DocSeeker: Structured Visual Reasoning and Evidence Localization to Tackle Long Document Understanding Challenges

The DocSeeker framework addresses the low signal-to-noise ratio and weak supervision signals of multimodal large models in long document understanding through a three-stage workflow of "Analysis-Localization-Reasoning" and a two-stage training strategy, enabling robust generalization from short-document training to ultra-long documents. This framework focuses on structured visual reasoning and evidence localization, providing an effective technical path for long document processing.

In long document understanding, the performance of existing multimodal large models drops sharply as document length increases, rooted in two points: 1. Signal-to-Noise Ratio Dilemma: Key information (signal) is overwhelmed by a large amount of irrelevant content (noise); 2. Scarcity of Supervision Signals: Existing datasets only provide final answers, lacking annotations of evidence sources, making it difficult for models to learn to localize evidence.

DocSeeker adopts a structured visual reasoning paradigm, with a three-stage workflow including: 1. Analysis Stage: Understand the problem requirements and form a search strategy; 2. Localization Stage: Explicitly output evidence positions (page/region/text level) to enhance interpretability and accuracy; 3. Reasoning Stage: Generate answers based on localized evidence. The training uses a two-stage strategy: first, fine-tune using supervision data generated from a teacher model via knowledge distillation, then optimize evidence localization and answer correctness using evidence-aware reinforcement learning. Innovations include Evidence-Guided Resolution Allocation (dynamically allocate computing resources) and Natural Synergy with RAG Systems (micro-localization + macro-retrieval).

DocSeeker performs excellently in multiple benchmark tests: 1. Leading Performance: Outperforms existing methods, especially with obvious advantages in complex localization problems; 2. Robust Generalization: After short-document training, it can generalize to ultra-long documents of hundreds of pages; 3. Domain Transfer: Performs well in out-of-domain tasks. Ablation experiments verify: Removing the explicit localization or reinforcement learning stage leads to a significant drop in performance; Uniform resolution processing is inferior to evidence-guided allocation in both efficiency and performance.

DocSeeker can be applied in multiple fields: 1. Legal Document Analysis: Quickly locate contract clauses and compare versions; 2. Financial Report Review: Extract key financial report indicators and identify risks; 3. Medical Record Processing: Locate patient medical history information to support clinical decisions; 4. Scientific Research Assistance: Assist in literature reviews and accelerate knowledge discovery.

Limitations: High computing cost for processing ultra-long documents; mainly supports English; complex multi-hop reasoning ability needs to be enhanced; real-time interaction response delay needs optimization. Future Directions: Optimize computing efficiency, expand multi-language support, enhance multi-hop reasoning, and improve real-time performance. Conclusion: DocSeeker effectively addresses the core challenges of long document understanding through structured reasoning and evidence localization, providing a foundation for building trustworthy AI systems and serving as an important technical path in the era of information explosion. Paper link: http://arxiv.org/abs/2604.12812v1