Replacing Multi-Step Assembly of Data Preparation Pipelines with One-Step LLM Pipeline Generation for Table QA
Operation-R1 trains lightweight LLMs via verifiable reinforcement learning to generate entire data preprocessing pipelines in one step, boosting table QA accuracy by 8.83%.
Key Findings
Methodology
This paper introduces Operation-R1, a framework employing a novel variant of reinforcement learning with verifiable rewards (RLVR) to train small LLMs like Qwen-4B/1.7B for one-shot pipeline generation. It integrates a self-supervised, fine-grained reward mechanism based on cell-focused QA, which evaluates each operation’s contribution by checking if answer-critical cells are preserved or removed. To improve training stability, a variance-aware group resampling strategy is used, which dynamically filters response groups based on reward variance. During inference, the model generates a complete data preprocessing pipeline in a single pass, combining structured operators (select, filter) executed directly and semantic operators (add_column) invoked via LLM calls. To enhance robustness, two mechanisms are employed: operation merge, which consolidates multiple candidate pipelines through consensus, and adaptive rollback, which reverts to earlier table states if key information is lost. Extensive experiments on WikiTQ and TabFact show accuracy gains of 8.83 and 4.44 points, with 79% table compression and 2.2× cost reduction.
Key Results
- Operation-R1 achieves 8.83% and 4.44% absolute accuracy improvements over multi-step baselines on WikiTQ and TabFact, respectively, while compressing tables by 79% and reducing costs by over twofold.
- The fine-grained, self-supervised reward mechanism and variance-aware resampling stabilize training and improve pipeline quality, especially in complex scenarios.
- Operation merge and adaptive rollback significantly improve pipeline robustness, reducing errors and information loss during preprocessing.
Significance
This work addresses the critical bottleneck of high latency and cost in traditional multi-step pipeline assembly for table QA. By enabling a single inference step to generate high-quality preprocessing pipelines, it paves the way for practical, real-time applications. The approach combines the interpretability of operator-centric methods with the efficiency of end-to-end models, making large-scale deployment feasible. Its robustness mechanisms ensure reliability in diverse and unseen scenarios, broadening the applicability of table QA systems in industry and research. Overall, it marks a significant step toward efficient, scalable, and interpretable natural language understanding over structured data.
Technical Contribution
The paper introduces a pioneering RLVR-based training paradigm for lightweight LLMs, integrating a self-supervised, fine-grained pipeline reward system and variance-aware response group resampling. It innovatively combines multiple candidate pipelines through consensus filtering and employs adaptive rollback to mitigate information loss. These techniques collectively address the challenges of reward sparsity, training instability, and robustness in one-shot pipeline generation. The framework demonstrates that high-quality, end-to-end data preprocessing can be achieved with significantly reduced inference costs, opening new avenues for efficient natural language reasoning over structured data.
Novelty
This is the first work to leverage verifiable rewards within reinforcement learning to train small LLMs for single-pass pipeline generation tailored to table QA. Unlike prior multi-step methods, it achieves end-to-end pipeline synthesis in one inference, drastically reducing latency and cost. The integration of self-supervised, operation-level rewards and robustness mechanisms such as operation merging and adaptive rollback represents a novel approach that bridges the gap between interpretability, efficiency, and reliability in data-driven NLP tasks.
Limitations
- The current reward mechanism is primarily designed for cell-focused QA scenarios; extending to more complex or abstract answers may require new reward formulations.
- Handling extremely large or high-dimensional tables remains challenging, as computational costs and model capacity limitations persist.
- While costs are reduced, the single-pass approach may still face difficulties in highly complex or ambiguous questions, necessitating further optimization.
Future Work
Future research will explore multi-modal extensions, integrating visual and structural data for richer reasoning. Developing more sophisticated reward signals for diverse answer types and expanding to multi-turn interactions are promising directions. Additionally, optimizing distributed execution and scaling to larger datasets will further enhance practical deployment, aiming toward fully autonomous, real-time structured data understanding systems.
