DocSkill: Skill-Integrated Agentic Reasoning for Multilingual Multi-Hop Document Question Answering
Summary
DocSkill提出一种带技能编排策略的智能体框架，用于多语言多跳文档问答，通过MCTS离线探索轨迹并结合SFT/DPO训练策略，同时发布包含60K QA对、覆盖10种语言的MMHop-DocBench数据集。
arXiv:2610.36759v1 Announce Type: new
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.
View Cached Full Text
Cached at: 10/28/26, 08:11 AM
# DocSkill: Skill-Integrated Agentic Reasoning for Multilingual Multi-Hop Document Question Answering Source: [https://arxiv.org/html/2610.36759](https://arxiv.org/html/2610.36759) CCS:Computing methodologies Artificial intelligenceCCS:Information systems Information retrievalJing Huang[https://orcid.org/0009-0005-5651-9033](https://orcid.org/0009-0005-5651-9033),Zhiya Tanemail:[zhiya\.tan@antgroup\.com](mailto:[email protected])Affiliation:Ant Digital Technologies, Ant Group,Singapore,SingaporeAffiliation:Nanyang Technological University,Singapore,Singapore,Changtao Miao[https://orcid.org/0000-0002-7634-9992](https://orcid.org/0000-0002-7634-9992)email:[miaochangtao\.mct@antgroup\.com](mailto:[email protected])Affiliation:Ant Digital Technologies, Ant Group,Hangzhou,China,Shutao Gongemail:[gongshutao\.gst@digital\-engine\.com](mailto:[email protected])Affiliation:Ant Digital Technologies, Ant Group,Changsha,China,Yang Wangemail:[wy462625@antgroup\.com](mailto:[email protected])Affiliation:Ant Digital Technologies, Ant Group,Beijing,China,Joey Tianyi Zhouemail:[tyzhou@smu\.edu\.sg](mailto:[email protected])Affiliation:Singapore Management University,Singapore,SingaporeAffiliation:Institute of Advanced Intelligence and Computing \(IAIC\), Agency for Science, Technology and Research \(A\*STAR\),Singapore,Singapore,Jianshu Liemail:[jianshu\.l@antgroup\.com](mailto:[email protected])Affiliation:Ant Digital Technologies, Ant Group,Singapore,SingaporeandYing Yanemail:[fuying\.yy@antgroup\.com](mailto:[email protected])Affiliation:Ant Digital Technologies, Ant Group,Singapore,Singapore © none ###### 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 presentDocSkill,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 introduceMMHop\-DocBench, a multilingual multi\-hop document QA dataset with training and test splits, containing 60K QA pairs across 10 languages and eight document domains\. ItsHopGenpipeline 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\.2F1\\mathrm\{F\}\_\{1\}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\. ###### Keywords: Multilingual Vision\-Language Models, Multilingual Multi\-hop Document Question Answering, Agentic Reasoning, Skill Orchestration Figure 1\.Overview of DocSkill and its comparison with VLM\-based and tool\-augmented methods\. DocSkill adaptively orchestrates four skills,Locate,Inspect,Integrate, andVerify, for cross\-page document reasoning and answering\.Diagram comparing VLM\-based methods, tool\-augmented methods, and DocSkill\. DocSkill adaptively selects four document skills to retrieve relevant pages, inspect fine\-grained content, integrate cross\-page evidence, and verify candidate answers\.## 1\.Introduction Multilingual multi\-hop document question answering \(MMHop\-DocQA\) is a cornerstone capability for intelligent systems that process multilingual, multi\-page documents\([Mathew et al\., 2021](https://arxiv.org/html/2610.36759#bib.bib1);[Tito et al\., 2023](https://arxiv.org/html/2610.36759#bib.bib2);[Ma et al\., 2024](https://arxiv.org/html/2610.36759#bib.bib3);[Zou et al\., 2024](https://arxiv.org/html/2610.36759#bib.bib4);[Huang et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib6);[Tang et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib5)\)\. The fundamental challenge lies in locating and integrating sparse, spatially distributed evidence across heterogeneous layouts and languages, requiring iterative document navigation, structural grounding, and multi\-step reasoning over long contexts\. As illustrated in Figure[1](https://arxiv.org/html/2610.36759#S0.F1), recent approaches can be broadly grouped into two paradigms with progressively increasing reasoning capabilities, yet each exhibiting inherent limitations\.\(a\) VLM\-based Methodsdirectly encode document pages and queries into a Vision\-Language Model \(VLM\)\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7);[OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9);[Anthropic, 2026](https://arxiv.org/html/2610.36759#bib.bib10);[Google, 2026](https://arxiv.org/html/2610.36759#bib.bib39)\)\. Although these methods benefit from end\-to\-end multimodal understanding, single\-pass inference over long document contexts can suffer from context overload, language inconsistency, and incomplete integration of sparse evidence across pages\.\(b\) Tool\-Augmented Methodsenhance document processing through OCR\([Cui et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib11)\), retrieval\-augmented generation\([Yu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib12);[Faysse et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib14);[Cho et al\., 2024](https://arxiv.org/html/2610.36759#bib.bib15);[Guo et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib16)\), or agentic tool use\([Han et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib18);[Jain et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib19);[Sun et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib13);[Zhang et al\., 2026](https://arxiv.org/html/2610.36759#bib.bib20)\)\. While these mechanisms improve evidence access and support adaptive tool use, learning to jointly coordinate evidence acquisition, cross\-page integration, and answer verification remains challenging\. These limitations are further amplified in multilingual settings, which require robust understanding across languages and reliable evidence grounding throughout the reasoning process\. To address these limitations, we introduceDocSkill, an agentic framework for learned skill orchestration in multilingual multi\-hop document QA\. Its key design shifts agent control from operation\-level tool selection to capability\-level skill orchestration\. Specifically, we distinguish a*skill*from a*tool*: a skill is a reusable capability abstraction that specifies what the agent should accomplish given the current interaction state, whereas a tool is a concrete executable operation that specifies how part of that capability is carried out\. Accordingly, a skill may combine model\-native reasoning with one or more tool calls, decoupling orchestration decisions from specific tool implementations\. As illustrated in Figure[1](https://arxiv.org/html/2610.36759#S0.F1), DocSkill organizes document interaction into four reusable skills:Locatecombines query refinement with page retrieval via Page\_Retriever;Inspectextracts and interprets fine\-grained multimodal content, invoking Page\_Parser or Zoom\_In when needed;Integrateconnects complementary findings across pages; andVerifychecks the grounding and consistency of candidate answers against accumulated evidence\. Conditioned on the evolving interaction state, the learned policy adaptively selects, skips, repeats, and composes these skills\. Learning adaptive skill orchestration requires trajectory\-level supervision because the effectiveness of each skill decision depends on its downstream contribution to evidence acquisition, reasoning, and answer generation\. We adopt an offline trajectory curation and preference optimization strategy\. We use bounded Monte Carlo Tree Search \(MCTS\)\([Silver et al\., 2017](https://arxiv.org/html/2610.36759#bib.bib21);[Yao et al\., 2023a](https://arxiv.org/html/2610.36759#bib.bib22)\)to explore alternative trajectories over the skill space\. The resulting trajectories are ranked by answer correctness and overall quality, assessed through language consistency, logical soundness, and skill\-orchestration effectiveness, to construct training data for supervised fine\-tuning \(SFT\) and trajectory\-level preference pairs\. We optimize the orchestration policy through SFT followed by Direct Preference Optimization \(DPO\)\([Rafailov et al\., 2024](https://arxiv.org/html/2610.36759#bib.bib23);[Xie et al\., 2024](https://arxiv.org/html/2610.36759#bib.bib24);[Chen et al\., 2024](https://arxiv.org/html/2610.36759#bib.bib25)\), which learns from preferred and dispreferred skill trajectories\. To address the scarcity of multilingual multi\-hop document QA data, we introduceMMHop\-DocBenchwith train and test splits to support policy learning and benchmark evaluation\. The dataset is constructed usingHopGen, an automatic curation pipeline that represents long documents as hierarchical Page–Element Graphs\([Ho et al\., 2020](https://arxiv.org/html/2610.36759#bib.bib26);[Xiao et al\., 2019](https://arxiv.org/html/2610.36759#bib.bib27)\)and derives traceable cross\-page evidence paths for grounded question generation\. MMHop\-DocBench contains 60K question–answer pairs across 10 languages and 8 document domains, together with explicit page\- and element\-level evidence annotations\. Experiments show that DocSkill outperforms strong baselines on MMHop\-DocBench and achieves competitive performance on both MMLongBench\-Doc and DocBench\. Our main contributions are summarized as follows: - •Agentic Skill\-Orchestration Framework\.We introduce DocSkill, which orchestrates four reusable reasoning and tool\-use skills\. Bounded MCTS generates trajectories for SFT and preference pairs for DPO\. - •Multilingual Multi\-Hop QA Dataset\.We introduce HopGen to construct MMHop\-DocBench, comprising 60K