DocDancer: Towards Agentic Document-Grounded Information Seeking

TL;DR

DocDancer advances long-context document question answering by introducing an end-to-end, open-source DocQA agent grounded in information-seeking. It employs a two-tool framework (Search for global retrieval and Read for local comprehension) within an Exploration-then-Synthesis data pipeline to generate high-quality training data from diverse PDFs. Evaluations on MMLongBench-Doc and DocBench show competitive to state-of-the-art performance, with DocDancer achieving $F_1$ of $56.8$ and LasJ of $67.6$ on one benchmark and $85.5$ on another, in some cases surpassing human baselines. The work provides actionable insights into document parsing, tool design, and synthetic data strategies for scalable, open-source agentic document understanding, while discussing limitations and ethical considerations.

Abstract

Document Question Answering (DocQA) focuses on answering questions grounded in given documents, yet existing DocQA agents lack effective tool utilization and largely rely on closed-source models. In this work, we introduce DocDancer, an end-to-end trained open-source Doc agent. We formulate DocQA as an information-seeking problem and propose a tool-driven agent framework that explicitly models document exploration and comprehension. To enable end-to-end training of such agents, we introduce an Exploration-then-Synthesis data synthesis pipeline that addresses the scarcity of high-quality training data for DocQA. Training on the synthesized data, the trained models on two long-context document understanding benchmarks, MMLongBench-Doc and DocBench, show their effectiveness. Further analysis provides valuable insights for the agentic tool design and synthetic data.

Figures (13)

• Figure 1: The overall of DocDancer for document-grounded information seeking, where search and read tools for effective document retrieval and comprehension over processed documents.
• Figure 2: Overall of the Exploration-then-Synthesis framework. (i) Exploration stage iteratively interacts with the source document through Action($u$)–Observation($y$)–Intent($i$) steps. (ii) Synthesis stage aggregates the collected evidence to generate the final question and answer. We present a concrete case illustrating the whole generation process in Appendix \ref{['app:case_study']}.
• Figure 3: Distribution of document used to synthesise.
• Figure 4: Ablation study on document parsing and tools.
• Figure 5: Performance comparison between models trained on our synthesized QA data and open-source QA data.