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 YanarXiv 2026 · arXiv

TL;DR

DocSkill is a new agentic framework designed to tackle the "needle in a haystack" problem of finding and integrating information across massive, multi-page, multilingual documents. By shifting from simple tool-calling to high-level "skill orchestration" and training on reasoning trajectories optimized via Monte Carlo Tree Search (MCTS), DocSkill significantly outperforms existing Vision-Language Models (VLMs) and tool-augmented agents on complex multi-hop reasoning tasks.

What Problem Does This Paper Address?

Answering a question like "List all courses in the Packaging Design specialization in chronological order" sounds simple, but for an AI, it’s a nightmare if the answer is scattered across a 100-page PDF in Indonesian. This is Multilingual Multi-Hop Document QA (MMHop-DocQA).

Current approaches generally fall into two camps, both with flaws:

  1. Vanilla VLMs: They try to "read" everything at once. This often leads to "context overload," where the model loses track of details or hallucinates when evidence is pages apart.
  2. Tool-Augmented Agents: These use tools (like OCR or search) but often struggle to coordinate them. They might find page 5 and page 50 but fail to realize they need to "Integrate" the two to form a complete answer.

The authors identify a gap in coordination: learning when to search, how to inspect a specific table, and how to verify that the final answer is actually supported by the text.

Methods

The core innovation of DocSkill is moving from operation-level tool selection to capability-level skill orchestration.

The Skill Library

Instead of just giving the model a "search" button, DocSkill defines four high-level skills:

  • Locate: Refines the query and retrieves candidate pages.
  • Inspect: Uses fine-grained parsing or "zooming in" to extract details from tables or charts.
  • Integrate: Connects findings across different pages.
  • Verify: A self-correction step to check for consistency and grounding.
Overview of the DocSkill framework showing the skill orchestration and training pipeline.

Learning to Reason via MCTS

To train the model's "brain" (the policy), the authors used Bounded Monte Carlo Tree Search (MCTS) offline. MCTS explores thousands of possible "trajectories" (sequences of skills) to find the most efficient path to the correct answer.

  • SFT (Supervised Fine-Tuning): The model learns from the successful paths found by MCTS.
  • DPO (Direct Preference Optimization): The model is shown "good" vs. "bad" trajectories (e.g., a path that verified the answer vs. one that skipped verification and got it wrong) to further refine its decision-making.

Datasets

The researchers introduced MMHop-DocBench, a massive new benchmark to fill the void of high-quality multi-hop document data:

  • Scale: 60,000 QA pairs across 10 languages (including English, Chinese, Thai, Vietnamese, and more).
  • Complexity: Documents average 68.2 pages, requiring genuine multi-hop navigation.
  • HopGen Pipeline: A novel automated pipeline that builds "Page-Element Graphs" to ensure questions actually require information from multiple pages to solve.
Figure 3: Overview of HopGen, which constructs hierarchical page–element graphs to generate and verify multilingual multi-hop QA pairs with explicit evidence references.

Key Insights and Ideas

  • Skills > Tools: Decoupling the what (the skill) from the how (the tool) allows the agent to focus on the reasoning state. If a tool changes, the high-level skill logic remains valid.
  • Verification is Non-Negotiable: Ablation studies showed that removing the "Verify" skill led to a significant drop in accuracy. Agents, like humans, perform better when they are forced to double-check their work.
  • MCTS as a "Data Engine": Using search to discover the best reasoning paths is far more effective than just using the "shortest" path. It captures the nuances of document navigation that simple heuristics miss.

Key Results

DocSkill (using a Qwen3-VL-8B backbone) set new state-of-the-art marks:

  • MMHop-DocBench: Achieved 41.0 Accuracy and 49.2 F1, outperforming the strongest baseline (DocLens) by over 5 points.
  • Public Benchmarks: It also achieved top-tier results on MMLongBench-Doc (69.2 Acc) and DocBench (86.8 LASJ), proving that training on their new multilingual dataset generalizes well to other tasks.
  • Qualitative Edge: As seen in Figure 1, while standard methods might miss a specific course or fail to integrate data across pages, DocSkill's adaptive policy allows it to revisit pages and verify its findings until the answer is complete.
Comparison of DocSkill against VLM-based and tool-augmented methods.

Limitations and Future Work

While DocSkill is a leap forward, the authors note a few areas for growth:

  • Skill Discovery: Currently, the four skills are predefined by humans. Future versions could potentially "discover" or refine their own skills for specialized domains (e.g., legal or medical).
  • Language Expansion: While 10 languages is a great start, the "long tail" of low-resource languages remains a challenge for document understanding.
  • Efficiency: Running MCTS for training is computationally expensive; finding more efficient ways to generate high-quality reasoning trajectories is an open research question.

Summary updated

Abstract

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.
BibTeX
@misc{huang2026docskill,
  author = {Jing Huang and Zhiya Tan and Changtao Miao and Shutao Gong and Yang Wang and Joey Tianyi Zhou and Jianshu Li and Ying Yan},
  title = {{DocSkill: Skill-Integrated Agentic Reasoning for Multilingual Multi-Hop Document Question Answering}},
  year = {2026},
  eprint = {2610.36759},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.36759},
}
Figure