cs.CL提交于 2026-10-28已译

DocSkill: 面向多语言多跳文档问答的技能集成智能体推理

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 · et al.

中文摘要

多语言多跳文档问答（MMHop-DocQA）需在复杂布局、多页文档及多种语言中定位并整合证据。现有方法如单次处理或自适应工具使用，在协调证据获取、跨页整合及答案验证方面仍具挑战。DocSkill 提出一种具备学习型技能编排策略的智能体框架，根据动态证据状态选择、重访或跳过技能。利用有界蒙特卡洛树搜索（MCTS）离线探索技能轨迹，为监督微调（SFT）提供演示，为直接偏好优化（DPO）提供偏好对。此外，构建了 MMHop-DocBench 数据集，包含 10 种语言、8 个文档领域的 60K 问答对，其 HopGen 流水线通过构建层级页面元素图生成带证据标注的问答对。实验显示，基于 Qwen3-VL-8B 的 DocSkill 在 MMHop-DocBench 上取得 41.0 准确率和 49.2 F1，分别超过最强基线 5.5 和 6.3 分，并在 MMLongBench-Doc 和 DocBench 上表现具有竞争力。

关键要点

01针对多语言多跳文档问答中证据获取与跨页整合协调困难的问题，提出技能集成智能体框架 DocSkill。
02利用有界 MCTS 离线生成技能轨迹，构建 SFT 演示数据与 DPO 偏好对以训练技能编排策略。
03发布 MMHop-DocBench 数据集，涵盖 10 种语言、8 个领域共 60K 问答对，支持层级页面元素证据标注。
04基于 Qwen3-VL-8B 的 DocSkill 在 MMHop-DocBench 上准确率 41.0、F1 49.2，显著优于现有最强基线。

解读

尚无解读。

原始英文摘要

arXiv:2610.36759v1 Announce Type: cross 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.