Paper deep dive

Hierarchical Evidence-Driven Reasoning for Long Document Understanding

Junyu Xiong, Yonghui Wang, Rongjian Gu, Chenyu Liu, Bing Yin, Wengang Zhou, Houqiang Li

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 61

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%

Last extracted: 7/7/2026, 2:56:41 PM

Summary

The paper introduces HIEVI-RAG, a hierarchical, evidence-driven multimodal Retrieval-Augmented Generation framework designed for long-document understanding. It addresses limitations of standard RAG pipelines, such as fetching answer-void distractors and suffering from cascading errors in single-pass retrieval. The framework decomposes complex queries into atomic sub-questions, performs coarse visual page retrieval, verifies evidence using a specialized multi-page verifier (EVIAGENT) trained with GRPO, and employs memory-guided iterative generation to dynamically reason over prioritized evidence. Extensive evaluations across four benchmarks demonstrate significant accuracy improvements over existing baselines.

Entities (12)

HIEVI-RAG · framework · 98%Coarse Visual Page Retrieval · module · 95%EVIAGENT · model · 95%FetaTab · dataset · 95%Hierarchical Question Decomposition · module · 95%LongDocURL · dataset · 95%Memory-Guided Iterative Generation · module · 95%MMLongBench · dataset · 95%PaperTab · dataset · 95%GRPO · algorithm · 92%Ops-ColQwen3-4B · retriever · 90%Qwen3-VL-8B-Instruct · model · 90%

Relation Signals (12)

Evidence-Aware Page Verification → employs → EVIAGENT

confidence 95% · fine-grained page verification via EVIAGENT, a specialized multi-page verifier trained with GRPO

HIEVI-RAG → evaluatedon → LongDocURL

confidence 95% · MMLongBench and LongDocURL, which feature extended multimodal documents

HIEVI-RAG → evaluatedon → PaperTab

confidence 95% · evaluate HIEVI-RAG on four benchmarks: PaperTab and FetaTab

HIEVI-RAG → evaluatedon → FetaTab

confidence 95% · evaluate HIEVI-RAG on four benchmarks: PaperTab and FetaTab

HIEVI-RAG → evaluatedon → MMLongBench

confidence 95% · MMLongBench and LongDocURL, which feature extended multimodal documents

HIEVI-RAG → uses → Hierarchical Question Decomposition

confidence 95% · HIEVI-RAG systematically factorizes complex queries into a cooperative four-stage pipeline: (1) hierarchical question decomposition

HIEVI-RAG → uses → Evidence-Aware Page Verification

confidence 95% · HIEVI-RAG systematically factorizes complex queries into a cooperative four-stage pipeline: (3) fine-grained page verification via EVIAGENT

HIEVI-RAG → uses → Memory-Guided Iterative Generation

confidence 95% · HIEVI-RAG systematically factorizes complex queries into a cooperative four-stage pipeline: (4) memory-guided iterative generation

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Retrieval-Augmented Generation (RAG) streamlines long-document understanding by leveraging retrieval mechanisms to restrict input images to a highly curated subset. However, existing multimodal RAG pipelines primarily face two critical challenges: first, standard semantic similarity retrievers frequently fetch topically overlapping yet answer-void distractor pages that mislead downstream generation; second, rigid single-pass pipelines heavily depend on initial retrieval success, where any omission of core evidence inevitably causes cascading errors. To address these challenges, we introduce HIEVI-RAG, a hierarchical, evidence-driven multimodal RAG framework for closed-domain document understanding. HIEVI-RAG systematically factorizes complex queries into a cooperative four-stage pipeline: (1) hierarchical question decomposition to break multi-hop root queries into atomic child questions; (2) coarse visual page retrieval leveraging a multimodal retriever to fetch candidate pages based on semantic similarity; (3) fine-grained page verification via EVIAGENT, a specialized multi-page verifier trained with GRPO to execute cross-page reasoning over multi-image blocks; and (4) memory-guided iterative generation that leverages accumulated sub-question context to execute multi-round, dynamic reasoning over the prioritized sequence. Extensive evaluations across four benchmarks demonstrate the robust efficacy and synergy of our framework, which significantly outperforms existing open-source baselines and exceeds the strongest reported baseline by an average of 8.05% in accuracy.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

59,223 characters extracted from source content.

