Deep Tabular Research via Continual Experience-Driven Execution
Junnan Dong, Chuang Zhou, Zheng Yuan, Yifei Yu, Qiufeng Wang, Yinghui Li, Siyu An, Di Yin, Xing Sun, Feiyue Huang
2026-03-23
Summary
This paper focuses on the difficulty large language models have when trying to analyze and answer questions based on complex tables, especially those that aren't organized in a standard way.
What's the problem?
Large language models struggle with tasks that require them to think through multiple steps and connect different parts of a table to find an answer. These tables often have headers that span multiple rows and columns, and aren't laid out in a simple, predictable format, making it hard for the models to understand the relationships between the data.
What's the solution?
The researchers created a new system that works like an 'agent' – it breaks down the problem of understanding a table into a series of decisions. First, it builds a map of how the different parts of the table relate to each other. Then, it uses a strategy to choose the most promising steps to take to find the answer. Finally, it remembers what it has already tried and uses that information to improve its approach over time, storing past results in a special memory system.
Why it matters?
This work is important because it shows that to effectively use large language models with complex tables, you need to separate the high-level planning of *how* to solve the problem from the low-level details of actually working with the table. This approach makes the models much better at handling these kinds of analytical tasks, which are common in many real-world applications.
Abstract
Large language models often struggle with complex long-horizon analytical tasks over unstructured tables, which typically feature hierarchical and bidirectional headers and non-canonical layouts. We formalize this challenge as Deep Tabular Research (DTR), requiring multi-step reasoning over interdependent table regions. To address DTR, we propose a novel agentic framework that treats tabular reasoning as a closed-loop decision-making process. We carefully design a coupled query and table comprehension for path decision making and operational execution. Specifically, (i) DTR first constructs a hierarchical meta graph to capture bidirectional semantics, mapping natural language queries into an operation-level search space; (ii) To navigate this space, we introduce an expectation-aware selection policy that prioritizes high-utility execution paths; (iii) Crucially, historical execution outcomes are synthesized into a siamese structured memory, i.e., parameterized updates and abstracted texts, enabling continual refinement. Extensive experiments on challenging unstructured tabular benchmarks verify the effectiveness and highlight the necessity of separating strategic planning from low-level execution for long-horizon tabular reasoning.