QA pairs across 10 languages and eight domains with page\- and element\-level evidence annotations\. - •Empirical Results and Analysis\.Experiments on three benchmarks evaluate DocSkill against strong baselines, with comparisons and ablations analyzing skill orchestration and trajectory supervision\. Figure 2\.Overview of the DocSkill framework\. During offline training, bounded MCTS explores alternative skill trajectories to construct training data for SFT and preference pairs for DPO\. During inference, the learned policy adaptively orchestrates four skills,Locate,Inspect,Integrate, andVerify, each combining model reasoning with optional tool use\.Framework diagram showing offline trajectory construction and inference\-time skill orchestration\. Bounded MCTS generates supervision for SFT and DPO\. At inference, the learned policy selects document skills based on the evolving interaction state\. Each skill performs model\-native reasoning, optionally invokes tools, and updates the state for subsequent decisions\. ## 2\.Related Work ### 2\.1\.Document Question Answering Document Question Answering \(DocQA\) involves extracting and reasoning over textual and visual evidence in documents\. VLM\-based approaches process document images directly\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7);[OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9);[Google, 2026](https://arxiv.org/html/2610.36759#bib.bib39)\), while OCR\-based pipelines\([Cui et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib11)\)extract text and layout information for subsequent reasoning\. Retrieval\-augmented generation \(RAG\)\([Yu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib12);[Faysse et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib14);[Cho et al\., 2024](https://arxiv.org/html/2610.36759#bib.bib15);[Guo et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib16)\)selects relevant evidence from searchable document indexes to support answer generation\. Document agents extend these approaches through iterative evidence acquisition and tool use\([Han et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib18);[Jain et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib19);[Sun et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib13);[Zhu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib34)\)\. DeepRead\([Li et al\., 2026b](https://arxiv.org/html/2610.36759#bib.bib47)\)supports structure\-aware retrieval and section reading, while ORCA\([Lassoued et al\., 2026](https://arxiv.org/html/2610.36759#bib.bib48)\)coordinates specialized agents through routing and debate\. Doc\-V∗V^\{\*\}\([Zheng et al\., 2026](https://arxiv.org/html/2610.36759#bib.bib49)\)combines OCR\-free page navigation with working memory, using SFT and GRPO for policy training\. Our work focuses on multilingual multi\-hop DocQA, learning state\-dependent orchestration over capability\-level skills through SFT and DPO on offline MCTS trajectories\. ### 2\.2\.Skill\-integrated Agentic Reasoning Skill learning for LLM agents focuses on acquiring structured reasoning and action strategies\. Prior work, including ReAct\([Yao et al\., 2023b](https://arxiv.org/html/2610.36759#bib.bib28)\)and tool\-oriented approaches such as Toolformer\([Schick et al\., 2023](https://arxiv.org/html/2610.36759#bib.bib29)\), ToolAlpaca\([Tang et al\., 2023](https://arxiv.org/html/2610.36759#bib.bib30)\), and Gorilla\([Patil et al\., 2023](https://arxiv.org/html/2610.36759#bib.bib31)\), demonstrates the effectiveness of interleaving reasoning with tool use\. These approaches establish reasoning and tool\-use interfaces\. Some recent works have explored skill\-integrated agent frameworks, such as AutoSkill\([Yang et al\., 2026](https://arxiv.org/html/2610.36759#bib.bib32)\)and XSkill\([Jiang et al\., 2026](https://arxiv.org/html/2610.36759#bib.bib33)\), where task\-level skills and action strategies are explicitly defined and composed to solve complex tasks\. In this work, we introduce DocSkill, a skill\-integrated agentic framework for multilingual multi\-hop document question answering that dynamically composes reasoning actions and tool usage based on evolving interaction states for iterative evidence acquisition and structured document reasoning\. ## 3\.Method Figure[2](https://arxiv.org/html/2610.36759#S1.F2)summarizes DocSkill’s offline training and adaptive inference\. Conditioned on the interaction state, the policy selects amongLocate,Inspect,Integrate, andVerify, each combining reasoning with optional tool execution\. To construct training data, bounded MCTS uses a frozen generator and evaluator to produce scored skill trajectories with explicit reasoning from MMHop\-DocBench QA instances\. The resulting trajectories provide SFT demonstrations and DPO preference pairs for policy training\. At inference, the learned policy orchestrates document skills without tree search\. ### 3\.1\.Skill\-Orchestrated Document Agent Given an inputx=\(𝒟,q\)x=\(\\mathcal\{D\},q\), where𝒟=\{pi\}i=1M\\mathcal\{D\}=\\\{p\_\{i\}\\\}\_\{i=1\}^\{M\}is anMM\-page multilingual document andqqis a question, the goal is to generate an answeryygrounded in cross\-page evidence\. Following ReAct\([Yao et al\., 2023b](https://arxiv.org/html/2610.36759#bib.bib28)\), we formulate DocSkill as a finite\-step interaction with the document environment, in which the agent acquires evidence, updates its reasoning state, and determines its subsequent actions\. #### 3\.1\.1\.Skill Abstraction In DocSkill, a skill is a reusable document\-level capability that organizes model reasoning and optional tool execution around a specific objective\. A tool, in contrast, provides an executable operation, such as page retrieval, document parsing, or region cropping\. This distinction concerns the abstraction used for agent control: the policy selects the document capability to invoke and specifies its arguments, while the corresponding skill carries out the reasoning and tool execution needed for that invocation\. Each skill is represented as \(1\)σk=⟨dk,𝒰k⟩,\\sigma\_\{k\}=\\langle d\_\{k\},\\mathcal\{U\}\_\{k\}\\rangle,wheredkd\_\{k\}describes the capability and𝒰k\\mathcal\{U\}\_\{k\}denotes its available tool set, which may be empty for reasoning\-only execution\. The skill library is \(2\)𝒮=\{𝖫𝗈𝖼𝖺𝗍𝖾,𝖨𝗇𝗌𝗉𝖾𝖼𝗍,𝖨𝗇𝗍𝖾𝗀𝗋𝖺𝗍𝖾,𝖵𝖾𝗋𝗂𝖿𝗒\}\.\\mathcal\{S\}=\\\{\\mathsf\{Locate\},\\mathsf\{Inspect\},\\mathsf\{Integrate\},\\mathsf\{Verify\}\\\}\.All four skills are predefined and fixed during policy training\. #### 3\.1\.2\.Interaction State and Policy At steptt, the interaction state is \(3\)zt=\(q,ℰt,y^t,vt,ℋt\),z\_\{t\}=\\left\(q,\\mathcal\{E\}\_\{t\},\\hat\{y\}\_\{t\},v\_\{t\},\\mathcal\{H\}\_\{t\}\\right\),whereqqis the original input question;ℰt\\mathcal\{E\}\_\{t\}contains accumulated page\- and element\-level evidence and connections across pages;y^t\\hat\{y\}\_\{t\}denotes the current candidate answer;vtv\_\{t\}records verification and reflection feedback; andℋt\\mathcal\{H\}\_\{t\}is the ordered history of preceding actions and their returned observations\. The input document𝒟\\mathcal\{D\}remains accessible through the document environment and is consulted via skill executions and their associated tools\. The policy operates over an action space𝒜\\mathcal\{A\}containing parameterized skill invocations and terminal answer outputs: \(4\)at∼πθ\(⋅∣zt\),at∈𝒜\.a\_\{t\}\\sim\\pi\_\{\\theta\}\(\\cdot\\mid z\_\{t\}\),\\qquad a\_\{t\}\\in\\mathcal\{A\}\.A skill invocation takes the format=\(st,ξt\)a\_\{t\}=\(s\_\{t\},\\xi\_\{t\}\), wherest∈𝒮s\_\{t\}\\in\\mathcal\{S\}specifies the skill andξt\\xi\_\{t\}contains its arguments, such as a retrieval query, target pages, or regions of interest\. Thus, invoking the same skill with different arguments constitutes different actions\. Alternatively, the policy emits a terminal action, which returns the final answeryyand ends the interaction\. #### 3\.1\.3\.Document Skills The four skills perform complementary operations on the interaction state\. Locateuses model reasoning to formulate or refine a query based on the question and interaction state, then invokesPage\_Retrieverto identify candidate evidence pages\. Returned page references and content are added to the accumulated evidence\. The policy can revisit this skill when further evidence is needed\. InspectinvokesPage\_Parserfor layout\-aware extraction and, when needed,Zoom\_Infor region cropping\. Model\-native reasoning interprets the returned content, and the resulting element\-level findings are added to the accumulated evidence\. Integrateuses model\-native reasoning to connect facts across pages and synthesize accumulated evidence to construct or refine the candidate answer, without additional retrieval or extraction\. Verifyuses model reasoning to assess the candidate answer for completeness, consistency, and evidence support\. Its feedback guides further evidence acquisition and answer revision\. #### 3\.1\.4\.Execution and Composition\. Executing a skill returns an observation, updates the relevant state components, and records the action and observation in the interaction history\. Trajectories retain the agent’s recorded reasoning, skill invocations, and native skill outputs, including candidate answers and verification feedback\. Native skill outputs serve as observations that condition subsequent decisions\. Tools are invoked within skills rather than exposed as independent top\-level actions\. Skills compose sequentially through the shared state\. At each step, the