Expand or collapse full text

HIEVI-RAG: Hierarchical Evidence-Driven Reasoning for Long Document Understanding Junyu Xiong 1 Yonghui Wang 1 Rongjian Gu 1 Chenyu Liu 2 Bing Yin 2 Wengang Zhou 1 * Houqiang Li 1 1 University of Science and Technology of China 2 iFLYTEK Research xiongjyu@mail.ustc.edu.cn, zhwg@ustc.edu.cn Abstract Retrieval-AugmentedGeneration(RAG) streamlines long-document understanding by leveraging retrieval mechanisms to restrict in- put images to a highly curated subset. However, existing multimodal RAG pipelines primarily face two critical challenges: first, standard semantic similarity retrievers frequently fetch topically overlapping yet answer-void distractor pages that mislead downstream generation; second, rigid single-pass pipelines heavily depend on initial retrieval success, where any omission of core evidence inevitably causes cascading errors. To address these challenges, we introduce HIEVI-RAG, a hierarchical,evidence-driven multimodal RAG framework for closed-domain document understanding. HIEVI-RAG systematically factorizes complex queries into a cooperative four-stage pipeline: (1) hierarchical question decomposition to break multi-hop root queries into atomic child questions; (2) coarse visual page retrieval leveraging a multimodal retriever to fetch candidate pages based on semantic similarity; (3) fine-grained page verification via EVIAGENT, a specialized multi-page verifier trained with GRPO to execute cross-page reasoning over multi-image blocks; and (4) memory-guided iterative generation that leverages accumulated sub-question context to execute multi-round, dynamic reasoning over the prioritized sequence. Extensive evaluations across four benchmarks demonstrate the robust efficacy and synergy of our framework, which significantly outperforms existing open-source baselines and exceeds the strongest reported baseline by an average of 8.05% in accuracy. * Corresponding author. 1 Introduction Closed-domain document understanding has be- come an indispensable core capability for practical AI systems deployed across various sectors, includ- ing science, finance, enterprise knowledge manage- ment, and technical support (Hui et al., 2024; Suri et al., 2025; Deng et al., 2025; Dong et al., 2024; Xia et al., 2024). In these scenarios, users typically pose queries over a proprietary document collec- tion, where the answers are sparsely distributed across hundreds of pages containing diverse and rich information such as plain text, tables, charts, and illustrations (Wang et al., 2025a; Chia et al., 2025). However, due to the prohibitive number of images in these closed-domain documents, feeding the entire collection into a model in an end-to-end manner easily exceeds the context window limits of existing LLMs. To address this bottleneck, a growing body of work on Retrieval-Augmented Generation (RAG) has emerged, which leverages retrieval mechanisms to restrict the number of input images to a highly curated subset prior to answer generation (Lewis et al., 2020; Cho et al., 2024; Suri et al., 2025; Xiong et al., 2026). Initially, to bypass tradi- tional OCR limitations and capture multimodal in- formation, ColPali (Faysse et al., 2025) extends late interaction to rendered page images, while M3DocRAG (Cho et al., 2024) establishes this method as the core interface for multi-page QA. Subsequently, research shifted toward filtering dis- tractors and optimizing ranking precision: MDocA- gent (Han et al., 2025) and SimpleDoc (Jain et al., 2025) refine results via module collaboration or iterative cascading; RagVL (Chen et al., 2025) and M-R5 (Xu et al., 2025) employ instruction- tuning or reinforcement learning to train strong arXiv:2607.04625v1 [cs.CV] 6 Jul 2026 VLM rerankers; and MoLoRAG (Wu et al., 2025) leverages page topology graphs for logic-aware traversal. Concurrently, DocAgent (Sun et al., 2025) and ALDEN (Yang et al., 2026) utilize mem- ory feedback or active exploration protocols to gather evidence dynamically, while DMAP (Fu et al., 2026) constructs human-aligned structural maps for global document comprehension. De- spite these advancements, two challenges remain: first, relying on semantic similarity misleads mod- els with topically related yet answer-void pages; second, single-pass pipelines rigidly depend on ini- tial retrieval, where any omission of core evidence makes correct generation impossible. To address these challenges, we propose HIEVI- RAG, a four-step framework for long-document understanding. First, hierarchical question de- composition breaks down the complex, multi-hop root question into multiple atomic sub-questions, reducing retrieval difficulty while assisting the root question’s resolution from different logical levels. Second, coarse-grained page retrieval utilizes a multimodal retriever (e.g., ops-ColQwen3) to fetch the top-kpages based on semantic similar- ity for each question, including both root and sub- questions. Third, for fine-grained evidence veri- fication, inspired by DocR1 (Xiong et al., 2025), we employ GRPO (Shao et al., 2024) to train EVI- AGENT, an evidence-aware multi-page reasoning model capable of cross-page reasoning and ground- ing over multi-image inputs to pinpoint question- relevant evidence. In this stage, for sub-questions, we strictly retain only the evidence pages identified by EVIAGENT; for the root question, EVIAGENT’s verified evidence pages are prioritized at the front, while non-evidence pages follow in order of their initial semantic similarity, thereby achieving pre- cise reranking. Fourth, during memory-guided iterative generation, an empty memory is initial- ized. The sub-questions and their corresponding evidence pages are fed into the model, with the sub-questions, generated reasoning traces, and in- termediate answers subsequently accumulated in the memory. For the root question, the model takes the root query, the accumulated memory, and a batch of images as input. If the current memory and image batch are insufficient to resolve the ques- tion, the system appends the latest reasoning trace to the memory and fetches the next batch of images for iterative reasoning. To verify the effectiveness of our approach, we conduct extensive experiments and analyses across four distinct datasets. The results demonstrate that our method significantly outperforms existing open- source baselines, exceeding the previous state-of- the-art by an average of 8.05%. Furthermore, we perform comprehensive ablation studies to validate the individual efficacy of each proposed module. Our contributions can be summarized as follows: •We propose HIEVI-RAG, a four-step multi- modal RAG framework for long-document QA that unifies hierarchical question decom- position, coarse-to-fine retrieval, and memory- guided iterative generation. •We introduce a fine-grained verification mech- anism via EVIAGENT, which performs cross- page reasoning over multi-image inputs to pre- cisely ground answer-supporting evidence. • We formulate a memory-guided iterative gen- eration process that utilizes sub-question rea- soning traces to assist the root query while dy- namically updating an explicit memory across multiple rounds to eliminate cascading re- trieval errors. 