D2-ScaleAgent: Dual-Dimensional Scaling for Long Document Understanding

The evolution of long text understanding has seen significant progress with models like Transformer-based architectures (Vaswani et al., 2017) and multimodal fusion techniques (Li et al., 2023; Liu et al., 2024). Early efforts focused on improving contextual encoding and structural modeling, but faced limitations in handling lengthy, visually rich documents due to fixed context windows and static retrieval workflows. Recent advances include multi-modal retrieval-augmented generation (Lewis et al., 2020; Wu et al., 2025), which enhances information access but still relies on static top-K retrieval strategies. Multi-agent systems (Yao et al., 2022; Han et al., 2025) introduced specialized sub-agents for complex reasoning, yet their fixed workflows hinder adaptive evidence management. The challenge remains in dynamically balancing evidence breadth and depth, especially under resource constraints, to achieve comprehensive understanding.

The core problem in long document understanding is evidence insufficiency—either evidence is scattered across multiple pages (breadth insufficiency) or the understanding is superficial and lacks detail (depth insufficiency). Existing methods often follow rigid workflows, failing to adaptively expand retrieval or deepen reasoning based on the query’s complexity. This results in incomplete evidence chains and unreliable answers, especially in multi-modal contexts where visual and structural cues are critical. Addressing this requires a system capable of real-time, evidence-driven adjustment, ensuring both comprehensive coverage and detailed understanding, which current approaches lack.

The key innovation of D2-ScaleAgent is the integration of a Verifier-driven feedback loop with dual-dimensional scaling. It introduces an Evidence Bank as a dynamic memory that tracks evidence states at multiple levels. The system employs attribute decomposition to guide adaptive retrieval expansion, decomposing complex queries into multiple attributes and performing parallel page retrievals with fusion and pruning. When evidence gaps are detected, the Verifier triggers inward deep reasoning, dynamically selecting hierarchical sub-agents—Global Surveyor, Region Locator, Fine-grained Extractor—to extract detailed evidence. This process iterates until evidence is logically closed, ensuring comprehensive and reliable understanding. Unlike prior static workflows, this approach enables real-time, evidence-informed adjustments, significantly improving performance in complex long-text scenarios.

• �� Initialize Evidence Bank with retrieved pages and evidence states.
• �� Decompose query into multiple attributes using a large language model.
• �� Perform parallel retrieval for each attribute, fuse scores, and apply adaptive pruning.
• �� Use Verifier to assess evidence completeness; detect gaps in the Evidence Bank.
• �� If evidence breadth is lacking, trigger Retrieval Scaling: decompose query attributes further, expand retrieval scope.
• �� If evidence depth is insufficient, activate Reasoning Scaling: select hierarchical sub-agents to extract detailed facts.
• �� Incrementally update Evidence Bank with new evidence, maintaining logical consistency.
• �� Repeat the verification and routing process until evidence is complete.
• �� Generate final answer based on the fully saturated Evidence Bank, ensuring logical closure.

The evaluation involved six diverse multi-modal long-document datasets, including MMLongBench-Doc and LongDocURL, covering open and domain-specific texts with complex layouts. Baselines included MDocAgent, MoLoRAG, and ViDoRAG. Metrics focused on QA accuracy, retrieval recall, precision, nDCG, and MRR. The experiments tested the impact of each component via ablation studies, analyzing the contribution of the Verifier, dual scaling modules, and hierarchical sub-agents. Hyperparameters were tuned for optimal evidence coverage and computational efficiency. Results demonstrated consistent improvements over baselines, with detailed analysis of retrieval quality and reasoning depth across scenarios.

D2-ScaleAgent achieved an average QA accuracy of 63.7%, surpassing MDocAgent (58.3%) and MoLoRAG (53.6%). In retrieval tasks, recall and nDCG increased by 5-10%, indicating more relevant evidence retrieval. Ablation studies showed that removing the Verifier or dual scaling modules caused accuracy drops of over 8%, confirming their importance. The system effectively handled dispersed and complex evidence scenarios, validating the dual scaling strategy’s robustness and adaptability. These results demonstrate that dynamic, evidence-driven routing significantly enhances long document understanding performance.

This framework can be applied to legal document analysis, financial report summarization, scientific literature review, and enterprise knowledge management, where comprehensive understanding of lengthy, multimodal texts is crucial. It supports multi-source evidence integration, enabling reliable decision-making and automated reasoning. Future integration with reinforcement learning could further optimize dynamic routing, making the system more autonomous and efficient for real-world deployment in high-stakes environments.

Despite its strengths, D2-ScaleAgent faces challenges in extremely complex or highly dispersed evidence scenarios, where complete logical closure is difficult. High computational costs limit real-time application in very large documents. Dependence on high-quality multimodal data makes it vulnerable to noise and bias. Further optimization is needed to improve efficiency, robustness, and generalization across diverse domains and data qualities.