policy selects a skill invocation or a terminal answer\. The updated state allows skills to be revisited or skipped without a prescribed order\. These sequences are selected dynamically rather than defined as separate compound skills\. A completed interaction withTTskill invocations is denoted as \(5\)τ=\(z0,a0,o1,z1,…,aT−1,oT,zT,y\),\\tau=\(z\_\{0\},a\_\{0\},o\_\{1\},z\_\{1\},\\ldots,a\_\{T\-1\},o\_\{T\},z\_\{T\},y\),whereata\_\{t\}denotes a skill invocation,o1,…,oTo\_\{1\},\\ldots,o\_\{T\}are the observations, andyyis the final answer\. Incomplete rollouts retain the recorded interaction prefix without a terminal answer\. These trajectories define the space explored during offline supervision construction\. ### 3\.2\.Offline MCTS\-Based Skill Trajectory Curation The parameterized interactions defined above induce a large trajectory space, where the value of a skill invocation may depend on its downstream effects on evidence acquisition, integration, and verification\. Independent rollouts can repeatedly explore similar paths and provide limited coverage of alternative decisions\. We therefore apply bounded Monte Carlo Tree Search \(MCTS\)\([Silver et al\., 2017](https://arxiv.org/html/2610.36759#bib.bib21);[Yao et al\., 2023a](https://arxiv.org/html/2610.36759#bib.bib22)\)to allocate a fixed search budget across candidate skill trajectories\. For each training instancexi=\(𝒟i,qi\)x\_\{i\}=\(\\mathcal\{D\}\_\{i\},q\_\{i\}\), the root noden0n\_\{0\}represents the initial statez0z\_\{0\}, and each subsequent node represents an interaction state\. An edge records a complete action\. For a skill invocationat=\(st,ξt\)a\_\{t\}=\(s\_\{t\},\\xi\_\{t\}\), executing the edge produces an observationot\+1o\_\{t\+1\}and transitions the agent fromztz\_\{t\}tozt\+1z\_\{t\+1\}\. A terminal answer action ends the corresponding branch\. Search depth counts policy decisions, with each edge representing one skill invocation or terminal answer\. Each edgeeemaintains an estimated valueQ\(e\)Q\(e\)and a visit countN\(e\)N\(e\), while each nodennmaintains a visit countN\(n\)N\(n\)\. ##### Selection\. Starting from the root, MCTS recursively selects the edge with the highest Upper Confidence Bound for Trees \(UCT\) score: \(6\)UCT\(e\)=Q\(e\)\+λln\(N\(n\)\+1\)N\(e\)\+1,\\operatorname\{UCT\}\(e\)=Q\(e\)\+\\lambda\\sqrt\{\\frac\{\\ln\\\!\\left\(N\(n\)\+1\\right\)\}\{N\(e\)\+1\}\},wherennis the parent node ofee,Q\(e\)Q\(e\)is the mean return of previous rollouts through the edge, andλ\\lambdabalances exploration and exploitation\. Selection stops at an unexpanded or terminal node\. ##### Expansion\. At an expandable leaf, the generator proposes feasible actions conditioned on the current state\. Skill invocations with different arguments form distinct branches, producing observations and child states through model reasoning and optional tool calls when executed\. Answer actions create terminal children\. ##### Simulation\. From an expanded child, the generator proposes and executes actions until it emits a terminal answer, reaches the maximum depth, or fails to acquire new page\- or element\-level evidence after an evidence\-acquisition action\. This evidence\-stagnation criterion is used to truncate offline rollouts and is not an inference\-time stopping rule\. The evaluator assigns the trajectory quality score: \(7\)C\(τ\)=wlangclang\(τ\)\+wlogicclogic\(τ\)\+wskillcskill\(τ\),C\(\\tau\)=w\_\{\\mathrm\{lang\}\}c\_\{\\mathrm\{lang\}\}\(\\tau\)\+w\_\{\\mathrm\{logic\}\}c\_\{\\mathrm\{logic\}\}\(\\tau\)\+w\_\{\\mathrm\{skill\}\}c\_\{\\mathrm\{skill\}\}\(\\tau\),where the three components measure language consistency, logical soundness, and skill\-orchestration effectiveness, respectively\. Each component takes a value in\{0,0\.5,1\}\\\{0,0\.5,1\\\}\. The weights are non\-negative and sum to one; we assign equal weight to the three components\. The evaluator assigns an outcome scoreO\(τ\)∈\{1,−1,0\}O\(\\tau\)\\in\\\{1,\-1,0\\\}for correct, incorrect, and incomplete trajectories, respectively\. The rollout return combines the outcome score with trajectory quality: \(8\)J\(τ\)=O\(τ\)\+C\(τ\)\.J\(\\tau\)=O\(\\tau\)\+C\(\\tau\)\. ##### Backpropagation\. The rollout returnJ\(τ\)J\(\\tau\)is propagated through the selected path\. For each visited nodenjn\_\{j\}and edgeeje\_\{j\}, we update \(9\)N\(nj\)←N\(nj\)\+1,N\(n\_\{j\}\)\\leftarrow N\(n\_\{j\}\)\+1,\(10\)N\(ej\)←N\(ej\)\+1,N\(e\_\{j\}\)\\leftarrow N\(e\_\{j\}\)\+1,and \(11\)Q\(ej\)←Q\(ej\)\+J\(τ\)−Q\(ej\)N\(ej\)\.Q\(e\_\{j\}\)\\leftarrow Q\(e\_\{j\}\)\+\\frac\{J\(\\tau\)\-Q\(e\_\{j\}\)\}\{N\(e\_\{j\}\)\}\.Thus,Q\(ej\)Q\(e\_\{j\}\)maintains the empirical mean return of rollouts passing through the edge\. We use up toB=3B=3candidate actions per expansion, a maximum search depth ofD=6D=6, andNsim=8N\_\{\\mathrm\{sim\}\}=8simulations per question\. Both the generator and evaluator use frozen GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)with role\-specific prompts and separate inference contexts\. The generator proposes actions and performs model\-native skill reasoning, while reference answers and evidence annotations are used only for offline evaluation\. For each question, the search returns scored trajectories containing recorded explicit reasoning, skill invocations, and execution outputs\. Successful trajectories provide SFT demonstrations, while trajectories with different scores for the same question form DPO preference pairs\. ### 3\.3\.Trajectory\-Level Preference Learning The scored trajectories produced by MCTS provide complementary supervision for imitation and preference learning\. ##### Supervised fine\-tuning\. We supervise successful trajectories, including reasoning, skill invocations, native skill outputs, and terminal answers\. External tool returns and deterministic state updates serve as context\. Let𝒟SFT\\mathcal\{D\}\_\{\\mathrm\{SFT\}\}denote these input–trajectory pairs andu1:Lτu\_\{1:L\_\{\\tau\}\}the serialized tokens ofτ\\tau\. We setmj=1m\_\{j\}=1for model\-generated outputs andmj=0m\_\{j\}=0for context tokens\. The SFT objective is \(12\)ℒSFT=−𝔼\(x,τ\)∼𝒟SFT\[∑j=1Lτmjlogpθ\(uj∣cj\)\],\\mathcal\{L\}\_\{\\mathrm\{SFT\}\}=\-\\mathbb\{E\}\_\{\(x,\\tau\)\\sim\\mathcal\{D\}\_\{\\mathrm\{SFT\}\}\}\\left\[\\sum\_\{j=1\}^\{L\_\{\\tau\}\}m\_\{j\}\\log p\_\{\\theta\}\(u\_\{j\}\\mid c\_\{j\}\)\\right\],wherepθp\_\{\\theta\}is the shared backbone’s token distribution andcjc\_\{j\}contains the role\-specific prompt, available interaction context, and preceding output tokens for the corresponding generation call\. ##### Preference construction\. For each inputxix\_\{i\}, two generated trajectories form an ordered preference pair when one receives a higher rollout score than the other: \(13\)𝒟pref=\{\(xi,τi\+,τi−\)\|J\(τi\+\)\>J\(τi−\)\}\.\\mathcal\{D\}\_\{\\mathrm\{pref\}\}=\\left\\\{\(x\_\{i\},\\tau\_\{i\}^\{\+\},\\tau\_\{i\}^\{\-\}\)\\;\\middle\|\\;J\(\\tau\_\{i\}^\{\+\}\)\>J\(\\tau\_\{i\}^\{\-\}\)\\right\\\}\.Both trajectories in each pair are generated for the same question, and equal\-score pairs are discarded\. ##### Trajectory\-level DPO Using the same supervision mask, we define the conditional log\-likelihood of the recorded model\-generated content as \(14\)ℓθ\(x,τ\)=∑j=1Lτmjlogpθ\(uj∣cj\)\.\\ell\_\{\\theta\}\(x,\\tau\)=\\sum\_\{j=1\}^\{L\_\{\\tau\}\}m\_\{j\}\\log p\_\{\\theta\}\(u\_\{j\}\\mid c\_\{j\}\)\.Terminal\-answer tokens contribute only when present, so this definition also applies to incomplete trajectory prefixes\. We optimize trajectory preferences using \(15\)ℒDPO=−𝔼\(x,τ\+,τ−\)∼𝒟pref\[logσ\(β\[rθ\(x,τ\+\)−rθ\(x,τ−\)\]\)\],\\displaystyle\\mathcal\{L\}\_\{\\mathrm\{DPO\}\}=\-\\mathbb\{E\}\_\{\(x,\\tau^\{\+\},\\tau^\{\-\}\)\\sim\\mathcal\{D\}\_\{\\mathrm\{pref\}\}\}\\Big\[\\log\\sigma\\big\(\\beta\[r\_\{\\theta\}\(x,\\tau^\{\+\}\)\-r\_\{\\theta\}\(x,\\tau^\{\-\}\)\]\\big\)\\Big\],whererθ\(x,τ\)=ℓθ\(x,τ\)−ℓref\(x,τ\)r\_\{\\theta\}\(x,\\tau\)=\\ell\_\{\\theta\}\(x,\\tau\)\-\\ell\_\{\\mathrm\{ref\}\}\(x,\\tau\),σ\\sigmais the logistic sigmoid, and the reference model is the frozen SFT checkpoint\. Both models use the same supervision mask and conditioning contexts\. Table 1\.Comparison of document QA datasets by language count, multi\-hop support, number of QA pairs, average document length, and source\. ## 4\.MMHop\-DocBench We introduce MMHop\-DocBench, a large\-scale benchmark for multilingual multi\-hop document question answering\. It contains long documents from diverse domains and provides page\- and element\-level evidence references for each question–answer pair\. The benchmark is constructed usingHopGen, an automatic pipeline that generates grounded questions from traceable cross\-page evidence paths\. Figure 3\.Overview of HopGen, which constructs hierarchical page–element graphs to generate and verify multilingual multi\-hop QA pairs with explicit evidence references\.HopGen pipeline for multilingual multi\-hop QA curation\. Page profiles and parsed elements form a hierarchical graph whose cross\-page evidence paths guide question–answer generation\. Candidates are evaluated, filtered, and deduplicated, with retained pairs linked to supporting pages and elements\.