph{arXiv preprint arXiv:2412.11990}.
\bibitem[{Yang et~al.(2025{\natexlab{c}})Yang, Zhang, Yang, Miao, Quan, Wu, Peng, Yang, Liu, Cui et~al.}]{xcoder}
Jian Yang, Wei Zhang, Jiaxi Yang, Yibo Miao, Shanghaoran Quan, Zhenhe Wu, Qiyao Peng, Liqun Yang, Tianyu Liu, Zeyu Cui, et~al. 2025{\natexlab{c}}.
\newblock Multi-agent collaboration for multilingual code instruction tuning.
\newblock \emph{arXiv preprint arXiv:2502.07487}.
\bibitem[{Yu et~al.(2018)Yu, Zhang, Yang, Yasunaga, Wang, Li, Ma, Li, Yao, Roman et~al.}]{yu2018spider}
Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, et~al. 2018.
\newblock Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task.
\newblock In \emph{EMNLP 2018}, pages 3911--3921.
\bibitem[{Zhang et~al.(2024)Zhang, Zhang, Ma, Li, Zhang, Li, Yao, Xu, Zhou, Zhang{-}Li, Yu, Zhao, Li, and Tang}]{tablellm}
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, and Jie Tang. 2024.
\newblock Tablellm: Enabling tabular data manipulation by llms in real office usage scenarios.
\newblock \emph{CoRR}, abs/2403.19318.
\bibitem[{Zhong et~al.(2017)Zhong, Xiong, and Socher}]{zhong2017seq2sql}
Victor Zhong, Caiming Xiong, and Richard Socher. 2017.
\newblock Seq2sql: Generating structured queries from natural language using reinforcement learning.
\newblock \emph{CoRR}.
\end{thebibliography}