Rethinking Table Pruning in TableQA: From Sequential Revisions to Gold Trajectory-Supervised Parallel Search
Proposes TabTrim, transforming table pruning from sequential revisions to gold trajectory-supervised parallel search, boosting TableQA accuracy by 3.2%.
Key Findings
Methodology
This paper introduces the TabTrim framework, which leverages intermediate sub-table trajectories derived from gold SQL queries as supervision signals. The approach involves training a pruner and verifier to align step-wise pruning with these trajectories. During inference, multiple candidate pruning trajectories are generated in parallel via beam search, with the verifier scoring each candidate based on a loss-aware function. The top trajectories are retained, and the best sub-table is selected for downstream reasoning. The core components include SQL decomposition for trajectory construction, preference optimization (DPO) for robust pruning, and multi-trajectory beam search to explore diverse pruning paths. This design addresses the limitations of traditional sequential revision methods, reducing error propagation and improving the preservation of answer-critical data.
Key Results
- On WikiTQ, TabTrim-8B achieves 73.5% average accuracy, surpassing the previous best (Table-Critic 70.3%) by 3.2%. It reaches 79.4% on WikiTQ and 61.2% on TableBench, demonstrating robustness across tasks. The model performs especially well on complex, multi-hop questions, with over 10% improvements in hard categories. Ablation studies confirm that preference optimization and multi-trajectory search are key to these gains, significantly reducing early pruning errors and preserving answer-critical cells.
- Across datasets, the method consistently outperforms baselines, with notable improvements in numerical reasoning and fact verification tasks. The multi-trajectory approach effectively avoids local optima, leading to more accurate and complete sub-tables. The results indicate that the approach scales well with larger models, with 8B parameter models outperforming 4B counterparts by 2.5% on average.
- The experiments validate the hypothesis that reliable supervision via gold SQL trajectories and exploration of multiple pruning paths significantly enhance pruning quality, especially in challenging scenarios. The approach's robustness is demonstrated through extensive ablations, showing that removing preference learning or multi-trajectory search causes accuracy drops of 4% or more.
Significance
This work advances the state-of-the-art in table pruning for question answering by shifting from error-prone sequential revision to a robust, parallel exploration paradigm. It addresses core issues such as unreliable critique signals and error accumulation, providing a scalable, high-precision solution. The framework's ability to preserve answer-critical data while simplifying tables has broad implications for improving large-scale knowledge retrieval, automated data analysis, and multi-hop reasoning. Its integration with SQL decomposition and preference learning offers a new blueprint for future research in robust, explainable table reasoning systems, potentially transforming how AI models interpret complex structured data.
Technical Contribution
The paper introduces a novel supervision mechanism based on gold SQL trajectories, enabling the training of a pruner that aligns with verified, answer-critical sub-tables. It combines this with a loss-aware verifier that scores sub-tables based on precision and recall metrics, emphasizing answer-critical data retention. The multi-trajectory beam search during inference allows exploration of diverse pruning paths, avoiding local optima typical in sequential methods. The integration of preference optimization (DPO) further refines the model's bias toward correct sub-tables, resulting in improved robustness and accuracy. These innovations collectively establish a new paradigm for scalable, reliable table pruning in complex question answering tasks.
Novelty
This is the first work to leverage gold SQL execution trajectories as supervision signals for step-wise table pruning, transforming the process into a multi-path exploration problem. Unlike prior methods relying on probabilistic judgments or error signals, this approach grounds pruning decisions in verified, intermediate sub-tables. The combination of SQL decomposition, preference learning, and multi-trajectory beam search creates a comprehensive framework that significantly outperforms existing single-trajectory, sequential methods, marking a new direction in robust table reasoning.
Limitations
- The reliance on gold SQL trajectories limits applicability to datasets with annotated SQL, reducing generalization to real-world, unlabeled tables. The computational overhead of multi-trajectory search increases inference time, especially for large tables or complex queries. The verifier's effectiveness depends on training data quality, which may introduce biases. Future work should focus on reducing computational costs, extending supervision signals, and improving generalization to broader data domains.
Future Work
Future research could explore unsupervised or weakly supervised methods for trajectory generation, reducing dependency on annotated SQL data. Enhancing search efficiency through adaptive pruning strategies or reinforcement learning could make the approach more scalable. Integrating multi-modal data, such as images or text, may expand applicability to more complex reasoning tasks. Additionally, extending the framework to multi-task settings, including fact verification and data summarization, could further demonstrate its versatility and impact.