AI Executive Summary
Table Question Answering (TQA) has become a vital area in natural language understanding, enabling users to query structured data using natural language. However, existing operator-centric methods, which generate multi-step pipelines for data preprocessing, face significant challenges in latency and computational cost due to repeated large model calls. These limitations hinder real-world deployment, especially in scenarios demanding quick responses and low resource consumption.
To address this, the authors propose Operation-R1, a novel framework that leverages lightweight large language models (LLMs) trained via a reinforcement learning paradigm with verifiable rewards (RLVR). The key innovation lies in enabling a single inference step to generate a complete, high-quality data preprocessing pipeline tailored to each question-table pair. This approach dramatically reduces latency and costs while maintaining or improving accuracy.
The core of Operation-R1 involves a self-supervised, fine-grained reward mechanism based on cell-focused QA, which evaluates each operation’s contribution by checking if answer-critical cells are preserved or removed. To stabilize training, a variance-aware group resampling strategy filters response groups based on reward variance, preventing noise amplification. During inference, the model produces multiple candidate pipelines, which are merged through a consensus algorithm, and employs an adaptive rollback mechanism to revert to earlier table states if key information is lost.
Extensive experiments on WikiTQ and TabFact demonstrate that Operation-R1 achieves accuracy improvements of 8.83 and 4.44 percentage points over multi-step baselines, with 79% table compression and 2.2× cost savings. The results validate that a single-pass, robust pipeline generation approach can outperform traditional multi-step methods, paving the way for more efficient, scalable, and interpretable table QA systems. This work significantly advances the practical deployment of end-to-end structured data understanding, with promising future directions including multi-modal reasoning and broader application scenarios.
Deep Analysis
Background
随着大规模预训练模型（如GPT-4、PaLM）在自然语言理解中的突破，表格问答（TQA）逐渐成为研究热点。早期方法多采用文本序列化或SQL生成，存在长表处理困难和错误易发的问题。近年来，操作中心的管道式方法通过逐步操控表格，提升了可解释性和性能，但多次调用模型带来高延迟和成本。为解决这一瓶颈，研究者开始探索单步生成策略，结合强化学习优化操作策略，推动了端到端系统的发展。代表性工作包括DeepSeek-R1和TableR1，强调在训练中引入奖励机制以提升生成质量，但仍面临奖励稀疏和训练不稳定的挑战。本文在此基础上，提出了更高效、鲁棒的单步生成框架，结合自监督奖励和重采样策略，极大改善了训练效果和推理效率。
Core Problem
传统多步骤管道组装虽能达到较高准确率，但在实际应用中存在明显的延迟和成本问题。每次问答都需多次调用大型模型，导致响应时间长、资源消耗大，难以满足实时需求。此外，复杂场景下模型容易生成错误操作，导致信息丢失或错误推理。如何在保证高性能的同时，显著降低调用次数和成本，成为亟待解决的难题。单步生成策略虽具潜力，但在训练中面临奖励稀疏、稳定性不足等问题，限制了其推广应用。
Innovation
本文提出Operation-R1框架，创新点主要包括：• 设计基于细粒度操作奖励的自监督机制，解决奖励稀疏问题，提升训练信号的丰富性；• 引入方差感知分组重采样，增强训练稳定性，避免噪声放大；• 结合操作合并与自适应回滚机制，提高管道生成的鲁棒性，确保信息完整性；• 利用单次推理生成完整数据预处理流程，极大降低成本与延迟。这些创新突破了传统多步方法的瓶颈，推动了端到端表格问答的实用化。
Methodology
- �� 构建基于RLVR的训练流程，采用自监督奖励机制对每个操作进行细粒度评估，奖励包括正确性和压缩效率；
