DocTrace: Towards Traceable Long Document VQA via Hierarchical Evidence Graph Reasoning
Proposes DocTrace, a hierarchical evidence graph reasoning model for LongDocVQA, achieving 14.4% accuracy improvement and explicit evidence provenance.
Key Findings
Methodology
This paper reformulates LongDocVQA as an explicit evidence graph reasoning task. The model comprises three stages: evidence localization, structured document parsing, and evidence graph construction. Using automatically generated supervision data, the approach combines joint supervised fine-tuning (SFT) with task-specific Group Relative Policy Optimization (GRPO). The architecture employs multi-task learning with reinforcement signals to optimize evidence accuracy, graph faithfulness, and answer correctness. The evidence graph explicitly models dependencies among evidence units, enabling transparent reasoning. Experiments on datasets MMLongBench-Doc, LongDocURL, and SlideVQA demonstrate superior performance over state-of-the-art models, with significant accuracy gains and node-level evidence traceability.
Key Results
- On MMLongBench-Doc, DocTrace improves accuracy from 38.5% (Qwen3-VL-8B) to 52.9%, a 14.4-point increase; on LongDocURL, from 45.1% to 56.4%; on SlideVQA, F1 score rises from 73.4 to 85.1. The model constructs explicit evidence graphs, ensuring transparent, verifiable reasoning paths.
- The two-stage training—joint SFT and task-specific GRPO—significantly enhances evidence localization and graph reasoning. Ablation studies show that explicit graph modeling boosts multi-hop reasoning, robustness, and interpretability in long documents.
- Across diverse scenarios, including multi-page and multimodal inputs, DocTrace maintains high performance, with evidence graphs providing clear reasoning trails. This approach addresses core challenges in long document understanding, especially in high-stakes domains requiring explainability.
Significance
This work advances long document multimodal reasoning by transforming implicit inference into explicit, traceable evidence graphs. The approach enhances transparency, accountability, and trustworthiness, crucial for applications like financial auditing, legal analysis, and medical diagnosis. The hierarchical framework and reinforcement learning training paradigm set new standards for explainable AI in complex, multi-hop reasoning tasks, bridging the gap between performance and interpretability, and paving the way for industry adoption of trustworthy AI systems.
Technical Contribution
The paper introduces a hierarchical evidence graph reasoning framework that explicitly models evidence dependencies across long documents. Key innovations include: • Multi-stage process combining evidence localization, structured parsing, and graph reasoning; • Automatic supervision data generation for training evidence localization and graph construction; • Reinforcement learning with dedicated rewards (localization, faithfulness, correctness) to optimize reasoning fidelity. These contributions enable scalable, interpretable, and high-performing long document reasoning, setting a new benchmark for traceable multimodal VQA.
Novelty
This research is the first to explicitly model evidence dependencies as a hierarchical graph in LongDocVQA, moving beyond traditional implicit representations. Unlike prior methods relying solely on latent features or coarse retrieval, DocTrace constructs detailed evidence graphs with node-level provenance, enabling transparent, multi-hop reasoning. The integration of automatic supervision with reinforcement learning for evidence graph optimization is a novel approach that significantly enhances interpretability and robustness, marking a breakthrough in explainable multimodal document understanding.
Limitations
- The model’s performance degrades on extremely long documents (>120 pages) due to challenges in evidence localization and graph construction, indicating the need for more scalable techniques.
- Supervision data generated automatically may contain biases or inaccuracies, affecting evidence quality and reasoning fidelity. Human-annotated data could further improve results.
- Training with reinforcement learning is computationally intensive, limiting real-time deployment. Future work should focus on efficiency and model compression to facilitate broader industry use.
Future Work
Future research will explore integrating external knowledge bases and causal reasoning to deepen understanding. Enhancing evidence localization accuracy for ultra-long documents and reducing computational costs through model compression are key directions. Additionally, extending the framework to other multimodal tasks, such as video or speech-based long-form understanding, will broaden its applicability. Emphasizing user-interpretable explanations and interactive reasoning interfaces will further promote trustworthy AI deployment.
AI Executive Summary
Long document visual question answering (LongDocVQA) presents unique challenges due to the dispersed, heterogeneous nature of evidence across multiple pages and modalities. Existing models often rely on implicit representations, which hinder transparency and verification. To address this, this paper introduces DocTrace, a hierarchical framework that explicitly constructs evidence graphs to model the reasoning process. The approach involves three key stages: coarse evidence localization, structured document parsing, and evidence graph reasoning. By automatically generating supervision data and employing a combination of supervised fine-tuning and reinforcement learning, the model learns to accurately localize relevant evidence, parse it into meaningful units, and build dependency graphs that trace the reasoning path.
Experiments on datasets such as MMLongBench-Doc, LongDocURL, and SlideVQA demonstrate that DocTrace outperforms current state-of-the-art models, achieving accuracy improvements of up to 14.4 percentage points. The evidence graphs provide node-level provenance, ensuring that each answer can be traced back to specific document elements, thus significantly enhancing interpretability. The model’s robustness across varying document lengths and complex reasoning scenarios highlights its scalability and practical value.
This work marks a significant step toward trustworthy AI in long-form multimodal understanding. By explicitly modeling evidence dependencies, it bridges the gap between high performance and explainability, crucial for high-stakes applications like finance, healthcare, and legal analysis. The hierarchical evidence graph reasoning paradigm opens new avenues for research and industry adoption, emphasizing transparency, verification, and user trust in AI systems. Future directions include integrating external knowledge, improving efficiency, and extending to other modalities, promising a comprehensive solution for complex long-document comprehension.