2 Related Work 2.1 Document Visual Question Answering Advancements in DocVQA have bifurcated into OCR-based and OCR-free paradigms. OCR-based methods convert pages into structured textual and spatial tokens to feed downstream reasoning frame- works (Tang et al., 2023; Wang et al., 2024a; Zhu et al., 2025), yet their performance remains inher- ently bounded by compounding parsing errors and the serialization loss of visual context. Conversely, OCR-free approaches directly train VLMs on raw document images, evolving from architectural scal- ing for resolution management (Ye et al., 2023a; Feng et al., 2023; Hu et al., 2024b) to evidence- aware reasoning via reinforcement learning (Xiong et al., 2025; Zheng et al., 2026). While these meth- ods substantially enhance single- or multi-page reader execution, effectively navigating and dis- tilling massive long-document repositories remains a largely orthogonal challenge. 2.2 Retrieval-Augmented Generation Multimodal RAG methodologies have transitioned from coarse-grained text-retrieval extensions to visual-centric, layout-aware document discovery. Early pipelines leverage visual retrievers and VLM- based rerankers to screen candidate pages inde- pendently based on superficial similarity scor- ing (Faysse et al., 2025; Cho et al., 2024; Chen et al., 2025; Xu et al., 2025). More recent efforts incorporate iterative multi-agent navigation, mem- ory feedback, or topological page graphs to cap- ture structural inter-page logic (Wu et al., 2025; Sun et al., 2025; Yang et al., 2026; Fu et al., 2026). Despite these scaling improvements, ex- isting pipelines still evaluate retrieved pages as iso- lated query-image pairs, leaving them highly sus- ceptible to topically overlapping but answer-void distractors. To bridge these gaps, HIEVI-RAG casts validation as a joint, cross-page verification problem and bypasses cascading failure modes via memory-guided iterative reasoning. Detailed re- lated works are deferred to Appendix A. 3 Method 3.1 Overview and Task Formulation Let a document be represented as an ordered se- quence of page imagesD =p 1 ,p 2 ,...,p N and qdenote a user question in a closed-domain set- ting. The final answer is assumed to be grounded inD, although the supporting evidence may be sparsely distributed across multiple non-contiguous pages. When the document lengthNscales to hun- dreds of pages, direct end-to-end VLM ingestion becomes computationally prohibitive due to con- text window constraints. To bypass this bottleneck, HIEVI-RAG decomposes long-document QA into four cooperative stages: (1) hierarchical question decomposition, (2) coarse visual page retrieval, (3) evidence-aware page verification, and (4) memory- guided iterative generation. A core design rationale of our framework is the multi-granularity opera- tion across stages: coarse retrieval is strictly recall- oriented, EVIAGENT conducts fine-grained verifi- cation over grouped page candidates via cross-page reasoning, and the final generator iteratively con- sumes the verified evidence sequence backed by an explicit memory. The overall pipeline is illustrated in Figure 1. 3.2 Hierarchical Question Decomposition Given a complex root questionq, this stage con- structs a shallow reasoning hierarchy, as illustrated in Figure 1(1). Specifically, a lightweight LLM ex- tracts key entities, structural attributes, constraints, and logical relations fromqto instantiate a set of atomic child questions: Q(q) =q 1 ,q 2 ,...,q m , m≤ 5.(1) Each child questionq i is formulated as a distinct natural query covering a non-redundant subset of the global information need, while strictly avoiding near-paraphrases of the full prompt. Crucially, partitioning the global query into lo- calized targets reduces the inherent difficulty of multi-hop reasoning. This upfront decomposition prevents retrieval omissions typical of standalone root queries, ensuring core evidence is captured before downstream generation. The correspond- ing prompt template and example are detailed in Appendix B.1 and E. 3.3 Coarse Visual Page Retrieval Followingdecomposition,weemploy Ops-ColQwen3-4B(OpenSearch-AI,2026), a ColPali-style late-interaction multimodal re- triever (Faysse et al., 2025; Khattab and Zaharia, 2020), to construct a candidate page pool from rendered document images.As illustrated in Figure 1(2), this retrieval stage is split into offline index construction and online query matching. Offline Index Construction. Each document page imagep i ∈ Dis encoded offline into a multi- vector representation to preserve fine-grained vi- sual features: v p i =E page (p i ), i = 1, 2,...,N.(2) The derived page embeddings are subsequently stored to build the global document index. Online Query Matching.Upon receiving a root questionq, the stage online encodes bothqand its derived atomic child questionsq i ∈Q(q)to obtain their respective text embeddings: v q =E query (q),v q i =E query (q i ).(3) These query embeddings are then matched against the offline page index using the late-interaction operator to compute similarity scores across all document pages. Based on these scores, we rank and retain the Top-K(K = 20) candidate pages for the root question and each child question, denoted as T q and T q i respectively. 1. Hierarchical Ques- tionDecomposition LLM 2. Coarse Visual Retrieval Encoder Vector DB ... $ ! $ " $ # $ & offline: index construction Encoder online: query matching ! ! " ! ! ! # ... match ... Top K (K=20) 3. Evidence-Aware Verification 1 ! 1 " 1 & 1 ' ... think evidence page ! / ...... Keep Discard ... EviAgent 1 ! 1 " 1 & 1 ' ... think evidence page Evidence First ... EviAgent ! Child question Root question ... ... Others Later 4.Memory-Guided Iterative Generation Final Answer ! 