PaperlayerView exact revision on arXiv

Source-linked AI summary

DocSkill: Skill-Integrated Agentic Reasoning for Multilingual Multi-Hop Document Question Answering

Jing Huang, Zhiya Tan, Changtao Miao, Shutao Gong, Yang Wang, Joey Tianyi Zhou, Jianshu Li, Ying Yan

arXiv:2610.36759v1cs.CLcs.CV

TL;DR

Multilingual multi-hop document QA must integrate sparse evidence across pages, layouts, and languages, while coordinating evidence acquisition, integration, and answer verification remains challenging. DocSkill learns to orchestrate reusable skills using offline MCTS trajectories, and its Qwen3-VL-8B configuration leads the reported MMHop-DocBench comparison.

  • Problem

    Multilingual multi-hop document QA requires locating and integrating sparse evidence across heterogeneous layouts and languages, while coordinating acquisition, cross-page integration, and verification remains challenging.

    Source: §1 · p. 1

  • Method

    DocSkill adaptively orchestrates four reusable document skills, with bounded MCTS trajectories providing SFT demonstrations and DPO preference pairs.

    Source: §1 · p. 2

  • Results

    41.0 Acc. and 49.2 F1 are achieved by DocSkill with Qwen3-VL-8B on MMHop-DocBench, exceeding DocLens with GPT-5.6-SOL by 5.5 Acc. and 6.3 F1 points.

    Source: §5.1 · p. 7

  • Takeaways & Limitations

    Experiments across MMHop-DocBench, MMLongBench-Doc, and DocBench show consistent improvements over strong baselines, while ablations support skill abstraction, adaptive orchestration, and trajectory supervision.

    Source: §6 · p. 8

    All four skills are predefined and fixed during policy training.

    Source: §3.1.1 · p. 3

Abstract

from arXiv · show

Multilingual multi-hop document question answering (MMHop-DocQA) requires locating and integrating evidence across complex layouts, document pages, and diverse languages. Existing approaches include single-pass processing and adaptive tool use, but learning to coordinate evidence acquisition, cross-page integration, and answer verification remains challenging. To address this challenge, we present DocSkill, an agentic framework with a learned skill-orchestration policy for adaptive evidence acquisition and reasoning. The policy selects, revisits, or skips skills according to the evolving evidence state. To construct supervision for skill orchestration, bounded Monte Carlo Tree Search (MCTS) explores alternative skill trajectories offline. These trajectories provide demonstrations for supervised fine-tuning (SFT) and preference pairs for Direct Preference Optimization (DPO). We further introduce MMHop-DocBench, a multilingual multi-hop document QA dataset with training and test splits, containing 60K QA pairs across 10 languages and eight document domains. Its HopGen pipeline builds hierarchical page-element graphs to derive traceable cross-page evidence paths and generate grounded QA pairs with page- and element-level evidence annotations. Experiments show that DocSkill with Qwen3-VL-8B achieves 41.0 accuracy and 49.2 F1 on MMHop-DocBench, exceeding the strongest evaluated baseline by 5.5 and 6.3 points, respectively, and achieving competitive performance on MMLongBench-Doc and DocBench.

1 Introduction

MMHop-DocQA requires locating and integrating sparse evidence across pages, layouts, and languages. DocSkill learns to orchestrate document skills, with MCTS-based trajectory supervision, while MMHop-DocBench provides grounded multilingual training and evaluation data.