Deep Analysis
Background
长文档理解作为多模态AI的核心任务之一，经历了从单页问答到跨页、多模态信息融合的演变。早期工作如Dude et al. 2023主要关注单页信息检索和问答，后续如InternVL3（Zhu et al. 2025）和Qwen3VL（Bai et al. 2025）扩展到多页、多模态场景，强调跨页信息整合和长距离推理。然而，这些方法多依赖隐式表示，难以实现推理路径的可追溯性。近年来，基于检索和交互的模型如VDocRAG、MoLoRAG等，提升了信息筛选效率，但仍未解决证据依赖关系的明确建模问题。高风险场景如金融审计、医学诊断等，要求模型不仅给出答案，还能追溯证据来源，确保推理的透明性和可信度。
Core Problem
长文档VQA的核心难点在于信息分散、推理路径复杂，传统模型难以明确追踪每一步推理所依据的具体证据。隐式表示导致推理过程不透明，难以验证和改进。现有方法在多页、多模态信息融合、跨页推理和证据依赖建模方面存在明显不足，限制了模型在高风险场景中的应用。如何构建可解释、可验证的推理路径，成为亟待解决的关键问题。
Innovation
本文提出层次化证据图推理架构，创新点包括：• 证据定位、结构化解析与图推理三阶段流程，确保推理路径明确；• 自动生成监督数据，结合微调和强化学习，优化证据的可靠性和推理连贯性；• 构建节点级证据溯源图，实现推理过程的透明化。这些创新突破了隐式推理的局限，为多跳推理提供了理论保证和工程实现路径。
Methodology
- �� 输入长文档和问题，模型首先在低分辨率图像上进行粗略证据定位，筛选出相关页面；• 选中的页面在高分辨率下结构化解析，提取文本、表格、图像等元素，形成基础证据池；• 构建有向证据图，节点代表证据块，边表示推理依赖关系；• 通过强化学习优化证据定位和图结构，确保推理路径的正确性和完整性；• 最终结合证据图生成答案，实现推理的可追溯性。
Experiments
在MMLongBench-Doc、LongDocURL和SlideVQA三大数据集上，模型采用微调和强化学习训练，评估指标包括准确率、F1分数、证据定位的Page F1和Coverage。对比基线模型如Qwen3-VL-8B-Instruct，结果显示本模型在所有指标上均优于对手，特别是在多页、多模态推理任务中表现出色。通过消融实验验证了证据图推理和结构化解析的贡献，模型在长文档中的鲁棒性和可解释性得到显著提升。
Applications
该模型可广泛应用于金融审计、法律文书分析、医学诊断等领域，满足对推理透明性和证据追溯的高要求。实现条件包括高质量的长文档扫描和多模态信息采集，结合行业特定知识库，可大幅提升决策的可信度。未来还可结合知识图谱和因果推理，增强模型的推理深度和解释能力，推动行业智能化升级。
Limitations & Outlook
模型在极长或极复杂的文档中仍存在证据定位不准确的问题，自动生成监督数据可能引入偏差，训练和推理过程计算成本较高，限制了实时应用。未来需优化模型结构和训练策略，提升效率和适应性。
Plain Language Accessible to non-experts
想象你在整理一本很厚的百科全书，要找到关于某个主题的所有相关信息。传统的方法就像随便翻翻，找到一些片段就给出答案，但你不知道这些片段是不是全部、是否正确。而新方法像是建立一张详细的思维导图，把每个信息点都标记清楚，知道每个结论是依据哪个具体的章节、图片或表格得来的。这样，不仅能找到答案，还能清楚告诉别人这个答案是怎么得来的，哪部分信息最重要。这就像是把复杂的知识变成一张有条理的证据链，让每一步都可以追溯，确保答案的可靠性和透明度。
ELI14 Explained like you're 14
想象你在学校图书馆找资料，老师让你写一篇关于某个话题的报告。以前，你可能只记得大概内容，写完后老师也不知道你是不是用了正确的资料。而现在，有了新方法，就像你在写报告时画出一张图，标明每个信息来自哪本书、哪一页、哪一段。这样，不仅你自己知道每个结论的来源，别人也能看懂你是怎么得出这个答案的。这就像在做一个超级详细的证据地图，确保每个结论都能追溯到具体的资料来源。这样一来，你的报告就变得更可靠，也更容易被别人相信。
Abstract
Long Document Visual Question Answering (LongDocVQA) requires Multimodal Large Language Models (MLLMs) to locate, integrate, and reason over heterogeneous document elements distributed across multiple pages. Existing approaches, including end-to-end MLLMs, retrieval-augmented generation (RAG) pipelines, and document agents, often lack explicit mechanisms to represent and verify how grounded evidence is progressively composed during reasoning, limiting both answer accuracy and traceability. In this paper, we cast LongDocVQA as an explicit evidence graph reasoning problem rather than implicit answer prediction. To this end, we propose DocTrace, a hierarchical framework that progressively performs evidence localization, structured document parsing, and evidence graph reasoning to enable explicit evidence provenance. To effectively learn these capabilities, we develop a two-stage training framework: joint Supervised Fine-Tuning (SFT) first initializes evidence localization and graph reasoning abilities, followed by task-specific Group Relative Policy Optimization (GRPO) with dedicated rewards to further optimize these capabilities. Extensive experiments on MMLongBench-Doc, LongDocURL, and SlideVQA demonstrate that DocTrace consistently outperforms both existing open-source baselines and proprietary MLLMs. Compared with the Qwen3-VL-8B-Instruct backbone, DocTrace achieves absolute improvements of 14.4, 11.3, and 11.7 points on the three benchmarks, respectively. Beyond competitive performance, DocTrace constructs traceable evidence graphs with explicit node-level provenance, enabling transparent and verifiable reasoning for long document understanding.