0 VLM answer Memory initialize ! Answer able yes fetch the next block. no u pdate memory Question: 2 ! ! ! ! " ! # ... "≤5 Initializing Memory Iterative Reasoning ! ! / Child question Root question Figure 1: Overview of HIEVI-RAG. The system first builds a shallow hierarchy from the original question to atomic child questions, retrieves Top-Kcandidate pages for both root and child questions with a visual retriever, verifies and reranks grouped-kcandidates with EVIAGENT, and performs memory-guided iterative reasoning over the verified evidence order. 3.4 Evidence-Aware Page Verification To filter out topically related but answer-void dis- tractors from the coarse retrieval pool, we introduce EVIAGENT, a specialized model trained to perform binary evidence page verification via cross-page reasoning. As illustrated in Figure 1(3), this stage takes the coarse candidate listsT q andT q i as input, partitions them into grouped-ksliding windows, and leverages the trained agent to output structured decisions. Training EVIAGENT with Evidence-Aware GRPO. Inspired by DocR1 (Xiong et al., 2025), we train EVIAGENT using an Evidence-Aware Group Relative Policy Optimization (EviGRPO) pipeline to specialize a VLM for multi-page docu- ment verification. The policy is optimized across six open-source multi-page datasets (Van Lan- deghem et al., 2023; Tito et al., 2022; Zhao et al., 2022; Schimanski et al., 2026; Yu et al., 2025a; Tanaka et al., 2023) with statistics detailed in Ap- pendix D. During the training phase, the VLM is con- strained to generate a structured content contain- ing three mandatory fields:o = (τ,y,a). The corresponding prompt template is detailed in Ap- pendix B.2. • Reasoning Trace (τ): Contained within the <think>tag,τencapsulates intra-page visual inspection and cross-page reasoning. • Evidence Decisions (y ∈ T,F k ): En- closed within the<evidence_page>tag,y enforces a strict one-to-one binary sequence matching the exact order of input images to prevent evaluation omissions. •Final Answer (a): Wrapped within the <answer>tag,adelivers the terminal text re- sponse derived from the verified evidence. The policy is optimized via a weighted joint re- ward function: R = λ fmt r fmt + λ evi r evi + λ acc r acc ,(4) wherer fmt enforces schema adherence, andr acc evaluates final answer correctness. The evidence rewardr evi computes theF 1 score against ground- truth page labels to accurately measure verification performance and prevent reward hacking. Fine-Grained Page Validation. During on- line inference, each retrieved candidate list is split intoL= ⌈K/k⌉blocks of sizek: b k t = p (t−1)k+1 ,...,p min(tk,K) , wheret ∈ 1,...,Lis the block index. Each block is veri- fied jointly by the trained EVIAGENT. To prevent information fragmentation across boundaries, all pages verified as true evidence are aggregated into a final consolidated block for a global cross-block synthesis. The refined signals are then routed to two distinct downstream policies: •Discarding for Child Questions: For each child questionq i , we extract only the verified evidence pages: E(q i ) =p∈ T q i :y p = T.(5) Non-evidence pages are discarded to restrict downstream generation strictly to localized, verified facts. •Reranking for the Root Question: For the root questionq, letE(q)denote the verified evidence page set. We construct a prioritized sequenceΠ(q)while preserving the original retriever ranking order within each partition: Π(q) = E(q)⊕ (T q \ E(q)),(6) where⊕denotes sequence concatenation. BothE(q)and the remaining subsetT q (q) strictly inherit the original ranking order de- termined by the coarse visual retriever. Non- evidence pages are retained as a backup suffix to secure recall for downstream iterative rea- soning. 3.5 Memory-Guided Iterative Generation This final stage leverages the resolved child ques- tions and the prioritized sequenceΠ(q)to gener- ate the terminal answer via a two-step, memory- enhanced pipeline. As illustrated in Figure 1(4), the process consists of initializing an explicit mem- ory layer with sub-question contexts and iteratively reasoning over the root question. Initializing Memory with Sub-Questions. Be- fore addressing the root query, the framework se- quentially resolves each child questionq i ∈Q(q) over its verified evidence blockE(q i )using EVIA- GENT: (τ i ,y i ,a i ) =A q i ,E(q i ) ,(7) whereAdenotes the generative policy of EVIA- GENT. Initializing the memory asM 0 = ∅, the state updates cumulatively by appending the com- plete triplets: M i = M i−1 ∪(q i ,τ i ,a i ).(8) This mechanism distills sub-question verification into informative textual facts, seeding the working memory with both mid-level execution traces and terminal local answers. Iterative Reasoning for the Root Question. Upon consolidating the sub-question memoryM m (wherem = |Q(q)|), HIEVI-RAG evaluates the root questionqby scanning the prioritized se- quenceΠ(q)via a sliding window of sizek. At ex- ecution roundt, EVIAGENT ingests the root query, the persistent memory state, and thet-th evidence block b k t ⊂ Π(q): (τ t ,y t ,a t ) =A q,b k t ,M m+t−1 .(9) The execution terminates immediately, returning a t as the global response ifa t ̸= α ∅ , where α ∅ denotes the designated abstention token (i.e., NOT_ANSWERABLE). Otherwise, the current reason- ing traceτ t is appended to the memory to guide the next window: M m+t = M m+t−1 ∪(q,τ t ).(10) This iterative accumulation continues until a valid answer is produced orΠ(q)is exhausted. By main- taining text-based memory across rounds, this re- current design forces the model to synthesize his- torical context globally, preventing critical informa- tion omission inherent in local window constraints. 4 Experiments Benchmark# QA SamplesMin. PagesMax. PagesAvg. Pages PaperTab393215210.7 FetaTab1016221616.3 MMLongBench1,082946847.8 LongDocURL2,3255114989.0 Table 1: Benchmark statistics for the four-dataset evalu- ation protocol. Average page counts are rounded to one decimal place. 4.1 Experimental Setup Benchmarks.We evaluate HIEVI-RAG on four benchmarks: PaperTab and FetaTab (Hui et al., MethodParam.PaperTabFetaTabMMLongBenchLongDocURL (Acc.)(Acc.)(Acc.)