### 4\.1\.HopGen: MMHop\-DocQA Pair Curation Figure[3](https://arxiv.org/html/2610.36759#S4.F3)provides an overview ofHopGen\. HopGen uses GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)for both generation and evaluation, with role\-specific prompts and separate inference contexts, while PP\-OCR\([Cui et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib11)\)provides layout\-aware parsing to extract fine\-grained elements from document pages\. For each document, the generator creates a compact profile of every page, including its summary, entities, topics, and layout structures\. It then assesses candidate page pairs for explicit semantic dependencies to construct a Page Graph\. The parser extracts fine\-grained document elements and attaches them to their corresponding source pages, forming a hierarchical Page–Element Graph\. Connected page sequences and their associated elements define candidate cross\-page evidence paths for grounded question generation and subsequent verification\. Given an evidence path and parsed elements, the generator produces questions, answers, and supporting page–element references in the document language using only the supplied evidence\. The generator is instructed to create questions requiring complementary evidence from at least two distinct pages, with no single page sufficient for the complete answer\. The evaluator scores candidates for answer correctness, evidence sufficiency, cross\-page necessity, and language consistency, retaining those meeting a predefined threshold\. Normalized question–answer duplicates are resolved by retaining the higher\-scoring candidate\. MethodENZHYUEIDMSTLTATHVIMYAcc\.𝐅𝟏\\boldsymbol\{\\mathrm\{F\}\_\{1\}\}Vanilla VLM MethodsNaive VLMw/ Qwen3\-VL\-4B\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7)\)18\.0/23\.215\.2/20\.38\.3/12\.016\.9/23\.614\.9/19\.69\.0/12\.92\.0/3\.35\.4/8\.010\.3/14\.02\.9/4\.210\.314\.1Naive VLMw/ Qwen3\-VL\-8B\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7)\)24\.0/28\.921\.0/25\.912\.0/15\.720\.6/28\.019\.8/23\.913\.5/16\.63\.6/5\.18\.2/10\.514\.3/18\.24\.3/5\.714\.117\.9Naive VLMw/ Qwen3\-VL\-32B\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7)\)24\.2/30\.021\.2/26\.115\.2/19\.925\.5/32\.722\.5/27\.815\.8/20\.85\.2/7\.612\.6/16\.118\.2/23\.15\.6/7\.516\.621\.2Naive VLMw/ Gemini\-3\.1\-Pro\([Google, 2026](https://arxiv.org/html/2610.36759#bib.bib39)\)25\.0/31\.324\.3/29\.820\.8/27\.525\.4/33\.723\.8/30\.520\.4/28\.612\.9/18\.621\.8/28\.222\.0/29\.36\.1/11\.820\.326\.9Naive VLMw/ Claude\-Sonnet\-5\([Anthropic, 2026](https://arxiv.org/html/2610.36759#bib.bib10)\)25\.6/32\.023\.8/29\.522\.1/29\.226\.0/34\.524\.8/31\.422\.5/31\.813\.5/19\.820\.8/28\.122\.4/29\.86\.4/12\.520\.827\.9Naive VLMw/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)26\.2/32\.624\.0/30\.121\.8/28\.826\.5/35\.224\.5/31\.222\.2/31\.514\.6/20\.521\.4/28\.422\.7/30\.16\.7/12\.921\.128\.1OCR\-Based MethodsTesseract\([Smith, 2007](https://arxiv.org/html/2610.36759#bib.bib40)\)w/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)23\.2/30\.121\.5/27\.614\.8/20\.523\.9/31\.222\.5/28\.819\.8/26\.911\.8/16\.517\.5/23\.418\.9/25\.67\.4/11\.618\.124\.2PP\-OCR\([Cui et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib11)\)w/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)27\.1/33\.524\.8/31\.221\.5/28\.427\.3/36\.125\.4/32\.322\.8/32\.114\.2/20\.122\.3/29\.523\.1/30\.87\.6/13\.821\.628\.8RAG\-Based MethodsVisRAG\([Yu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib12)\)w/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)28\.5/35\.225\.8/32\.422\.1/29\.127\.8/36\.525\.9/33\.123\.2/32\.614\.5/20\.322\.6/29\.823\.5/31\.27\.8/14\.122\.229\.4ColPali\([Faysse et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib14)\)w/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)29\.2/36\.126\.4/33\.223\.1/29\.828\.5/37\.426\.5/34\.023\.8/33\.214\.8/20\.823\.1/30\.424\.0/31\.88\.0/14\.522\.730\.1Qwen3\-VL\-Emb\-8B\([Li et al\., 2026a](https://arxiv.org/html/2610.36759#bib.bib17)\)w/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)30\.8/38\.528\.2/35\.622\.6/32\.130\.1/39\.228\.1/36\.225\.4/35\.116\.2/22\.624\.8/32\.525\.6/33\.98\.9/15\.824\.132\.2Agent\-Based MethodsDocAgent\([Sun et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib13)\)w/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)27\.9/34\.225\.5/31\.922\.2/29\.028\.0/36\.726\.1/32\.923\.5/32\.714\.9/20\.823\.0/30\.123\.8/31\.48\.2/14\.422\.329\.4SimpleDoc\([Jain et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib19)\)w/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)28\.6/35\.026\.2/32\.822\.8/29\.928\.7/37\.626\.8/33\.824\.2/33\.615\.5/21\.623\.6/30\.924\.4/32\.28\.6/15\.022\.930\.2DocLens\([Zhu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib34)\)w/ Gemini\-3\.1\-Pro\([Google, 2026](https://arxiv.org/html/2610.36759#bib.bib39)\)38\.6/45\.536\.3/43\.233\.6/40\.938\.3/45\.636\.9/44\.234\.7/42\.225\.9/32\.634\.3/41\.635\.2/42\.918\.6/25\.233\.240\.4DocLens\([Zhu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib34)\)w/ Claude\-Sonnet\-5\([Anthropic, 2026](https://arxiv.org/html/2610.36759#bib.bib10)\)40\.3/47\.437\.6/44\.734\.7/42\.039\.6/47\.038\.0/45\.536\.3/44\.127\.3/34\.335\.6/43\.136\.7/44\.620\.3/27\.034\.642\.0DocLens\([Zhu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib34)\)w/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)41\.0/48\.338\.3/45\.634\.5/41\.741\.4/49\.039\.1/46\.737\.3/45\.128\.1/35\.336\.3/43\.737\.6/45\.621\.1/27\.935\.542\.9DocSkill \(Ours\)w/ Gemini\-3\.1\-Pro\([Google, 2026](https://arxiv.org/html/2610.36759#bib.bib39)\)41\.2/48\.538\.5/45\.835\.8/43\.240\.8/48\.239\.1/46\.537\.2/44\.828\.5/35\.636\.8/44\.237\.5/45\.621\.2/28\.535\.743\.1w/ Claude\-Sonnet\-5\([Anthropic, 2026](https://arxiv.org/html/2610.36759#bib.bib10)\)42\.5/49\.839\.8/47\.236\.5/44\.142\.1/49\.540\.5/47\.838\.5/46\.229\.8/37\.237\.8/45\.538\.6/46\.822\.5/30\.136\.944\.4w/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)43\.8/51\.241\.2/48\.636\.2/43\.843\.5/51\.041\.8/49\.239\.8/47\.831\.2/38\.837\.5/45\.239\.8/48\.223\.8/31\.637\.945\.5w/ Qwen3\-VL\-4B\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7)\)\(SFT\)42\.5/49\.839\.5/46\.834\.2/41\.841\.8/49\.240\.2/47\.537\.8/45\.528\.5/35\.836\.2/43\.838\.2/46\.521\.5/29\.236\.043\.6w/ Qwen3\-VL\-4B\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7)\)\(SFT \+ DPO\)44\.2/51\.841\.5/49\.235\.8/43\.543\.8/51\.542\.5/50\.239\.5/47\.830\.8/38\.538\.2/46\.240\.2/48\.823\.5/31\.838\.045\.9w/ Qwen3\-VL\-8B\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7)\)\(SFT\)45\.2/52\.842\.8/50\.537\.5/45\.244\.8/52\.543\.5/51\.240\.8/49\.232\.5/40\.239\.5/47\.841\.5/50\.225\.2/33\.539\.347\.3w/ Qwen3\-VL\-8B\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7)\)\(SFT \+ DPO\)46\.8/54\.544\.5/52\.239\.2/47\.146\.2/54\.245\.2/53\.142\.5/51\.234\.2/42\.541\.2/49\.543\.2/52\.127\.1/35\.841\.049\.2Human BaselineHuman Experts––––––––––61\.763\.9 Table 2\.Performance on MMHop\-DocBench across ten languages, reported as per\-language and macro\-averaged Acc/𝐅𝟏\\boldsymbol\{\\mathrm\{F\}\_\{1\}\}\. Bold values indicate the best results within each method category, while underlined bold values denote the best model results overall\.MethodMMLongBench\-DocDocBenchAcc\.𝐅𝟏\\boldsymbol\{\\mathrm\{F\}\_\{1\}\}LASJLASJDocAgent\([Sun et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib13)\)w/ GPT\-4o\([OpenAI, 2024](https://arxiv.org/html/2610.36759#bib.bib8)\)51\.849\.1–79\.9DocAgent\([Sun et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib13)\)w/ Claude\-3\.5\-Sonnet\([Anthropic, 2024](https://arxiv.org/html/2610.36759#bib.bib42)\)57\.354\.1––DocDancer\([Zhang et al\., 2026](https://arxiv.org/html/2610.36759#bib.bib20)\)w/ GPT\-4o\([OpenAI, 2024](https://arxiv.org/html/2610.36759#bib.bib8)\)52\.350\.859\.273\.5DocDancer\([Zhang et al\., 2026](https://arxiv.org/html/2610.36759#bib.bib20)\)w/ Gemini\-2\.5\-Pro\([Comanici and others, 2025](https://arxiv.org/html/2610.36759#bib.bib43)\)56\.355\.365\.979\.9DocDancer\([Zhang et al\., 2026](https://arxiv.org/html/2610.36759#bib.bib20)\)w/ GPT\-5\.2\([OpenAI, 2025](https://arxiv.org/html/2610.36759#bib.bib44)\)57\.056\.867\.685\.5SimpleDoc\([Jain et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib19)\)w/ Claude\-4\-Sonnet\([Anthropic, 2025](https://arxiv.org/html/2610.36759#bib.bib41)\)58\.6–––SimpleDoc\([Jain et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib19)\)w/ GPT\-5\.2\([OpenAI, 2025](https://arxiv.org/html/2610.36759#bib.bib44)\)57\.453\.864\.281\.8DocLens\([Zhu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib34)\)w/ Claude\-4\-Sonnet\([Anthropic, 2025](https://arxiv.org/html/2610.36759#bib.bib41)\)63\.3–––DocLens\([Zhu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib34)\)w/ Gemini\-2\.5\-Pro\([Comanici and others, 2025](https://arxiv.org/html/2610.36759#bib.bib43)\)67\.6–––DocLens\([Zhu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib34)\)w/ GPT\-5\.2\([OpenAI, 2025](https://arxiv.org/html/2610.36759#bib.bib44)\)67\.855\.068\.283\.9DocSkill \(Ours\)w/ Gemini\-2\.5\-Pro\([Comanici and others, 2025](https://arxiv.org/html/2610.36759#bib.bib43)\)68\.066\.467\.884\.5w/ GPT\-5\.2\([OpenAI, 2025](https://arxiv.org/html/2610.36759#bib.bib44)\)68\.566\.867\.985\.9w/ GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\)68\.867\.168\.586\.6w/ Qwen3\-VL\-4B\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7)\)\(SFT \+ DPO\)65\.864\.564\.882\.6w/ Qwen3\-VL\-8B\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7)\)\(SFT \+ DPO\)69\.267\.368\.986\.8 Table 3\.Performance comparison on MMLongBench\-Doc and DocBench\. Underlined bold values indicate the best model results\. Unreported results are marked with “–”\. ### 4\.2\.Dataset Statistics MMHop\-DocBench contains 60K question–answer