- �� 设计方差感知分组重采样策略，动态筛选响应组，确保训练中奖励的多样性和稳定性；
- �� 训练过程中，模型生成多个候选管道，通过操作合并算法进行共识过滤，提升生成一致性；
- �� 在推理阶段，模型在一次前向中输出完整管道，结合操作执行（结构化操作直接执行，语义操作调用LLM）实现端到端推理；
- �� 引入自适应回滚机制，确保关键信息不丢失，提升系统鲁棒性。
Experiments
采用WikiTQ和TabFact两个公开数据集，比较基线多步骤方法和单步生成模型的性能。指标包括准确率、表格压缩率和成本。设置超参数如奖励权重和采样组大小，进行消融实验验证奖励机制和重采样策略的有效性。模型训练采用多候选管道生成与共识过滤，确保生成质量。评估时，重点分析模型在复杂问答场景中的表现和鲁棒性，验证其在不同表格规模和问题类型下的适应性。
Results
实验结果显示，Operation-R1在WikiTQ上准确率提升8.83%，在TabFact上提升4.44%，显著优于传统多步方法。表格压缩率达79%，成本降低2.2倍，验证了单步推理的高效性。消融分析表明，奖励机制和重采样策略对训练稳定性和生成质量起到关键作用。操作合并和回滚机制显著减少了操作错误，提升了系统鲁棒性。整体而言，该方法在保持高准确率的同时，大幅降低了推理成本，展现出强大的实用潜力。
Applications
该技术适用于企业级自动化数据分析、智能问答系统和实时决策支持。只需提供结构化表格和自然语言问题，系统即可快速生成预处理流程，提升数据处理效率。未来可结合多模态信息，实现更复杂场景的智能问答，推动智能数据分析的普及。
Limitations & Outlook
目前方法主要针对细粒度答案场景，面对更复杂或模糊答案时，奖励机制可能不足。模型在超大规模表格或高维数据中仍存在性能瓶颈。虽然成本降低显著，但在某些高复杂度任务中，单次推理仍存在一定的计算压力。未来需优化算法以应对更大数据规模，并扩展多模态融合能力。
Plain Language Accessible to non-experts
想象你在厨房做菜，准备一道复杂的菜肴。传统方法是逐步添加调料、切菜、煮汤，每一步都需要不停地去看食谱、调整火候。这就像用很多步骤逐个操作，虽然可以做出好菜，但花费时间长、容易出错。现在，假设你有一个智能厨师，只需一次性告诉它所有步骤，它就能在后台快速规划出完整的做菜流程，然后一次性完成所有准备工作。这就像本文提出的Operation-R1，用一个聪明的模型在一次操作中完成所有数据预处理，既快又准，避免了繁琐的多次操作，也减少了出错的可能。这个方法让复杂的厨房工作变得简单高效，节省时间又保证质量。
ELI14 Explained like you're 14
想象你在学校里准备一个大项目，平时你会一步步做：查资料、写草稿、修改、整理。每次都要花很多时间，还可能因为忘记某个步骤而出错。现在，假如你有个超级助手，只要告诉它你的目标，它就能在一瞬间帮你规划好所有步骤，甚至一次性完成所有准备工作！这就像论文写作的AI助手，只需一句话，它就能帮你整理资料、写提纲、排版，节省了很多时间，也不容易出错。本文的Operation-R1就像这个超级助手，用一个聪明的模型在一次操作中完成所有数据预处理，让表格问答变得又快又准。这样一来，处理复杂问题就像和朋友合作一样轻松愉快，效率大大提高！
Abstract
Table Question Answering (TQA) aims to answer natural language questions over structured tables. Large Language Models (LLMs) enable promising solutions to this problem, with operator-centric solutions that generate table manipulation pipelines in a multi-step manner offering state-of-the-art performance. However, these solutions rely on multiple LLM calls, resulting in prohibitive latencies and computational costs. We propose Operation-R1, the first framework that trains lightweight LLMs (e.g., Qwen-4B/1.7B) via a novel variant of reinforcement learning with verifiable rewards to produce high-quality data-preparation pipelines for TQA in a single inference step. To train such an LLM, we first introduce a self-supervised rewarding mechanism to automatically obtain fine-grained pipeline-wise supervision signals for LLM training. We also propose variance-aware group resampling to mitigate training instability. To further enhance robustness of pipeline generation, we develop two complementary mechanisms: operation merge, which filters spurious operations through multi-candidate consensus, and adaptive rollback, which offers runtime protection against information loss in data transformation. Experiments on two benchmark datasets show that, with the same LLM backbone, Operation-R1 achieves average absolute accuracy gains of 8.83 and 4.44 percentage points over multi-step preparation baselines, with 79\% table compression and a 2.2$\times$ reduction in monetary cost.