1 Introduction

  • The task requires iterative navigation, structural grounding, and multi-step reasoning over sparse evidence distributed across heterogeneous layouts and languages.VLM single-pass approaches can face context overload and incomplete cross-page evidence integration, while coordinating acquisition, integration, and verification remains challenging for tool-augmented methods.Source: §1 · pp. 1–2
  • DocSkill shifts agent control from individual operations to four reusable skills that the policy can select, skip, repeat, and compose as the interaction state evolves.Locate retrieves pages, Inspect interprets fine-grained content, Integrate connects findings across pages, and Verify checks answer grounding and consistency.Source: Figure 1 · p. 1, §1 · p. 2
  • Bounded MCTS explores skill trajectories that are ranked for correctness and quality, producing demonstrations for SFT and preference pairs for DPO.Trajectory quality includes language consistency, logical soundness, and orchestration effectiveness.Source: §1 · p. 2
  • HopGen builds hierarchical page-element graphs and traceable cross-page evidence paths to create MMHop-DocBench, with 60K QA pairs across 10 languages and eight domains.The dataset includes page- and element-level evidence annotations and training and test splits.Source: §1 · p. 2
  • DocSkill outperforms strong baselines on MMHop-DocBench and achieves competitive performance on MMLongBench-Doc and DocBench.The introduction reports experiments on three benchmarks but does not give the specific comparison values.Source: §1 · p. 2

2 Related Work

Prior document QA approaches combine direct VLM processing, OCR, retrieval, or iterative agent tools. Skill-integrated agent work defines and composes task-level capabilities, while DocSkill applies state-dependent orchestration to multilingual multi-hop document QA.

2.1 Document Question Answering

  • VLMs process document images directly, while OCR pipelines extract text and layout for later reasoning and RAG retrieves evidence from searchable indexes.Document agents extend these approaches with iterative evidence acquisition and tool use.Source: §2.1 · p. 2
  • DeepRead supports structure-aware retrieval and section reading, ORCA coordinates specialized agents through routing and debate, and Doc-V* combines OCR-free page navigation with working memory.Doc-V* uses SFT and GRPO for policy training.Source: §2.1 · p. 2

2.2 Skill-integrated Agentic Reasoning

  • ReAct and tool-oriented systems establish reasoning and tool-use interfaces, while AutoSkill and XSkill explicitly define and compose task-level skills.DocSkill applies dynamic skill composition to multilingual multi-hop QA based on evolving interaction states.Source: §2.2 · p. 2

3 Method

DocSkill learns to orchestrate reusable document skills that acquire and reason over cross-page evidence. Bounded MCTS curates skill trajectories for supervised fine-tuning and preference learning, while the learned policy selects skills at inference.

3.1.1 Skill Abstraction.

  • Skills are reusable document-level capabilities that organize reasoning and optional tool execution, while tools are executable operations used to carry them out.The four predefined skills remain fixed during policy training.Source: §3.1.1 · p. 3

3.1.2 Interaction State and Policy.

  • DocSkill frames answering as an interaction in which a policy selects parameterized skill invocations or a terminal answer based on evolving state.The state includes accumulated page- and element-level evidence, candidate answers, verification feedback, and interaction history.Source: §3.1 · p. 2, §3.1.2 · p. 3

3.1.3 Document Skills.

  • Locate retrieves candidate evidence pages, Inspect extracts layout-aware element-level findings, Integrate connects cross-page facts, and Verify checks answer completeness, consistency, and evidence support.Verification feedback guides further evidence acquisition and answer revision.Source: §3.1.3 · pp. 3–4

3.1.4 Execution and Composition.

  • Skills compose sequentially through shared state, allowing the policy to revisit or skip them rather than follow a prescribed order.Each execution returns an observation that updates state and is recorded in the interaction history.Source: §3.1.4 · p. 4

3.2 Offline MCTS-Based Skill Trajectory Curation

  • The MCTS search allocates a fixed budget across trajectories because a skill’s value may depend on downstream evidence acquisition, integration, and verification.Search expands feasible actions and truncates offline rollouts when they reach maximum depth or evidence acquisition yields no new page- or element-level evidence.Source: §3.2 · p. 4
  • Bounded MCTS explores candidate skill trajectories, scoring them for answer outcome and trajectory quality to create SFT demonstrations and DPO preference pairs.Search uses up to 3 candidate actions per expansion, depth 6, and 8 simulations per question.Source: §3.2 · p. 4

3.3 Trajectory-Level Preference Learning

  • SFT supervises successful trajectories, including reasoning, skill invocations, native skill outputs, and terminal answers.External tool returns and deterministic state updates serve as context.Source: §3.3 · p. 4
  • DPO preference pairs compare different-scoring trajectories generated for the same question, discarding equal-score pairs.The trajectory-level objective applies to incomplete prefixes as well as trajectories with terminal answers.Source: §3.3 · p. 5

4 MMHop-DocBench

MMHop-DocBench is a multilingual, multi-domain benchmark for multi-hop document QA, built from traceable cross-page evidence paths and annotated with page- and element-level references. HopGen constructs and verifies grounded QA pairs, while a human audit checks their quality.