pairs spanning ten languages: English \(EN\), Chinese \(ZH\), Cantonese \(YUE\), Indonesian \(ID\), Burmese \(MY\), Filipino \(TL\), Tamil \(TA\), Thai \(TH\), Vietnamese \(VI\), and Malay \(MS\)\. The documents cover eight domains: Business and Finance, Government, Academic, Education, Technical, Medical, Legal, and Media\. Within the comparison in Table[1](https://arxiv.org/html/2610.36759#S3.T1), MMHop\-DocBench provides the widest language coverage and the highest average page count, at 68\.2 pages per document\. Each pair includes page\- and element\-level evidence annotations, enabling evaluation of evidence localization and answer grounding\. ### 4\.3\.Human Audit of Dataset Quality We conduct a human quality audit of 1,200 retained QA pairs, sampling 80 training and 40 test pairs from each of the ten languages, with stratification across document domains\. Two target\-language annotators independently assess each pair for question validity, reference\-answer correctness, evidence sufficiency, and cross\-page necessity\. Disagreements are adjudicated by a third annotator\. Of the 1,200 audited pairs, 1,144 \(95\.3%\) satisfy all four criteria, while 56 \(4\.7%\) fail at least one\. Figure 4\.Qualitative comparison of DocAgent and DocSkill on an Indonesian document\. In this example, DocAgent returns an incorrect answer with incorrect evidence citations, whereas DocSkill produces the correct answer with supporting evidence through adaptive skill selection\. Intermediate skill invocations and the terminal answer action are omitted for brevity\.Diagram showing a document and reference QA above two agent trajectories\. DocAgent returns incorrect strategy names and evidence citations\. DocSkill displays Locate, Inspect, and Verify steps, with intermediate calls omitted, and returns the five reference strategies with supporting evidence from pages 18, 31, and 38\. ## 5\.Experiments ##### Implementation Details\. We train Qwen3\-VL\-4B and Qwen3\-VL\-8B using LoRA\([Hu et al\., 2021](https://arxiv.org/html/2610.36759#bib.bib45)\)on eight NVIDIA A100 80GB GPUs\. SFT runs for three epochs with an effective global batch size of 32 across all GPUs, an image resolution of768×768768\\times 768, and AdamW with a learning rate of2×10−42\\times 10^\{\-4\}\. LoRA uses a rank of 16, a scaling factor of 32, and a dropout rate of 0\.05\. DPO uses the same LoRA configuration with an effective global batch size of 16 andβ=0\.1\\beta=0\.1, where each preferred–dispreferred trajectory pair counts as one training instance\. We use Qwen3\-VL\-Embedding\-8B\([Li et al\., 2026a](https://arxiv.org/html/2610.36759#bib.bib17)\)for top\-5 page retrieval and PP\-OCR\([Cui et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib11)\)for layout\-aware parsing, while region\-level zooming is implemented through programmatic cropping\. ##### Benchmarks and Baselines\. We evaluate DocSkill on the MMHop\-DocBench test set and two public long\-document QA benchmarks, MMLongBench\-Doc\([Ma et al\., 2024](https://arxiv.org/html/2610.36759#bib.bib3)\)and DocBench\([Zou et al\., 2024](https://arxiv.org/html/2610.36759#bib.bib4)\)\. On MMHop\-DocBench, we compare against direct VLM inference, OCR\-based pipelines, multimodal RAG systems, and document agents, using official implementations and recommended settings when available\. Each DocSkill variant uses its listed backbone for both orchestration and skill execution, with the specified SFT/DPO checkpoints for Qwen variants and prompting for commercial variants\. ##### Evaluation Metrics\. For MMHop\-DocBench, we report accuracy \(Acc\.\) and token\-levelF1\\mathrm\{F\}\_\{1\}per\-language, with macro averages over all ten languages\. Accuracy is computed from normalized predictions and references, while language\-aware tokenization is applied for multilingualF1\\mathrm\{F\}\_\{1\}\. For MMLongBench\-Doc, we report accuracy,F1\\mathrm\{F\}\_\{1\}, and LLM\-as\-a\-Judge \(LASJ\), while DocBench is evaluated using LASJ\. All LASJ scores follow the judge prompts and configurations provided in the respective official code repositories\. ##### Human Baseline\. We establish a human baseline using 1,300 test questions, with 130 questions sampled from each of the ten languages\. Model results are evaluated on the full test set, while human results provide a reference on this subset\. ### 5\.1\.Main Results ##### Results on MMHop\-DocBench\. Table[2](https://arxiv.org/html/2610.36759#S4.T2)reports results on MMHop\-DocBench\. In same\-backbone comparisons across Gemini\-3\.1\-Pro\([Google, 2026](https://arxiv.org/html/2610.36759#bib.bib39)\), Claude\-Sonnet\-5\([Anthropic, 2026](https://arxiv.org/html/2610.36759#bib.bib10)\), and GPT\-5\.6\-SOL\([OpenAI, 2026](https://arxiv.org/html/2610.36759#bib.bib9)\), DocSkill improves over the corresponding DocLens\([Zhu et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib34)\)variants by 2\.3–2\.5 Acc\. and 2\.4–2\.7F1\\mathrm\{F\}\_\{1\}points\. DocSkill with Qwen3\-VL\-8B\([Bai et al\., 2025](https://arxiv.org/html/2610.36759#bib.bib7)\), trained through SFT followed by DPO\([Rafailov et al\., 2024](https://arxiv.org/html/2610.36759#bib.bib23)\), achieves the highest macro\-averaged scores among the evaluated model configurations: 41\.0 Acc\. and 49\.2F1\\mathrm\{F\}\_\{1\}\. It exceeds DocLens with GPT\-5\.6\-SOL by 5\.5 Acc\. and 6\.3F1\\mathrm\{F\}\_\{1\}points\. This comparison evaluates complete configurations with different backbones and training conditions\. Human experts achieve 61\.7 Acc\. and 63\.9F1\\mathrm\{F\}\_\{1\}on the sampled test subset, providing a reference\. ##### Generalization to Public Benchmarks\. Table[3](https://arxiv.org/html/2610.36759#S4.T3)reports results on MMLongBench\-Doc and DocBench\. DocSkill with Qwen3\-VL\-8B, trained on MMHop\-DocBench through SFT followed by DPO, achieves 69\.2 Acc\., 67\.3F1\\mathrm\{F\}\_\{1\}, and 68\.9 LASJ on MMLongBench\-Doc, and 86\.8 LASJ on DocBench\. Compared with the strongest competing baseline listed for each metric, these results yield gains of 1\.4, 10\.5, 0\.7, and 1\.3 points, respectively\. Using GPT\-5\.2 as the controller, DocSkill exceeds the reported results of DocDancer\([Zhang et al\., 2026](https://arxiv.org/html/2610.36759#bib.bib20)\)by 11\.5 Acc\., 10\.0F1\\mathrm\{F\}\_\{1\}, and 0\.3 LASJ on MMLongBench\-Doc, and by 0\.4 LASJ on DocBench\. ### 5\.2\.Qualitative Analysis Figure[4](https://arxiv.org/html/2610.36759#S4.F4)illustrates a case from a 67\-page Indonesian document\. The question requires combining complementary evidence from text and tables across pages 18, 31, and 38\. DocAgent searches and inspects pages sequentially but stops after collecting incomplete evidence from pages 17 and 31\. Consequently, it introduces unsupported content and cites incorrect evidence\. In contrast, DocSkill gathers and integrates evidence across all required pages and verifies the answer, enabling complete, grounded cross\-page reasoning\. ### 5\.3\.Ablation Studies ##### Effect of Skill Abstraction and Adaptive Orchestration\. We compare DocSkill with adaptive direct tool selection, a fixed\-order skill policy, and four leave\-one\-out variants\. All variants use Qwen3\-VL\-8B with SFT and DPO\. Direct tool selection exposes the underlying tools as actions while retaining reasoning for cross\-page integration, answer verification, and reflection, without explicit interfaces for the four document skills\. The fixed policy allows repeated cycles ofLocate→\\rightarrowInspect→\\rightarrowIntegrate→\\rightarrowVerify, while maintaining the prescribed execution order\. Each leave\-one\-out variant removes one skill during search and training\.RpageR\_\{\\mathrm\{page\}\}andRelemR\_\{\\mathrm\{elem\}\}measure the percentages of annotated supporting pages and elements recovered, respectively\. Table[4](https://arxiv.org/html/2610.36759#S5.T4)shows that DocSkill outperforms direct tool selection by 11\.2 Acc\. and 11\.0F1\\mathrm\{F\}\_\{1\}points, and the fixed policy by 7\.5 Acc\. and 7\.4F1\\mathrm\{F\}\_\{1\}points\. RemovingLocatecauses the largest Acc\. drop \(17\.6 points\), reducing page recall from 82\.4 to 41\.5; removingInspectlowers element recall to 43\.5\. This pattern suggests that relevant\-page retrieval alone is insufficient without fine\-grained evidence inspection\. RemovingIntegrateorVerifyreduces Acc\. by 9\.2 and 4\.5 points, respectively\. These results support the effectiveness of the full DocSkill configuration under the evaluated settings\. VariantAcc\.