(Acc.) M3DocRAG (Cho et al., 2024)7B28.563.836.249.0 MDocAgent (Han et al., 2025)7B30.066.338.546.9 MoLoRAG+ (Wu et al., 2025)7B31.0 69.241.051.9 ALDEN (Yang et al., 2026)7B24.562.339.255.1 URaG (Shi et al., 2026)7B–33.852.2 Doc-V ⋆ (Zheng et al., 2026)7B–42.156.3 HIEVI-RAG8B44.072.948.265.7 Table 2: Main experimental results across the four benchmarks. The best results are highlighted in bold, and the second-bestresults are underlined. The symbol “–” indicates that the baseline result is not reported. ALDEN uses the ColQwen+ColBERT pipeline, and Doc-V ⋆ refers to the GRPO-optimized model. VariantPaperTabFetaTabMMLongBenchLongDocURLAvg. HIEVI-RAG44.072.948.265.757.7 w/o Evidence-Aware Verification39.7 (-4.3)69.6 (-3.3)44.5 (-3.7)59.1 (-6.6)53.2 (-4.5) w/o Hierarchical Decomposition41.3 (-2.7)70.5 (-2.4)45.8 (-2.4)62.4 (-3.3)55.0 (-2.7) w/o Iterative Reasoning41.9 (-2.1)72.0 (-0.9)46.1 (-2.1)63.2 (-2.5)55.8 (-1.9) Table 3: Ablation analysis of individual architectural modules. For w/o Evidence-Aware Verification, HIEVI-RAG’s Step 3 is bypassed; sub-questions in Step 4 initialize memory directly using the unverified Top-Kcoarse retrieval, and the root question iteratively reasons over the un-reranked sequence. For w/o Hierarchical Decomposition, HIEVI-RAG’s Step 1 and the memory initialization phase in Step 4 are omitted, while the iterative reasoning loop remains active. For w/o Iterative Reasoning, the sliding window in HIEVI-RAG’s Step 4 is deactivated, force-feeding the Top-5 candidate pages into the generator in a single-pass manner. 2024), which target question answering over scientific and Wikipedia-style tables; and M- LongBench (Wang et al., 2025a) and Long- DocURL (Deng et al., 2025), which feature ex- tended multimodal documents stressing cross-page retrieval, visual grounding, and multi-hop reason- ing. Structural statistics are provided in Table 1. Implementation Details. For hierarchical ques- tion decomposition, a lightweightQwen3.5-4B model is deployed by default to generate at most five child questions. The baseline coarse retriever is configured asops-ColQwen3-4B(OpenSearch- AI, 2026). Both root and child queries retrieve the Top-K(K = 20) candidate pages, which are subsequently verified using a grouped window size ofk = 7. We instantiate EVIAGENT using Qwen3-VL-8B-Instruct(Qwen Team, 2025) as the foundational backbone. For the joint reward function, the balancing scaling weights are explic- itly set as(λ fmt ,λ evi ,λ acc ) = (0.1, 0.5, 0.4), priori- tizing the evidence verification performance. Train- ing is executed on an8×NVIDIA A100 GPU clus- ter. Metrics.For MMLongBench and LongDocURL, we follow their original evaluation protocols, em- ploying standard accuracy and rule-based string matching tailored to various answer types. For Pa- perTab and FetaTab, we adopt an LLM-as-a-judge paradigm leveraging Qwen3.5-27B as the evalua- tor to compute binary accuracy (0 or 1) by assess- ing whether the generated response semantically matches the ground truth. Additionally, retrieval- related performance is quantified via Recall@K, NDCG@K, and MRR@K, with detailed mathe- matical formulations deferred to Appendix C. 4.2 Main Results Table 2 presents a performance comparison across the four benchmarks. HIEVI-RAG consistently outperforms all competitive baselines, validating the synergistic effects of its core architectural com- ponents. On PaperTab and FetaTab,HIEVI-RAG achieves 44.0% and 72.9% accuracy, yielding ab- solute improvements of +13.0% and +3.7% over the strongest baseline (MoLoRAG+), respectively. Tabular documents typically require exact cell alignment; standard retrievers frequently introduce MethodGeneratorRetrieverPaperTabFetaTabMMLongBenchLongDocURL (Acc.)(Acc.)(Acc.)(Acc.) MoLoRAG+ (Wu et al., 2025) Qwen2.5-VL-7B-Instruct ColQwen2.531.069.241.051.9 HIEVI-RAG Qwen2.5-VL-7B-Instruct ColQwen2.534.269.142.655.4 Qwen2.5-VL-7B-Instruct Ops-ColQwen3-4B38.371.445.060.2 Qwen3-VL-8B-Instruct ColQwen2.536.070.843.857.1 Qwen3-VL-8B-Instruct Ops-ColQwen3-4B44.072.948.265.7 Table 4: Sensitivity analysis over generation backbones and visual retrievers. MoLoRAG+ is included as the external baseline. The last row corresponds to the main HIEVI-RAG configuration in Table 2. The remaining HIEVI-RAG rows are controlled variants that change only the listed generator or retriever while keeping the same evidence-verification and iterative-generation pipeline. false positives due to repetitive schema structures. HIEVI-RAG mitigates this bottleneck by execut- ing fine-grained verification via EVIAGENT, effec- tively isolating answer-void distractors. Similarly,on MMLongBench and Long- DocURL, HIEVI-RAG secures 48.2% and 65.7% accuracy, outperforming Doc-V ⋆ by +6.1% and +9.4%, respectively. These tasks necessitate cross- page retrieval and multi-hop reasoning, where single-pass pipelines easily omit critical informa- tion. The consistent performance gains demon- strate the efficacy of our memory-guided iterative generation. 4.3 Ablation Studies Contribution of Individual Modules.To isolate the empirical impact of each architectural phase in HIEVI-RAG, we evaluate three ablation variants across all benchmarks (Table 3). Disabling any single stage yields a consistent performance degra- dation, validating the synergy of our core compo- nents. Specifically, w/o Evidence-Aware Verification drops the average score by -4.5% absolutely. With- out Step 3 verification, answer-void distractors oc- cupy the front of the sequence, misleading the gen- erator into premature early stopping and missing genuine evidence. Furthermore, initializing mem- ory over unverified Top-Kinputs introduces catas- trophic cascading noise. Meanwhile, w/o Hierar- chical Decomposition compromises the average score by -2.7%. This demonstrates that execut- ing complex queries as standalone prompts without HIEVI-RAG’s Step 1 induces severe retrieval omis- sions, whereas atomic decomposition guarantees the upfront recall of dispersed evidence. Lastly, w/o Iterative Reasoning consistently degrades re- sults by -1.9% on average. This variant deactivates the sliding window in HIEVI-RAG’s Step 4 and force-feeds the Top-5 pages in a single pass. The result proves that while verification serves as an effective filter, iterative reasoning acts as an indis- pensable safety net against information omission under local window constraints. Backbone and Retriever Sensitivity. To evalu- ate the robustness of HIEVI-RAG under varying underlying architectures, we perform a grid sensi- tivity analysis by cross-combining different gener- ative backbones and visual retrievers (Table 4). Specifically, when deploying the identical back- bone and retriever configuration as MoLoRAG+ (Qwen2.5-VL-7B-Instruct+ColQwen2.5), HIEVI-RAG outperforms this baseline in almost all benchmarks. It delivers clear absolute