4 MMHop-DocBench

  • MMHop-DocBench uses HopGen to generate grounded questions from traceable cross-page evidence paths and provides page- and element-level references for each pair.The benchmark comprises long documents from diverse domains.Source: §4 · p. 5

4.1 HopGen: MMHop-DocQA Pair Curation

  • HopGen builds page and page–element graphs to derive cross-page evidence paths, generate grounded QA pairs, and evaluate answer correctness, evidence sufficiency, cross-page necessity, and language consistency.Generation requires complementary evidence from at least two pages; retained candidates meet a predefined evaluation threshold, and normalized duplicates keep the higher-scoring candidate.Source: §4.1 · p. 5, Figure 3 · p. 5

4.2 Dataset Statistics

  • 60K QA pairs span ten languages and eight domains; the benchmark has the widest language coverage and highest average page count in the Table 1 comparison, at 68.2 pages per document.Each pair includes page- and element-level evidence annotations for evaluating evidence localization and answer grounding.Source: §4.2 · p. 5

4.3 Human Audit of Dataset Quality

  • 95.3% of the 1,200 audited QA pairs satisfied all four human-assessed criteria: question validity, answer correctness, evidence sufficiency, and cross-page necessity.The audit sampled training and test pairs across ten languages and document domains, with disagreements adjudicated by a third annotator.Source: §4.3 · p. 5

5 Experiments

Experiments compare DocSkill with strong baselines across MMHop-DocBench, MMLongBench-Doc, and DocBench, and examine skill orchestration and trajectory supervision [p17, p21, p23, p27, p32, p34].

5.1 Main Results

  • 41.0 Acc. and 49.2 F1 are DocSkill Qwen3-VL-8B’s highest evaluated configuration scores on MMHop-DocBench, exceeding DocLens with GPT-5.6-SOL by 5.5 Acc. and 6.3 F1 points.The comparison uses different backbones and training conditions; same-backbone comparisons against DocLens improve by 2.3–2.5 Acc. and 2.4–2.7 F1 points.Source: Table 2 · p. 6, §5.1 · p. 7
  • On MMLongBench-Doc, Qwen3-VL-8B DocSkill scores 69.2 Acc., 67.3 F1, and 68.9 LASJ; on DocBench, it scores 86.8 LASJ.Compared with the strongest listed baseline for each metric, gains are 1.4, 10.5, 0.7, and 1.3 points, respectively.Source: Table 3 · p. 6, §5.1 · p. 7

5 Experiments

  • The experiments evaluate MMHop-DocBench with macro-averaged accuracy and multilingual token-level F1, alongside accuracy, F1, and LASJ on MMLongBench-Doc and LASJ on DocBench.Human results use a 1,300-question subset, with 130 questions sampled from each language, while model results use the full test set.Source: §5 · pp. 6–7

5.2 Qualitative Analysis

  • In a 67-page Indonesian example, DocSkill integrates text and table evidence from pages 18, 31, and 38, while DocAgent stops with incomplete evidence from pages 17 and 31.DocAgent gives an unsupported answer with incorrect citations; DocSkill verifies its grounded answer.Source: §5.2 · p. 7, Figure 4 · p. 7

