DocSkill: Learned Skill Orchestration for Multilingual Multi-Hop Document QA
ORIGINAL / DocSkill: Skill-Integrated Agentic Reasoning for Multilingual Multi-Hop Document Question Answering
The work presents DocSkill, an agentic framework whose learned skill-orchestration policy adaptively selects, revisits, or skips evidence-acquisition skills across multi-page, multilingual documents while tracking evolving evidence state; it also releases MMHop-DocBench, a 60K QA-pair benchmark spanning 10 languages and eight document domains with page- and element-level evidence annotations.
01 ABSTRACT
The source notes that multilingual multi-hop document QA requires locating and integrating evidence across complex layouts, pages, and languages, and that single-pass processing and adaptive tool use struggle to coordinate evidence acquisition, cross-page integration, and answer verification. DocSkill uses bounded Monte Carlo Tree Search offline to explore skill trajectories, yielding demonstrations for supervised fine-tuning and preference pairs for Direct Preference Optimization. Experiments report that DocSkill with Qwen3-VL-8B reaches 41.0 accuracy and 49.2 F1 on MMHop-DocBench and is competitive on MMLongBench-Doc and DocBench.
02 KEY FINDINGS
- DocSkill's policy chooses, revisits, or skips skills based on the evolving evidence state rather than a fixed pipeline.
- Bounded MCTS explores skill trajectories offline, providing SFT demonstrations and DPO preference pairs.
- MMHop-DocBench contains 60K QA pairs across 10 languages and eight document domains with page- and element-level evidence annotations.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.CL