𝑭𝟏\\boldsymbol\{F\_\{1\}\}𝑹𝐩𝐚𝐠𝐞\\boldsymbol\{R\_\{\\mathrm\{page\}\}\}𝑹𝐞𝐥𝐞𝐦\\boldsymbol\{R\_\{\\mathrm\{elem\}\}\}Full DocSkill41\.049\.282\.468\.5Direct Tool Selection29\.838\.268\.549\.3Fixed Skill Order33\.541\.870\.456\.5DocSkill w/oLocate23\.430\.241\.533\.2DocSkill w/oInspect28\.236\.172\.843\.5DocSkill w/oIntegrate31\.839\.576\.259\.8DocSkill w/oVerify36\.544\.879\.665\.2Table 4\.Ablation of DocSkill on MMHop\-DocBench using Qwen3\-VL\-8B \(SFT \+ DPO\)\. Scores are percentages\.RpageR\_\{\\mathrm\{page\}\}andRelemR\_\{\\mathrm\{elem\}\}denote supporting\-page and supporting\-element recall\. ##### Effect of Trajectory Preference Supervision\. We examine whether suboptimal trajectories provide useful supervision by comparing SFT with SFT followed by DPO\. Since MCTS naturally produces ranked trajectory pairs, DPO can directly optimize this offline supervision without additional on\-policy rollouts or reward\-model evaluations\. As shown in Table[2](https://arxiv.org/html/2610.36759#S4.T2), DPO improves Qwen3\-VL\-4B from 36\.0 to 38\.0 Acc\. and from 43\.6 to 45\.9F1\\mathrm\{F\}\_\{1\}\. For Qwen3\-VL\-8B, it improves Acc\. from 39\.3 to 41\.0 andF1\\mathrm\{F\}\_\{1\}from 47\.3 to 49\.2\. These gains show that preferences between successful and suboptimal trajectories provide complementary supervision beyond imitating successful trajectories alone\. Table 5\.Equal\-budget trajectory construction comparison\. Succ\. means the questions with a correct trajectory \(%\)\. Search metrics are mean±\\pmSD over three runs; Acc\./F1\\mathrm\{F\}\_\{1\}are MMHop\-DocBench scores for DocSkill w/ Qwen3\-VL\-8B \(SFT \+ DPO\)\. ##### Effect of MCTS\-Based Trajectory Construction\. Under matched rollout and evaluation budgets, we compare MCTS with Best\-of\-NNsampling\([Cobbe et al\., 2021](https://arxiv.org/html/2610.36759#bib.bib46)\)and evaluator\-guided beam search\([Xie et al\., 2023](https://arxiv.org/html/2610.36759#bib.bib35)\), using the same frozen GPT\-5\.6\-SOL generator and evaluator\. Seq\. Div\. is the mean per\-question fraction of distinct skill sequences among valid trajectories, while Pairs/Q is the average number of preference pairs per question\. Table[5](https://arxiv.org/html/2610.36759#S5.T5)shows that MCTS leads in success rate \(65\.8%\), diversity, and Pairs/Q\. Its Qwen3\-VL\-8B policy trained through SFT and DPO reaches 41\.0 Acc\. and 49\.2F1\\mathrm\{F\}\_\{1\}, outperforming Best\-of\-NNand beam search by 3\.2/2\.7 and 2\.4/2\.1 Acc\./F1\\mathrm\{F\}\_\{1\}points, respectively\. These gains also reflect higher supervision yield from MCTS\. ## 6\.Conclusion We presented DocSkill, a document agent that learns to orchestrate four reusable capability\-level skills for multilingual multi\-hop reasoning\. Bounded MCTS constructs skill trajectories for SFT and DPO\. We also introduced MMHop\-DocBench, containing 60K grounded QA pairs across 10 languages and eight document domains, with page\- and element\-level evidence annotations\. Experiments on MMHop\-DocBench, MMLongBench\-Doc, and DocBench demonstrate consistent improvements over strong baselines, while ablations support skill abstraction, adaptive orchestration, and trajectory supervision\. Future work will expand language and domain coverage and explore automatic discovery and refinement of reusable document skills for more complex document scenarios\. ## References - Anthropic \(2024\)AnthropicClaude 3\.5 Sonnet Model Card\.Note:[https://www\.anthropic\.com/system\-cards](https://www.anthropic.com/system-cards)Cited by:[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.4.1)\. - Anthropic \(2025\)AnthropicClaude 4 System Model Card\.Note:[https://www\.anthropic\.com/system\-cards](https://www.anthropic.com/system-cards)Cited by:[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.10.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.8.1)\. - Anthropic \(2026\)AnthropicClaude Sonnet 5 System Card\.Note:[https://www\.anthropic\.com/system\-cards](https://www.anthropic.com/system-cards)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.20.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.24.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.7.1),[§5\.1](https://arxiv.org/html/2610.36759#S5.SS1.SSS0.Px1.p1.1)\. - Baiet al\.\(2025\)S\. Bai, Y\. Cai, R\. Chen, K\. Chen, X\. Chen, Z\. Cheng, L\. Deng, W\. Ding, C\. Gao, C\. Ge, W\. Ge, Z\. Guo, Q\. Huang, J\. Huang, F\. Huang, B\. Hui, S\. Jiang, Z\. Li, M\. Li, M\. Li, K\. Li, Z\. Lin, J\. Lin, X\. Liu, J\. Liu, C\. Liu, Y\. Liu, D\. Liu, S\. Liu, D\. Lu, R\. Luo, C\. Lv, R\. Men, L\. Meng, X\. Ren, X\. Ren, S\. Song, Y\. Sun, J\. Tang, J\. Tu, J\. Wan, P\. Wang, P\. Wang, Q\. Wang, Y\. Wang, T\. Xie, Y\. Xu, H\. Xu, J\. Xu, Z\. Yang, M\. Yang, J\. Yang, A\. Yang, B\. Yu, F\. Zhang, H\. Zhang, X\. Zhang, B\. Zheng, H\. Zhong, J\. Zhou, F\. Zhou, J\. Zhou, Y\. Zhu, and K\. ZhuQwen3\-vl technical report\.External Links:2511\.21631,[Link](https://arxiv.org/abs/2511.21631)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.26.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.27.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.28.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.29.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.3.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.4.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.5.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.17.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.18.1),[§5\.1](https://arxiv.org/html/2610.36759#S5.SS1.SSS0.Px1.p1.1)\. - Chenet al\.\(2024\)G\. Chen, M\. Liao, C\. Li, and K\. FanStep\-level value preference optimization for mathematical reasoning\.External Links:2406\.10858,[Link](https://arxiv.org/abs/2406.10858)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p4.1)\. - Choet al\.\(2024\)J\. Cho, D\. Mahata, O\. Irsoy, Y\. He, and M\. BansalM3DocRAG: multi\-modal retrieval is what you need for multi\-page multi\-document understanding\.External Links:2411\.04952,[Link](https://arxiv.org/abs/2411.04952)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1)\. - Cobbeet al\.\(2021\)K\. Cobbe, V\. Kosaraju, M\. Bavarian, M\. Chen, H\. Jun, L\. Kaiser, M\. Plappert, J\. Tworek, J\. Hilton, R\. Nakano, C\. Hesse, and J\. SchulmanTraining verifiers to solve math word problems\.External Links:2110\.14168,[Link](https://arxiv.org/abs/2110.14168)Cited by:[§5\.3](https://arxiv.org/html/2610.36759#S5.SS3.SSS0.Px3.p1.1)\. - Comaniciet al\.\(2025\)G\. Comaniciet al\.Gemini 2\.5: pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities\.External Links:2507\.06261,[Link](https://arxiv.org/abs/2507.06261)Cited by:[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.11.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.14.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.6.1)\. - Cuiet al\.\(2025\)C\. Cui, T\. Sun, M\. Lin, T\. Gao, Y\. Zhang, J\. Liu, X\. Wang, Z\. Zhang, C\. Zhou, H\. Liu, Y\. Zhang, W\. Lv, K\. Huang, Y\. Zhang, J\. Zhang, J\. Zhang, Y\. Liu, D\. Yu, and Y\. MaPaddleOCR 3\.0 technical report\.External Links:2507\.05595,[Link](https://arxiv.org/abs/2507.05595)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2610.36759#S4.SS1.p1.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.11.1.1),[§5](https://arxiv.org/html/2610.36759#S5.SS0.SSS0.Px1.p1.1)\. - Faysseet al\.\(2025\)M\. Faysse, H\. Sibille, T\. Wu, B\. Omrani, G\. Viaud, C\. Hudelot, and P\. ColomboColPali: efficient document retrieval with vision language models\.External Links:2407\.01449,[Link](https://arxiv.org/abs/2407.01449)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.14.1)\. - Google \(2026\)GoogleGemini 3\.1 pro model card\.Note:[https://storage\.googleapis\.com/deepmind\-media/Model\-Cards/Gemini\-3\-1\-Pro\-Model\-Card\.pdf](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-1-Pro-Model-Card.pdf)Published February 2026Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.19.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.23.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.6.1),[§5\.1](https://arxiv.org/html/2610.36759#S5.SS1.SSS0.Px1.p1.1)\. - Guoet al\.\(2025\)Z\. Guo, X\. Ren, L\. Xu, J\. Zhang, and C\. HuangRAG\-anything: all\-in\-one rag framework\.External Links:2510\.12323,[Link](https://arxiv.org/abs/2510.12323)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1)\. - Hanet al\.\(2025\)S\. Han, P\. Xia, R\. Zhang, T\. Sun, Y\. Li, H\. Zhu, and H\. YaoMDocAgent: a multi\-modal multi\-agent framework for document understanding\.External Links:2503\.13964,[Link](https://arxiv.org/abs/2503.13964)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1)\. - Hoet al\.\(2020\)X\. Ho, A\. D\. Nguyen, S\. Sugawara, and A\. AizawaConstructing a multi\-hop qa dataset for comprehensive evaluation of reasoning steps\.External Links:2011\.01060,[Link](https://arxiv.org/abs/2011.01060)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p5.1)\. - Huet al\.\(2021\)E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. ChenLoRA: low\-rank adaptation of large language models\.External Links:2106\.09685,[Link](https://arxiv.org/abs/2106.09685)Cited by:[§5](https://arxiv.org/html/2610.36759#S5.SS0.SSS0.Px1.p1.1)\. - Huanget al\.\(2025\)J\. Huang, Z\. Tan, S\. Gong, F\. Zeng, J\. T\. Zhou, C\. Miao, H\. Tan, W\. Yao, and J\. LiLaV\-cot: language\-aware visual cot with multi\-aspect reward optimization for real\-world multilingual vqa\.External Links:2509\.10026,[Link](https://arxiv.org/abs/2509.10026)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p1.1)\. - Jainet al\.\(2025\)C\. Jain, Y\. Wu, Y\. Zeng, J\. Liu, S\. hengyu Dai, Z\. Shao, Q\. Wu, and H\. WangSimpleDoc: multi\-modal document understanding with dual\-cue page retrieval and iterative refinement\.External Links:2506\.14035,[Link](https://arxiv.org/abs/2506.14035)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.18.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.8.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.9.1)\. - Jianget al\.\(2026\)G\. Jiang, Z\. Su, X\. Qu, and Y\. R\. FungXSkill: continual learning from experience and skills in multimodal agents\.External Links:2603\.12056,[Link](https://arxiv.org/abs/2603.12056)Cited by:[§2\.2](https://arxiv.org/html/2610.36759#S2.SS2.p1.1)\. - Lassouedet al\.\(2026\)A\. Lassoued, M\. A\. Souibgui, and Y\. KessentiniORCA: orchestrated reasoning with collaborative agents for document visual question answering\.External Links:2603\.02438,[Link](https://arxiv.org/abs/2603.02438)Cited by:[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1)\. - Liet al\.