AI Executive Summary
Table question answering (TableQA) has gained prominence with the development of large pre-trained models, but effective reasoning over large or complex tables remains challenging. Traditional table pruning methods, which aim to extract compact sub-tables, often rely on sequential, error-prone revision processes driven by unreliable critique signals. These approaches can inadvertently discard critical answer data, leading to degraded performance, especially on intricate queries requiring multi-step reasoning.
To address these limitations, this paper introduces TabTrim, a novel framework that transforms the pruning process into a parallel, multi-trajectory search guided by gold SQL query trajectories. The core idea is to leverage intermediate sub-tables generated during gold SQL execution as reliable supervision signals, enabling the training of a pruner that aligns with verified, answer-critical sub-tables. Alongside, a loss-aware verifier assesses the quality of candidate sub-tables based on precision and recall metrics, emphasizing the preservation of answer-relevant data.
During inference, TabTrim employs beam search to generate multiple candidate pruning trajectories simultaneously. Each candidate is scored by the verifier, and the top trajectories are retained for further expansion. This parallel exploration significantly reduces the risk of early pruning errors and local optima, resulting in more accurate sub-table extraction. Extensive experiments on datasets like WikiTQ, TableBench, and TabFact demonstrate that TabTrim-8B achieves an average accuracy of 73.5%, outperforming existing methods by over 3%. Notably, the method excels in complex, multi-hop reasoning tasks, where preserving answer-critical information is crucial.
This work marks a significant step forward in robust table pruning, offering a scalable, effective solution that enhances the interpretability and accuracy of TableQA systems. Its innovative combination of SQL-based supervision, preference learning, and multi-path search paves the way for future advancements in structured data understanding, with broad implications for AI-driven data analysis, automated reasoning, and knowledge retrieval. Despite increased computational costs, the approach's robustness and scalability suggest promising avenues for real-world deployment and further research in multi-modal, multi-task AI systems.
Deep Analysis
Background
随着自然语言处理和深度学习的发展，表格问答（TableQA）逐渐成为信息检索和智能问答的重要方向。早期工作如TaBERT、TAPEX等，主要关注表格结构理解和语义匹配，但在处理大规模复杂表格时，推理效率和准确率仍受限。表格剪枝技术旨在通过筛除冗余信息，提取关键子表，从而简化推理过程。现有方法多采用程序化逐步修正（如SQL分解）或多步推理（如Chain-of-Thought），但都存在剪枝不鲁棒、信息丢失和错误传播的问题。近年来，利用大语言模型（LLMs）进行剪枝评价的研究逐步兴起，试图通过模型自我批评改善剪枝质量。然而，单一轨迹的修正策略易陷入局部最优，难以应对复杂场景。本文在此基础上提出创新框架，结合SQL轨迹构建、多轨迹搜索和偏好学习，推动表格推理技术向更鲁棒、更高效的方向发展。
Core Problem