im- provements of +3.2% on PaperTab and +3.5% on LongDocURL, demonstrating that our primary advantages stem strictly from architectural innovations rather than the scaling of foundational models.Furthermore, scaling up either the retriever toOps-ColQwen3-4Bor the generator to Qwen3-VL-8B-Instructprovides independent performance increments across benchmarks. The combination of both upgraded components yields the best performance, proving that our framework effectively leverages stronger retrieval and generation capabilities to maximize final accuracy. Effectiveness of Fine-Grained Page Validation. To evaluate the impact of the verification stage on retrieval quality, we compare the retrieval metrics of the baseline against the reranked outputs gener- ated by our verifier both before and after training under various Top-Ksettings on MMLongBench and LongDocURL (Table 5). Specifically, deploying an untuned Qwen3-VL- 8B-Instruct yields marginal or even detrimental performance shifts. While it provides negligi- ble fluctuations on MMLongBench (e.g., a mod- est+0.2point increase in NDCG@1 at Top-1), it Top-KMethod MMLongBenchLongDocURL RecallNDCGMRRRecallNDCGMRR 1 Ops-ColQwen3-4B (baseline)54.270.470.453.573.473.4 + Qwen3-VL-8B-Instruct54.2 (+0.0)70.6 (+0.2)70.6 (+0.2)52.9 (−0.6)72.1 (−1.3)72.1 (−1.3) + EVIAGENT59.4 (+5.2)78.3 (+7.9)78.3 (+7.9)57.3 (+3.8)78.8 (+5.4)78.8 (+5.4) 3 Ops-ColQwen3-4B (baseline)75.273.877.274.873.781.1 + Qwen3-VL-8B-Instruct75.6 (+0.4)74.4 (+0.6)77.6 (+0.4)73.7 (−1.1)72.5 (−1.2)79.8 (−1.3) + EVIAGENT80.1 (+4.9)80.6 (+6.8)83.7 (+6.5)77.5 (+2.7)77.3 (+3.6)84.9 (+3.8) 5 Ops-ColQwen3-4B (baseline)81.476.278.281.076.482.0 + Qwen3-VL-8B-Instruct81.8 (+0.4)76.6 (+0.4)78.5 (+0.3)80.2 (−0.8)75.4 (−1.0)80.7 (−1.3) + EVIAGENT85.2 (+3.8)82.2 (+6.0)84.4 (+6.2)83.2 (+2.2)79.7 (+3.3)85.6 (+3.6) Table 5: Retrieval performance on MMLongBench and LongDocURL under different Top-Ksettings. Values in parentheses denote the absolute difference compared with the Ops-ColQwen3-4B baseline under the same Top-K setting. Figure 2: Effect of block sizekon MMLongBench retrieval quality. A moderate block size provides the best balance between cross-page comparison and visual-context overload. consistently degrades retrieval precision on Long- DocURL, with NDCG slipping by−1.3points at Top-1 and−1.0points at Top-5. This demonstrates that off-the-shelf VLMs lack the necessary align- ment to distinguish authentic evidence from com- plex multimodal distractors. In sharp contrast, our trained EVIAGENT consistently and substantially outperforms the baseline. It elevates the NDCG@1 from 70.4% to 78.3% on MMLongBench, and from 73.4% to 78.8% on LongDocURL. Furthermore, at Top-5, EVIAGENT pushes the final Recall to 85.2% and 83.2% on the respective benchmarks. These comprehensive gains across Recall, NDCG, and MRR solidly confirm EVIAGENT’s superior capac- ity in identifying authentic evidence and filtering out answer-void noise. Sensitivity to Block Size.To evaluate the impact of the block sizekdefined in HIEVI-RAG’s Step 3, we analyze the retrieval quality under varying scales across MMLongBench (Figure 2). The em- pirical results indicate that both excessively large and excessively small values ofkdegrade retrieval quality. Specifically, a moderate block size (e.g.,k = 7) achieves the optimal balance and yields the high- est retrieval metrics. Settingktoo small isolates candidate pages. This constrains the model’s ca- pacity for cross-page reasoning and comparison. Conversely, expandingkexcessively degrades per- formance. This drop is driven by visual context overload, where the excessive input scale dilutes model attention. Therefore, maintaining a tightly bounded block sizekis essential to maximize page validation accuracy. 5 Conclusion In this work, we presented HIEVI-RAG, a hierar- chical, evidence-driven multimodal RAG frame- work for closed-domain long-document under- standing, whose core innovation lies in its cooper- ative four-stage pipeline. Specifically, the frame- work first employs hierarchical question decom- position to partition complex multi-hop queries into atomic child questions, drastically reducing initial retrieval difficulty. Second, coarse visual retrieval leverages rendered page indexing to guar- antee upfront evidence recall. Third, fine-grained page verification utilizes the EVIAGENT to perform cross-page reasoning over multi-image blocks, pre- cisely suppressing topically related yet answer- void distractors. Fourth, memory-guided iterative generation accumulates sub-question traces into an explicit text-based memory, dynamically exe- cuting sliding-window reasoning to eliminate cas- cading information omissions. Extensive experi- ments across four benchmarks demonstrate that HIEVI-RAG significantly outperforms existing open-source baselines. Furthermore, comprehen- sive ablation studies analyses solidly validate the individual efficacy and systemic synergy of each cooperative phase within our pipeline. Limitations Despite the state-of-the-art performance achieved by HIEVI-RAG, several inherent limitations war- rant further investigation. First, our framework introduces non-trivial com- putational overhead during the inference stage. Specifically, the hierarchical question decompo- sition, the multi-block page verification via the EVIAGENT, and the subsequent memory-guided generation inherently mandate multiple sequential LLM/VLM invocations. This compound multi- stage architecture inevitably increases cumulative response latency and inference costs compared to standard single-pass pipelines. Second, the memory-guided generation mecha- nism relies heavily on the quality of textual interme- diate reasoning traces; if the initial sub-questions yield severely hallucinated or biased summaries, such errors may cascade into the persistent mem- ory layer and occasionally misguide the terminal root-question synthesis. Third, our framework was optimized and evalu- ated predominantly on born-digital electronic doc- uments with pristine rendering. Its zero-shot gener- alization capabilities across real-world handwritten inputs, heavily degraded physical scans with text artifacts, or scene-text images captured in natural environments require broader empirical validation in future work. Ethical Considerations HIEVI-RAG is designed for closed-domain long- document understanding and may be applied to documents containing sensitive or proprietary in- formation. All experiments in this paper are con- ducted on publicly available benchmarks and open- source datasets. We do not collect new user data or perform human-subject annotation. Neverthe- less, deployment in real-world enterprise or finan- cial scenarios should ensure proper access control, data privacy protection, and auditing of generated answers. Because the system may still produce incorrect or unsupported answers when retrieval or evidence verification fails, we recommend using the framework as an assistive tool rather than an au- tonomous decision-making system in high-stakes domains. References Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R. Manmatha. 