\(2026a\)M\. Li, Y\. Zhang, D\. Long, K\. Chen, S\. Song, S\. Bai, Z\. Yang, P\. Xie, A\. Yang, D\. Liu, J\. Zhou, and J\. LinQwen3\-vl\-embedding and qwen3\-vl\-reranker: a unified framework for state\-of\-the\-art multimodal retrieval and ranking\.External Links:2601\.04720,[Link](https://arxiv.org/abs/2601.04720)Cited by:[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.15.1),[§5](https://arxiv.org/html/2610.36759#S5.SS0.SSS0.Px1.p1.1)\. - Liet al\.\(2026b\)Z\. Li, H\. Tian, L\. Luo, Y\. Cao, and P\. LuoDeepRead: document structure\-aware reasoning to enhance agentic search\.External Links:2602\.05014,[Link](https://arxiv.org/abs/2602.05014)Cited by:[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1)\. - Maet al\.\(2024\)Y\. Ma, Y\. Zang, L\. Chen, M\. Chen, Y\. Jiao, X\. Li, X\. Lu, Z\. Liu, Y\. Ma, X\. Dong, P\. Zhang, L\. Pan, Y\. Jiang, J\. Wang, Y\. Cao, and A\. SunMMLongBench\-doc: benchmarking long\-context document understanding with visualizations\.External Links:2407\.01523,[Link](https://arxiv.org/abs/2407.01523)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p1.1),[Table 1](https://arxiv.org/html/2610.36759#S3.T1.2.1.8.1),[§5](https://arxiv.org/html/2610.36759#S5.SS0.SSS0.Px2.p1.1)\. - Masryet al\.\(2022\)A\. Masry, D\. X\. Long, J\. Q\. Tan, S\. Joty, and E\. HoqueChartQA: a benchmark for question answering about charts with visual and logical reasoning\.External Links:2203\.10244,[Link](https://arxiv.org/abs/2203.10244)Cited by:[Table 1](https://arxiv.org/html/2610.36759#S3.T1.2.1.2.1)\. - Mathewet al\.\(2021\)M\. Mathew, D\. Karatzas, and C\. V\. JawaharDocVQA: a dataset for vqa on document images\.External Links:2007\.00398,[Link](https://arxiv.org/abs/2007.00398)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p1.1),[Table 1](https://arxiv.org/html/2610.36759#S3.T1.2.1.5.1)\. - OpenAI \(2024\)OpenAIGPT\-4o system card\.Note:[https://openai\.com/index/gpt\-4o\-system\-card/](https://openai.com/index/gpt-4o-system-card/)Cited by:[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.3.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.5.1)\. - OpenAI \(2025\)OpenAIUpdate to GPT\-5 system card: GPT\-5\.2\.External Links:[Link](https://openai.com/index/gpt-5-system-card-update-gpt-5-2/)Cited by:[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.12.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.15.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.7.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.9.1)\. - OpenAI \(2026\)OpenAIGPT\-5\.6 System Card\.Note:[https://deploymentsafety\.openai\.com/gpt\-5\-6](https://deploymentsafety.openai.com/gpt-5-6)Published July 2026Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1),[§3\.2](https://arxiv.org/html/2610.36759#S3.SS2.SSS0.Px4.p2.1),[§4\.1](https://arxiv.org/html/2610.36759#S4.SS1.p1.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.10.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.11.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.13.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.14.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.15.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.17.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.18.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.21.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.25.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.8.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.16.1),[§5\.1](https://arxiv.org/html/2610.36759#S5.SS1.SSS0.Px1.p1.1)\. - Patilet al\.\(2023\)S\. G\. Patil, T\. Zhang, X\. Wang, and J\. E\. GonzalezGorilla: large language model connected with massive apis\.External Links:2305\.15334,[Link](https://arxiv.org/abs/2305.15334)Cited by:[§2\.2](https://arxiv.org/html/2610.36759#S2.SS2.p1.1)\. - Rafailovet al\.\(2024\)R\. Rafailov, A\. Sharma, E\. Mitchell, S\. Ermon, C\. D\. Manning, and C\. FinnDirect preference optimization: your language model is secretly a reward model\.External Links:2305\.18290,[Link](https://arxiv.org/abs/2305.18290)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p4.1),[§5\.1](https://arxiv.org/html/2610.36759#S5.SS1.SSS0.Px1.p1.1)\. - Schicket al\.\(2023\)T\. Schick, J\. Dwivedi\-Yu, R\. Dessì, R\. Raileanu, M\. Lomeli, L\. Zettlemoyer, N\. Cancedda, and T\. ScialomToolformer: language models can teach themselves to use tools\.External Links:2302\.04761,[Link](https://arxiv.org/abs/2302.04761)Cited by:[§2\.2](https://arxiv.org/html/2610.36759#S2.SS2.p1.1)\. - Silveret al\.\(2017\)D\. Silver, T\. Hubert, J\. Schrittwieser, I\. Antonoglou, M\. Lai, A\. Guez, M\. Lanctot, L\. Sifre, D\. Kumaran, T\. Graepel, T\. Lillicrap, K\. Simonyan, and D\. HassabisMastering chess and shogi by self\-play with a general reinforcement learning algorithm\.External Links:1712\.01815,[Link](https://arxiv.org/abs/1712.01815)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p4.1),[§3\.2](https://arxiv.org/html/2610.36759#S3.SS2.p1.1)\. - Smith \(2007\)R\. SmithAn overview of the tesseract ocr engine\.IEEE\.Cited by:[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.10.1.1)\. - Sunet al\.\(2025\)L\. Sun, L\. He, S\. Jia, Y\. He, and C\. YouDocAgent: an agentic framework for multi\-modal long\-context document understanding\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.893)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.17.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.3.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.4.1)\. - Tanakaet al\.\(2023\)R\. Tanaka, K\. Nishida, K\. Nishida, T\. Hasegawa, I\. Saito, and K\. SaitoSlideVQA: a dataset for document visual question answering on multiple images\.External Links:2301\.04883,[Link](https://arxiv.org/abs/2301.04883)Cited by:[Table 1](https://arxiv.org/html/2610.36759#S3.T1.2.1.7.1)\. - Tanget al\.\(2025\)J\. Tang, Q\. Liu, Y\. Ye, J\. Lu, S\. Wei, C\. Lin, W\. Li, M\. F\. F\. B\. Mahmood, H\. Feng, Z\. Zhao, Y\. He, K\. Lu, Y\. Wang, Y\. Liu, H\. Liu, X\. Bai, and C\. HuangMTVQA: benchmarking multilingual text\-centric visual question answering\.External Links:2405\.11985,[Link](https://arxiv.org/abs/2405.11985)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p1.1),[Table 1](https://arxiv.org/html/2610.36759#S3.T1.2.1.4.1)\. - Tanget al\.\(2023\)Q\. Tang, Z\. Deng, H\. Lin, X\. Han, Q\. Liang, B\. Cao, and L\. SunToolAlpaca: generalized tool learning for language models with 3000 simulated cases\.External Links:2306\.05301,[Link](https://arxiv.org/abs/2306.05301)Cited by:[§2\.2](https://arxiv.org/html/2610.36759#S2.SS2.p1.1)\. - Titoet al\.\(2023\)R\. Tito, D\. Karatzas, and E\. ValvenyHierarchical multimodal transformers for multi\-page docvqa\.External Links:2212\.05935,[Link](https://arxiv.org/abs/2212.05935)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p1.1),[Table 1](https://arxiv.org/html/2610.36759#S3.T1.2.1.6.1)\. - Wanget al\.\(2020\)X\. Wang, Y\. Liu, C\. Shen, C\. C\. Ng, C\. Luo, L\. Jin, C\. S\. Chan, A\. van den Hengel, and L\. WangOn the general value of evidence, and bilingual scene\-text visual question answering\.External Links:2002\.10215,[Link](https://arxiv.org/abs/2002.10215)Cited by:[Table 1](https://arxiv.org/html/2610.36759#S3.T1.2.1.3.1)\. - Xiaoet al\.\(2019\)Y\. Xiao, Y\. Qu, L\. Qiu, H\. Zhou, L\. Li, W\. Zhang, and Y\. YuDynamically fused graph network for multi\-hop reasoning\.External Links:1905\.06933,[Link](https://arxiv.org/abs/1905.06933)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p5.1)\. - Xieet al\.\(2024\)Y\. Xie, A\. Goyal, W\. Zheng, M\. Kan, T\. P\. Lillicrap, K\. Kawaguchi, and M\. ShiehMonte carlo tree search boosts reasoning via iterative preference learning\.External Links:2405\.00451,[Link](https://arxiv.org/abs/2405.00451)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p4.1)\. - Xieet al\.\(2023\)Y\. Xie, K\. Kawaguchi, Y\. Zhao, X\. Zhao, M\. Kan, J\. He, and Q\. XieSelf\-evaluation guided beam search for reasoning\.External Links:2305\.00633,[Link](https://arxiv.org/abs/2305.00633)Cited by:[§5\.3](https://arxiv.org/html/2610.36759#S5.SS3.SSS0.Px3.p1.1)\. - Yanget al\.\(2026\)Y\. Yang, J\. Li, Q\. Pan, B\. Zhan, Y\. Cai, L\. Du, J\. Zhou, K\. Chen, Q\. Chen, X\. Li, B\. Zhang, and L\. HeAutoSkill: experience\-driven lifelong learning via skill self\-evolution\.External Links:2603\.01145,[Link](https://arxiv.org/abs/2603.01145)Cited by:[§2\.2](https://arxiv.org/html/2610.36759#S2.SS2.p1.1)\. - Yaoet al\.\(2023a\)S\. Yao, D\. Yu, J\. Zhao, I\. Shafran, T\. L\. Griffiths, Y\. Cao, and K\. NarasimhanTree of thoughts: deliberate problem solving with large language models\.External Links:2305\.10601,[Link](https://arxiv.org/abs/2305.10601)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p4.1),[§3\.2](https://arxiv.org/html/2610.36759#S3.SS2.p1.1)\. - Yaoet al\.\(2023b\)S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. CaoReAct: synergizing reasoning and acting in language models\.External Links:2210\.03629,[Link](https://arxiv.org/abs/2210.03629)Cited by:[§2\.2](https://arxiv.org/html/2610.36759#S2.SS2.p1.1),[§3\.1](https://arxiv.org/html/2610.36759#S3.SS1.p1.1)\. - Yuet al\.\(2025\)S\. Yu, C\. Tang, B\. Xu, J\. Cui, J\. Ran, Y\. Yan, Z\. Liu, S\. Wang, X\. Han, Z\. Liu, and M\. SunVisRAG: vision\-based retrieval\-augmented generation on multi\-modality documents\.External Links:2410\.10594,[Link](https://arxiv.org/abs/2410.10594)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.13.1)\. - Zhanget al\.