5.3 Ablation Studies

  • DocSkill exceeds adaptive direct tool selection by 11.2 Acc. and 11.0 F1 points, and a fixed-order skill policy by 7.5 Acc. and 7.4 F1 points.Removing Locate causes the largest Acc. drop (17.6 points), with page recall falling from 82.4 to 41.5; removing Inspect lowers element recall to 43.5. Removing Integrate or Verify reduces Acc. by 9.2 or 4.5 points.Source: §5.3 · pp. 7–8, Table 4 · p. 8
  • For Qwen3-VL-8B, adding DPO after SFT raises MMHop-DocBench scores from 39.3 to 41.0 Acc. and from 47.3 to 49.2 F1.DPO also improves Qwen3-VL-4B from 36.0 to 38.0 Acc. and from 43.6 to 45.9 F1.Source: §5.3 · p. 8
  • Under matched rollout and evaluation budgets, MCTS leads in success rate (65.8%), diversity, and preference pairs per question; its trained policy beats Best-of-N and beam-search policies on Acc./F1.The differences are 3.2/2.7 Acc./F1 points over Best-of-N and 2.4/2.1 points over beam search.Source: Table 5 · p. 8, §5.3 · p. 8

6 Conclusion

The conclusion presents DocSkill, its MCTS-based trajectory supervision, and MMHop-DocBench, then summarizes benchmark results and future directions.

6 Conclusion

  • DocSkill learns to orchestrate four reusable capability-level skills for multilingual multi-hop reasoning, with bounded MCTS constructing trajectories for SFT and DPO.MMHop-DocBench contains 60K grounded QA pairs across 10 languages and eight domains, with page- and element-level evidence annotations.Source: §6 · p. 8
  • Experiments on MMHop-DocBench, MMLongBench-Doc, and DocBench demonstrate consistent improvements over strong baselines, while ablations support skill abstraction, adaptive orchestration, and trajectory supervision.Source: §6 · p. 8
  • Future work will expand language and domain coverage and explore automatic discovery and refinement of reusable document skills for more complex scenarios.Source: §6 · p. 8

A More Details of MCTS Trajectory Generation

The appendix identifies separate prompt materials for generating and evaluating MCTS rollouts.

A. More Details of MCTS Trajectory Generation

B.1 HopGen Construction Pipeline

HopGen automatically curates grounded multilingual multi-hop QA pairs from long documents using hierarchical graphs, cross-page evidence paths, candidate generation, and quality filtering.

B.1 HopGen Construction Pipeline

  • HopGen’s automatic curation procedure constructs hierarchical graphs, extracts cross-page evidence paths, generates candidate questions, and filters them by quality.Source: §B.1 · p. 10
  • The Page Graph captures dependencies between page profiles, while the Page–Element Graph grounds them in localized document elements; evidence paths span two to four pages.Edges require explicit semantic and element-level support; at most three QA candidates are generated per path, and candidates must meet a 0.875 quality threshold.Source: §B.1 · p. 10

B.2 Prompt for HopGen Pipeline

C.1 Dataset Size and Splits

MMHop-DocBench contains 2,299 documents and 60K QA pairs, divided into training and test splits at the source-document level.

C.1 Dataset Size and Splits

  • MMHop-DocBench contains 2,299 documents and 60K QA pairs across its training and test splits.The training split has 1,599 documents and 40,924 QA pairs; the test split has 700 documents and 19,132 QA pairs.Source: §C.1 · p. 10
  • All QA pairs derived from a source document remain together in the same split.Source: §C.1 · p. 11
  • Training QA counts are approximately balanced across languages, while retained questions per document vary after filtering.Source: §C.1 · p. 11

C.2 Dataset Samples

Representative samples show multilingual questions grounded in cross-page textual, tabular, and visual evidence, with translations and explicit evidence annotations.

C.2 Dataset Samples

  • Examples from Chinese, Indonesian, Thai, and Vietnamese combine cross-page textual, tabular, and visual evidence.Source: §C.2 · p. 11
  • Each sample includes a target-language QA pair, its English translation, and page- and element-level evidence annotations.Source: §C.2 · p. 11
  • Figure 10 presents document and QA-pair counts by language for the training and test splits.Source: Figure 10 · p. 11
Loading 2610.36759v1…