HierDoc: Hierarchical Page-to-Region Evidence Routing for Long-Document Visual Question Answering

Technical Summary: HierDoc

Problem Statement
Multi-page Document Visual Question Answering (DocVQA) is evolving from single-page reading to long-document reasoning, where answer evidence is often sparse and distributed across multiple pages and visual elements. Existing approaches typically operate at only one level of granularity:

• Page-centric methods focus on retrieving relevant pages but treat region-level operations merely as navigation aids, failing to formulate fine-grained evidence selection as a distinct, optimized action.
• Region-centric methods assume relevant pages are already supplied, neglecting the critical upstream task of page acquisition.
  Consequently, page and region selection remain disconnected rather than functioning as successive, optimized stages of a unified evidence acquisition process.

Methodology: HierDoc Framework
HierDoc proposes a hierarchical evidence-routing framework that models long-document evidence acquisition as a two-stage set prediction problem: from pages to regions. The system operates in three distinct stages:

1. Page Routing (Page Policy):

   • Given a question and a rendered document, a trainable page policy selects a structured set of evidence pages.
   • To handle documents longer than a single visual context, the document is partitioned into non-overlapping windows. The policy acts independently on each window, and results are merged.
   • A bounded reflection mechanism is triggered if the initial selection exceeds a threshold, allowing the policy to prune redundant pages without requiring a separate training objective.
   • The policy is optimized using Group Relative Policy Optimization (GRPO) with a structured-set reward that balances evidence coverage (Recall/F1) against context volume and penalizes invalid actions.

2. Semantic Parsing & Region Routing (Region Policy):

   • Selected pages are parsed by MinerU2.5-Pro into candidate semantic regions (paragraphs, tables, figures, etc.), each assigned a discrete alias, bounding box, and OCR/table text. This creates a parser-native, discrete action space.
   • A region policy conditions on the question, the selected page images (with overlaid aliases), and the candidate list to output a structured set of evidence-region IDs.
   • Similar to page routing, a reflection mechanism refines the region set to the smallest sufficient subset.
   • The region policy is also optimized via GRPO with granularity-specific rewards that balance recall, precision, and F1, while penalizing surplus actions and invalid aliases.

3. Grounded Answering:

   • A separate answer model receives the selected full pages (preserving global layout and cross-element relations) alongside the selected region crops and their associated textual metadata (OCR/table text).
   • This design maintains global context while emphasizing fine-grained local evidence. If no regions are selected, the model operates on full pages alone.

Key Contributions

1. Unified Hierarchical Formulation: HierDoc is the first framework to represent page and semantic-region acquisition as successive, answer-agnostic structured-set policies. This makes both routing decisions explicit and independently analyzable.
2. Stage-wise Training Strategy: The authors develop a training strategy using stage-specific GRPO with granularity-specific set rewards. Crucially, they map spatial supervision (bounding boxes) to a discrete semantic action space (parser-native aliases), enabling efficient learning of region selection.
3. Comprehensive Evaluation: The framework is evaluated across five multi-page and long-document VQA benchmarks. The study includes controlled routing and evidence-composition ablations to isolate the value of each component.

Experimental Results

• Performance: HierDoc achieves state-of-the-art or competitive performance among open-weight systems. On LongDocURL, it improves relative accuracy by 16.87% over the strongest reported open-weight baseline. On MMLongBench-Doc, it improves by 27.36% over the baseline (Doc-V*).
• Retrieval Quality: The page policy significantly outperforms existing page-selection components, improving Page F1 by 38.63% on MMLongBench-Doc and 56.76% on LongDocURL, while constructing more compact evidence sets (fewer average selected pages).
• Ablation Insights:
  • Training: GRPO training of the page policy improves downstream QA accuracy by 5.26% and F1 by 4.40% compared to an untrained backbone.
  • Region Routing: Adding trained region routing to a page-only system improves QA accuracy by 5.51% and F1 by 4.82%.
  • Evidence Composition: Combining selected region crops with selected OCR yields the best results. However, removing full pages and relying only on crops and OCR reduces performance, confirming that global page context and fine-grained region evidence are complementary.

Significance and Limitations
The paper claims that organizing coarse page routing and fine-grained region routing as successive, separately optimized stages effectively addresses the challenge of sparse evidence in long documents. The results demonstrate that explicit, answer-agnostic routing policies can significantly enhance evidence localization and downstream answering accuracy.

The authors acknowledge several limitations:

• Error Propagation: The hierarchical design is irreversible; evidence omitted during page routing cannot be recovered in the region stage.
• Dependency: Region routing relies on MinerU for parsing; thus, layout-parsing errors, OCR noise, and heuristic alignment constraints can limit the action space.
• Optimization Gap: Policies are optimized with set-level objectives rather than end-to-end answer feedback, meaning improvements in selection metrics may not always translate proportionally to answer quality.

Future work suggested by the authors includes incorporating uncertainty-aware recovery, jointly optimizing evidence acquisition with answering, and evaluating trade-offs across broader document domains.