\(2026\)Q\. Zhang, X\. Lv, J\. Wu, B\. Li, Z\. Tao, G\. Yan, H\. Zhang, B\. Wang, J\. Xu, H\. Mi, and W\. ZhangDocDancer: towards agentic document\-grounded information seeking\.External Links:2601\.05163,[Link](https://arxiv.org/abs/2601.05163)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p2.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.5.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.6.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.7.1),[§5\.1](https://arxiv.org/html/2610.36759#S5.SS1.SSS0.Px2.p1.1)\. - Zhenget al\.\(2026\)Y\. Zheng, P\. Fu, H\. Li, Z\. Wang, Y\. Zhang, W\. Ruan, X\. Zhang, Z\. Wei, Z\. Luo, J\. Luan, W\. Chen, and X\. BaiDoc\-v\*:coarse\-to\-fine interactive visual reasoning for multi\-page document vqa\.External Links:2604\.13731,[Link](https://arxiv.org/abs/2604.13731)Cited by:[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1)\. - Zhuet al\.\(2025\)D\. Zhu, R\. Meng, J\. Chen, S\. Li, T\. Pfister, and J\. YoonDocLens : a tool\-augmented multi\-agent framework for long visual document understanding\.External Links:2511\.11552,[Link](https://arxiv.org/abs/2511.11552)Cited by:[§2\.1](https://arxiv.org/html/2610.36759#S2.SS1.p1.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.19.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.20.1),[Table 2](https://arxiv.org/html/2610.36759#S4.T2.2.1.21.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.10.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.11.1),[Table 3](https://arxiv.org/html/2610.36759#S4.T3.2.1.12.1),[§5\.1](https://arxiv.org/html/2610.36759#S5.SS1.SSS0.Px1.p1.1)\. - Zouet al\.\(2024\)A\. Zou, W\. Yu, H\. Zhang, K\. Ma, D\. Cai, Z\. Zhang, H\. Zhao, and D\. YuDOCBENCH: a benchmark for evaluating llm\-based document reading systems\.External Links:2407\.10701,[Link](https://arxiv.org/abs/2407.10701)Cited by:[§1](https://arxiv.org/html/2610.36759#S1.p1.1),[Table 1](https://arxiv.org/html/2610.36759#S3.T1.2.1.9.1),[§5](https://arxiv.org/html/2610.36759#S5.SS0.SSS0.Px2.p1.1)\. ## Appendix AMore Details of MCTS Trajectory Generation ### A\.1\.Prompt for MCTS Rollout Generation Figure 5\.Instructions for generating MCTS rollouts\.Prompt instructing the MCTS generator to propose skill invocations or a terminal answer based on the current interaction state\.Figure 6\.Instructions for evaluating MCTS rollouts\.Prompt instructing the MCTS evaluator to assess terminal outcomes, language consistency, logical soundness, and skill\-use effectiveness\. ## Appendix BMore Details of HopGen ### B\.1\.HopGen Construction Pipeline HopGen automatically constructs grounded multilingual multi\-hop QA pairs from long documents\. Algorithm[1](https://arxiv.org/html/2610.36759#alg1)formalizes the automatic curation procedure, comprising hierarchical graph construction, cross\-page evidence\-path extraction, candidate generation, and quality\-based filtering\. Algorithm 1HopGen: MMHop\-DocQA Curation0:Multilingual document DD, target language ℓ\\ell, OCR module focrf\_\{\\mathrm\{ocr\}\}, generator fgenf\_\{\\mathrm\{gen\}\}, evaluator fevalf\_\{\\mathrm\{eval\}\}, quality threshold τ\\tau 0:Verified QA set 𝒬=\{⟨q,a,Ep,Ee⟩\}\\mathcal\{Q\}=\\\{\\langle q,a,E^\{p\},E^\{e\}\\rangle\\\} 1: 𝒬←∅\\mathcal\{Q\}\\leftarrow\\emptyset 2: 𝒫←SplitPages\(D\)\\mathcal\{P\}\\leftarrow\\textsc\{SplitPages\}\(D\) 3:foreach page pi∈𝒫p\_\{i\}\\in\\mathcal\{P\}do 4: hi←CreatePageProfile\(pi,ℓ,fgen\)h\_\{i\}\\leftarrow\\textsc\{CreatePageProfile\}\(p\_\{i\},\\ell;f\_\{\\mathrm\{gen\}\}\) 5:endfor 6: ℋ←\{hi\}i=1\|𝒫\|\\mathcal\{H\}\\leftarrow\\\{h\_\{i\}\\\}\_\{i=1\}^\{\|\\mathcal\{P\}\|\} 7: 𝒢p←BuildPageGraph\(𝒫,ℋ,fgen\)\\mathcal\{G\}\_\{p\}\\leftarrow\\textsc\{BuildPageGraph\}\(\\mathcal\{P\},\\mathcal\{H\};f\_\{\\mathrm\{gen\}\}\) 8:foreach page pi∈𝒫p\_\{i\}\\in\\mathcal\{P\}do 9: ℰi←focr\(pi\)\\mathcal\{E\}\_\{i\}\\leftarrow f\_\{\\mathrm\{ocr\}\}\(p\_\{i\}\) 10:endfor 11: ℰ←⋃i=1\|𝒫\|ℰi\\mathcal\{E\}\\leftarrow\\displaystyle\\bigcup\_\{i=1\}^\{\|\\mathcal\{P\}\|\}\\mathcal\{E\}\_\{i\} 12: 𝒢pe←BuildPageElementGraph\(𝒢p,ℰ,fgen\)\\mathcal\{G\}\_\{pe\}\\leftarrow\\textsc\{BuildPageElementGraph\}\(\\mathcal\{G\}\_\{p\},\\mathcal\{E\};f\_\{\\mathrm\{gen\}\}\) 13: Π←ExtractCrossPagePaths\(𝒢pe\)\\Pi\\leftarrow\\textsc\{ExtractCrossPagePaths\}\(\\mathcal\{G\}\_\{pe\}\) 14:foreach evidence path π∈Π\\pi\\in\\Pido 15: 𝒞π←GenerateCandidates\(D,π,ℓ,fgen\)\\mathcal\{C\}\_\{\\pi\}\\leftarrow\\textsc\{GenerateCandidates\}\(D,\\pi,\\ell;f\_\{\\mathrm\{gen\}\}\) 16:foreach candidate c=⟨q,a,Ep,Ee⟩∈𝒞πc=\\langle q,a,E^\{p\},E^\{e\}\\rangle\\in\\mathcal\{C\}\_\{\\pi\}do 17: r←EvaluateCandidate\(D,c,ℓ,feval\)r\\leftarrow\\textsc\{EvaluateCandidate\}\(D,c,\\ell;f\_\{\\mathrm\{eval\}\}\) 18:if Quality\(r\)≥τ\\textsc\{Quality\}\(r\)\\geq\\tauthen 19: 𝒬←𝒬∪\{c\}\\mathcal\{Q\}\\leftarrow\\mathcal\{Q\}\\cup\\\{c\\\} 20:endif 21:endfor 22:endfor 23:return 𝒬\\mathcal\{Q\} The Page Graph captures semantic dependencies between page profiles, while the Page–Element Graph grounds them in localized text, tables, figures, captions, and equations\. Each evidence path spans two to four pages, with edges retained only when supported by explicit semantic and grounded element\-level relations\. HopGen generates at most three QA candidates per path\. The evaluator scores answer correctness, evidence sufficiency, cross\-page necessity, and language consistency, retaining candidates with thresholdτ=0\.875\\tau=0\.875\. After question and answer normalization, duplicates are merged by retaining the highest\-scoring candidate\. Each final QA pair preserves its supporting page and element annotations\. ### B\.2\.Prompt for HopGen Pipeline Figure 7\.Instructions for HopGen QA generation\.Prompt instructing the HopGen generator to produce grounded question\-answer pairs in the target language from cross\-page evidence, with supporting page and element IDs\.Figure 8\.Representative multilingual samples from MMHop\-DocBench\. Each example shows selected document pages, a target\-language QA pair, its English translation, and supporting page\- and element\-level annotations\. The examples illustrate cross\-page evidence composition over text, tables, and images\.Multilingual document QA examples showing selected pages, target\-language questions and answers, English translations, and supporting page and element annotations\. The examples combine evidence across pages from text, tables, and images\.Figure 9\.Instructions for HopGen QA evaluation\.Prompt instructing the HopGen evaluator to assess answer correctness, evidence sufficiency, cross\-page necessity, and language consistency, with a score and justification for each criterion\. ## Appendix CMore Details of MMHop\-DocBench ### C\.1\.Dataset Size and Splits Figure[10](https://arxiv.org/html/2610.36759#A3.F10)summarizes the document and QA\-pair counts across ten languages\. MMHop\-DocBench contains 2,299 documents and 60K QA pairs\. The training split contains 1,599 documents and 40,924 QA pairs, while the test split contains 700 documents and 19,132 QA pairs\. We split the dataset at the source\-document level, assigning all QA pairs derived from each document to the same split\. Training QA counts are approximately balanced across languages, although the number of retained questions per document varies after filtering\. Figure 10\.Document and QA\-pair counts by language in the training and test splits of MMHop\-DocBench\.Document and QA\-pair counts across ten languages, shown separately for the training and test splits\. ### C\.2\.Dataset Samples Figure[8](https://arxiv.org/html/2610.36759#A2.F8)shows representative Chinese, Indonesian, Thai, and Vietnamese examples combining cross\-page textual, tabular, and visual evidence\. Each includes a target\-language QA pair, its English translation, and explicit page\- and element\-level evidence annotations\.
Similar Articles
TS-Skill: A Benchmark for Evaluating Analytical Skills in Time-Series Question Answering
TS-Skill introduces a controlled benchmark with three composable analytical skills for time-series question answering, and experiments on LLMs and TSLMs reveal significant capability gaps, especially in cross-interval integration.
COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation
This paper presents COLLEAGUE.SKILL, an open-source system for automatically distilling person-grounded AI skills from heterogeneous traces into inspectable, correctable, and portable skill packages, enabling LLM agents to carry bounded representations of human expertise and interaction style.
Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction
Work2Skill 提出一个框架，通过 WorkRecorder 收集人类 GUI 操作演示，构建 WorkDemo-350 多模态数据集，并用智能体式技能归纳方法从人类经验中提炼可复用的智能体技能，在 MyPCBench 和 OSWorld-Verified 上显著提升了多种执行模型的任务表现。
Evolving from Lessons: Skill-Augmented Table Graph Reasoning for Operation-wise Table Question Answering
The paper introduces Operation-wise TableQA, a new task with fine-grained question taxonomy, and proposes SkillTGR, a skill-augmented table graph reasoning framework that uses graph traversal and a hierarchical SkillBank for self-evolving reasoning, achieving superior performance and efficiency.
Prompt2Skill: Unsupervised Skill Optimization From Natural Language Instructions
Prompt2Skill is a framework that automatically builds and refines LLM skills from a natural-language task description alone, using dataset synthesis and reflective editing to outperform direct prompting by an average of 10.8% across four domains.