传统的表格剪枝方法多依赖逐步修正机制，受限于不可靠的批评信号，难以确保答案关键数据的完整性。序贯修正容易受到早期错误的影响，导致最终子表信息缺失，影响推理准确性。尤其在处理复杂、多步骤查询时，单一轨迹难以跳出局部最优，错误积累严重。如何设计一种既能保证信息完整，又能高效探索多路径的剪枝策略，成为亟待解决的核心问题。这不仅关系到模型的推理能力，也影响到实际应用中的效率和可靠性。
Innovation
本文提出TabTrim框架，核心创新包括：1）利用金SQL查询的中间子表轨迹作为可靠的监督信号，确保每一步剪枝都符合正确的推理路径；2）引入偏好优化（DPO），强化模型偏向于正确轨迹，减少语义错误；3）采用多轨迹并行搜索（beam search），同时探索多个候选路径，避免陷入局部最优。这一机制显著提升剪枝的鲁棒性和多样性，确保答案关键数据的完整保留。不同于传统单一轨迹修正，方法更具探索性和容错性，为复杂场景下的表格推理提供了新思路。
Methodology
- �� 利用SQL分解，将金SQL逐步执行，生成对应的中间子表轨迹，作为监督信号。
- �� 训练剪枝器：通过监督微调（LSFT）和偏好优化（DPO），使模型沿着正确轨迹剪枝，并能从错误轨迹中恢复。
- �� 训练验证器：学习评估子表质量的损失感知评分，确保剪枝过程中保留答案关键数据。
- �� 推理阶段：采用beam search，生成多个候选子表轨迹，利用验证器筛选出最优路径，最终输出最佳子表。
- �� 关键机制：结合SQL分解、轨迹构建、偏好学习、多轨迹搜索，形成完整的端到端剪枝优化流程。
Experiments
采用WikiTQ、TableBench和TabFact等公开数据集，比较多种基线，包括程序基础、LLM基础和批评机制。模型训练采用80K样本，调优参数包括beam宽度、最大深度等。评估指标为准确率和子表信息完整性。通过消融实验验证偏好优化和多轨迹搜索的贡献，分析不同难度层级的表现差异，确保方法在复杂场景中的鲁棒性。
Results
TabTrim-8B在WikiTQ达到73.5%的平均准确率，优于最强基线（70.3%）3.2个百分点。在复杂问题中，提升尤为明显，极难类别准确率提升超10%。多轨迹搜索显著降低早期剪枝错误，提升答案关键数据的保留率。消融实验显示，偏好优化和多轨迹机制是性能提升的关键因素，整体验证了方法的有效性和鲁棒性。
Applications
该技术适用于自动化表格问答、数据分析和知识库检索等场景。通过提升剪枝鲁棒性和推理效率，可应用于企业智能问答系统、自动报告生成和大规模数据检索。未来结合多模态信息，将实现更复杂的跨模态表格理解与推理，推动行业智能化升级。
Limitations & Outlook
当前方法依赖SQL轨迹的构建，训练数据的依赖性较强，泛化能力有限。多轨迹搜索带来较高的计算成本，推理时间较长，难以实时应用。验证器性能受训练样本质量影响，可能引入偏差。未来需优化搜索效率，减少计算资源消耗，并扩展到无标注或少标注场景。
Plain Language Accessible to non-experts
想象你在厨房里准备一道菜。每次做菜都需要从一堆食材中挑出关键的，比如蔬菜、肉和调料。传统方法就像你逐步试错，先挑一些材料，发现少了重要的调料，又得重新挑选。这既费时又可能漏掉关键的东西。现在，有了新方法，就像你提前知道哪些材料必须用，还同时准备多个不同的组合，最后挑出最合适的那一份。这样不仅省时，还能确保菜味道正宗。这就像论文中的多轨迹搜索，提前规划多个方案，最后选出最优的，保证答案完整又准确。
ELI14 Explained like you're 14
嘿，你知道做菜的时候，有时候会试错，先放点盐，结果发现还不够咸，又得重新调料。传统的做法就像这样，一次次试，可能会漏掉一些重要的调料，最后做出来的菜不够完美。现在，想象你有个神奇的厨师助手，它提前告诉你哪些调料一定要用，还帮你准备多个不同的调料组合，然后你只需试试哪个最好。这样一来，不仅省时间，还能做出超级好吃的菜！这就像论文里的新方法，用多个方案同时试，最后挑出最棒的那一个，确保答案既完整又准确。是不是很酷？
Abstract
Table Question Answering (TableQA) benefits significantly from table pruning, which extracts compact sub-tables by eliminating redundant cells to streamline downstream reasoning. However, existing pruning methods typically rely on sequential revisions driven by unreliable critique signals, often failing to detect the loss of answer-critical data. To address this limitation, we propose TabTrim, a novel table pruning framework which transforms table pruning from sequential revisions to gold trajectory-supervised parallel search. TabTrim derives a gold pruning trajectory using the intermediate sub-tables in the execution process of gold SQL queries, and trains a pruner and a verifier to make the step-wise pruning result align with the gold pruning trajectory. During inference, TabTrim performs parallel search to explore multiple candidate pruning trajectories and identify the optimal sub-table. Extensive experiments demonstrate that TabTrim achieves state-of-the-art performance across diverse tabular reasoning tasks: TabTrim-8B reaches 73.5% average accuracy, outperforming the strongest baseline by 3.2%, including 79.4% on WikiTQ and 61.2% on TableBench.