2024. Doc- Formerv2: Local features for document understand- ing. In Proceedings of the AAAI Conference on Arti- ficial Intelligence. Zhanpeng Chen, Chengjin Xu, Yiyan Qi, Xuhui Jiang, and Jian Guo. 2025. VLM is a strong reranker: Advancing multimodal retrieval-augmented genera- tion via knowledge-enhanced reranking and noise- injected training. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 8140–8158, Suzhou, China. Association for Compu- tational Linguistics. Yew Ken Chia, Liying Cheng, Hou Pong Chan, Mao- jia Song, Chaoqun Liu, Mahani Aljunied, Soujanya Poria, and Lidong Bing. 2025. M-LongDoc: A benchmark for multimodal super-long document un- derstanding and a retrieval-aware tuning framework. In Proceedings of the 2025 Conference on Empiri- cal Methods in Natural Language Processing, pages 9233–9250, Suzhou, China. Association for Compu- tational Linguistics. Jaemin Cho, Debanjan Mahata, Ozan Irsoy, Yujie He, and Mohit Bansal. 2024. M3DocRAG: Multi- modal retrieval is what you need for multi-page multi-document understanding.arXiv preprint arXiv:2411.04952. Chao Deng, Jiale Yuan, Pi Bu, Peijie Wang, Zhong- Zhi Li, Jian Xu, Xiao-Hui Li, Yuan Gao, Jun Song, Bo Zheng, and Cheng-Lin Liu. 2025. LongDocURL: a comprehensive multimodal long document bench- mark integrating understanding, reasoning, and locat- ing. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pages 1135–1159, Vienna, Austria. Association for Computational Linguistics. Haoyu Dong, Jianbo Zhao, Yuzhang Tian, Junyu Xiong, Mengyu Zhou, Yun Lin, José Cambronero, Yeye He, Shi Han, and Dongmei Zhang. 2024. Encoding spreadsheets for large language models. In Proceed- ings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 20728–20748. Manuel Faysse, Hugues Sibille, Tony Wu, Bilel Omrani, Gautier Viaud, Céline Hudelot, and Pierre Colombo. 2025. ColPali: Efficient document retrieval with vision language models. In The Thirteenth Interna- tional Conference on Learning Representations. Hao Feng, Qi Liu, Hao Liu, Jingqun Tang, Wengang Zhou, Houqiang Li, and Can Huang. 2023. DocPe- dia: Unleashing the power of large multimodal model in the frequency domain for versatile document un- derstanding. arXiv preprint arXiv:2311.11810. ShunLiang Fu, Yanxin Zhang, Yixin Xiang, Xiaoyu Du, and Jinhui Tang. 2026. DMAP: Human-aligned structural document map for multimodal document understanding. In Proceedings of the ACM Web Con- ference 2026. Siwei Han, Peng Xia, Ruiyi Zhang, Tong Sun, Yun Li, Hongtu Zhu, and Huaxiu Yao. 2025. MDocAgent: A multi-modal multi-agent framework for document understanding. arXiv preprint arXiv:2503.13964. Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, and Fei Huang. 2024a. mPLUG-DocOwl 1.5: Unified struc- ture learning for OCR-free document understanding. arXiv preprint arXiv:2403.12895. Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, and Fei Huang. 2024b. mPLUG-DocOwl2: High-resolution compressing for OCR-free multi-page document un- derstanding. arXiv preprint arXiv:2409.03420. Yulong Hui, Yao Lu, and Huanchen Zhang. 2024. UDA: A benchmark suite for retrieval augmented generation in real-world document analysis. In Advances in Neural Information Processing Systems 37: Datasets and Benchmarks Track. Chelsi Jain, Yiran Wu, Yifan Zeng, Jiale Liu, Shengyu Dai, Zhenwen Shao, Qingyun Wu, and Huazheng Wang. 2025. SimpleDoc: Multi-modal document un- derstanding with dual-cue page retrieval and iterative refinement. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Process- ing, pages 28410–28427, Suzhou, China. Association for Computational Linguistics. Omar Khattab and Matei Zaharia. 2020. ColBERT: Effi- cient and effective passage search via contextualized late interaction over BERT. In Proceedings of the 43rd International ACM SIGIR Conference on Re- search and Development in Information Retrieval, pages 39–48. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge- intensive NLP tasks. In Advances in Neural Informa- tion Processing Systems 33. Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xu- Cheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. 2024. TextMonkey: An OCR-free large mul- timodal model for understanding document. arXiv preprint arXiv:2403.04473. Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. 2024. LayoutLLM: Lay- out instruction tuning with large language models for document understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition. Tengchao Lv, Yupan Huang, Jingye Chen, Yuzhong Zhao, Yilin Jia, Lei Cui, Shuming Ma, Yaoyao Chang, Shaohan Huang, Wenhui Wang, Li Dong, Weiyao Luo, Shaoxiang Wu, Guoxin Wang, Cha Zhang, and Furu Wei. 2023. KOSMOS-2.5: A multimodal liter- ate model. arXiv preprint arXiv:2309.11419. Ahmad Mohammadshirazi, Pinaki Prasad Guha Neogi, Ser-Nam Lim, and Rajiv Ramnath. 2024. DLaVA: Document language and vision assistant for answer localization with enhanced interpretability and trust- worthiness. arXiv preprint arXiv:2412.00151. Mor Shpigel Nacson, Aviad Aberdam, Roy Ganz, Elad Ben Avraham, Alona Golts, Yair Kittenplon, Shai Mazor, and Ron Litman. 2025. DocVLM: Make your VLM an efficient reader. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 29005–29015. OpenSearch-AI. 2026.Ops-Colqwen3: State-of- the-Art Multimodal Embedding Model for Visual Document Retrieval.https://huggingface.co/ OpenSearch-AI/Ops-Colqwen3-4B. Qwen Team. 2025. Qwen3-VL technical report. arXiv preprint arXiv:2511.21631. Tobias Schimanski, Imene Kolli, Yu Fan, Ario Saeid Vaghefi, Jingwei Ni, Elliott Ash, and Markus Leip- pold. 2026. pdfQA: Diverse, challenging, and real- istic question answering over PDFs. arXiv preprint arXiv:2601.02285. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300. Yaya Shi, Jingyun Zhang, Yuliang Liu, Hongliang Li, and Xiang Bai. 2026. URaG: Unified retrieval and generation for multimodal document understanding. arXiv preprint arXiv:2603.12189. Li Sun, Liu He, Shuyue Jia, Yangfan He, and Chenyu You. 2025. DocAgent: An agentic framework for multi-modal long-context document understanding. In Proceedings of the 2025 Conference on Empiri- cal Methods in Natural Language Processing, pages 17701–17716, Suzhou, China. Association for Com- putational Linguistics. Manan Suri, Puneet Mathur, Franck Dernoncourt, Kanika Goswami, Ryan A. Rossi, and Dinesh Manocha. 2025. VisDoM: Multi-document QA with visually rich elements using multimodal retrieval- augmented generation. In Proceedings of the 2025 Conference of the Nations of the Americas Chap- ter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Pa- pers), pages 6088–6109, Albuquerque, New Mexico. Association for Computational Linguistics. Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. 2023. SlideVQA: A dataset for document visual ques- tion answering on multiple images. arXiv preprint arXiv:2301.04883. Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. 2023. UDOP: Unified document processing with vision, text and layout. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 24423– 24433. Rubèn Tito, Dimosthenis Karatzas, and Ernest Val- veny. 2022.Hierarchical multimodal transform- ers for multi-page DocVQA.arXiv preprint arXiv:2212.05935. Jordy Van Landeghem, Rubén Tito, Łukasz Borchmann, Michał Pietruszka, Paweł Józiak, Rafał Powalski, Dawid Jurkiewicz, Mickaël Coustaty, Bertrand Ack- aert, Ernest Valveny, Matthew Blaschko, Sien Moens, and Tomasz Stanisławek. 2023. Document under- standing dataset and evaluation (DUDE). arXiv preprint arXiv:2305.08455. Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. 2024a. DocLLM: A layout-aware generative language model for multimodal document understanding.arXiv preprint arXiv:2401.00908. Yonghui Wang, Shi-Yong Chen, Zhenxing Zhou, Siyi Li, Haoran Li, Wengang Zhou, and Houqiang Li. 2024b. Root: Vlm based system for indoor scene understand- ing and beyond. arXiv preprint arXiv:2411.15714. Yonghui Wang, Shaokai Liu, Li Li, Wengang Zhou, and Houqiang Li. 2024c. Swinshadow: Shifted win- dow for ambiguous adjacent shadow detection. ACM Transactions on Multimedia Computing, Communi- cations and Applications, 20(11):1–20. Yonghui Wang, Wengang Zhou, Hao Feng, Keyi Zhou, and Houqiang Li. 2023a.Towards im- proving document understanding: An exploration on text-grounding via mllms.arXiv preprint arXiv:2311.13194. Yonghui Wang, Wengang Zhou, Zhenbo Lu, and Houqiang Li. 2022. Udoc-gan: Unpaired document illumination correction with background light prior. In Proceedings of the 30th ACM International Con- ference on Multimedia, pages 5074–5082. Yonghui Wang, Wengang Zhou, Yunyao Mao, and Houqiang Li. 2023b. Detect any shadow: Segment anything for video shadow detection. IEEE Transac- tions on Circuits and Systems for Video Technology, 34(5):3782–3794. Zhaowei Wang, Wenhao Yu, Xiyu Ren, Jipeng Zhang, Yu Zhao, Rohit Saxena, Liang Cheng, Ginny Wong, Simon See, Pasquale Minervini, Yangqiu Song, and Mark Steedman. 2025a.MMLongBench: Benchmarking long-context vision-language mod- els effectively and thoroughly.arXiv preprint arXiv:2505.10610. Zining Wang, Tongkun Guan, Pei Fu, Chen Duan, Qianyi Jiang, Zhentao Guo, Shan Guo, Junfeng Luo, Wei Shen, and Xiaokang Yang. 2025b. Marten: Vi- sual question answering with mask generation for multi-modal document understanding. In Proceed- ings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition. Haoran Wei, Chenglong Liu, Jinyue Chen, Jia Wang, Lingyu Kong, Yanming Xu, Zheng Ge, Liang Zhao, Jianjian Sun, Yuang Peng, Chunrui Han, and Xiangyu Zhang. 2024. General OCR theory: Towards OCR- 2.0 via a unified end-to-end model. arXiv preprint arXiv:2409.01704. Xixi Wu, Yanchao Tan, Nan Hou, Ruiyang Zhang, and Hong Cheng. 2025. MoLoRAG: Bootstrapping doc- ument understanding via multi-modal logic-aware retrieval. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 14024–14045, Suzhou, China. Association for Computational Linguistics. Shiyu Xia, Junyu Xiong, Haoyu Dong, Jianbo Zhao, Yuzhang Tian, Mengyu Zhou, Yeye He, Shi Han, and Dongmei Zhang. 2024. Vision language models for spreadsheet understanding: Challenges and op- portunities. In Proceedings of the 3rd Workshop on Advances in Language and Vision Research (ALVR), pages 116–128. Junyu Xiong, Yuan Pu, Jia Tang, and Yazhe Niu. 2026.Priorzero: Bridging language priors and world models for decision making. arXiv preprint arXiv:2605.12289. Junyu Xiong, Yonghui Wang, Weichao Zhao, Chenyu Liu, Bing Yin, Wengang Zhou, and Houqiang Li. 2025. DocR1: Evidence page-guided GRPO for multi-page document understanding. arXiv preprint arXiv:2508.07313. Mingjun Xu, Jinhan Dong, Jue Hou, Zehui Wang, Si- hang Li, Zhifeng Gao, Renxin Zhong, and Hengxing Cai. 2025. M-R5: Multimodal reasoning-enhanced reranker via reinforcement learning for document re- trieval. arXiv preprint arXiv:2506.12364. Tianyu Yang, Terry Ruas, Yijun Tian, Jan Philip Wahle, Daniel Kurzawe, and Bela Gipp. 2026. ALDEN: Reinforcement learning for active navigation and evi- dence gathering in long documents. arXiv preprint arXiv:2510.25668. Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Dan, Chenliang Zhao, Guohai Xu, Chen Li, Junfeng Tian, Qi Qian, Ji Zhang, and Fei Huang. 2023a.UReader: Universal OCR-free visually- situated language understanding with multimodal large language model. In Findings of the Associ- ation for Computational Linguistics: EMNLP 2023, pages 2841–2858. Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chen Li, Junfeng Tian, Qi Qian, Ji Zhang, and Fei Huang. 2023b. mPLUG- DocOwl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499. Wenhan Yu, Zhaoxi Zhang, Wang Chen, Guanqiang Qi, Weikang Li, Lei Sha, Deguo Xia, and Jizhou Huang. 2025a. SciEGQA: A dataset for scientific evidence- grounded question answering and reasoning. arXiv preprint arXiv:2511.15090. Xinlei Yu, Chengming Xu, Zhangquan Chen, Yudong Zhang, Shilin Lu, Cheng Yang, Jiangning Zhang, Shuicheng Yan, and Xiaobin Hu. 2025b. Visual doc- ument understanding and reasoning: A multi-agent collaboration framework with agent-wise adaptive test-time scaling. arXiv preprint arXiv:2508.03404. Yilun Zhao, Lyuhao Chen, Arman Cohan, and Chen Zhao. 2024. TaPERA: Enhancing faithfulness and in- terpretability in long-form table QA by content plan- ning and execution-based reasoning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12824–12840, Bangkok, Thailand. Association for Computational Linguistics. Yilun Zhao, Yunxiang Li, Chenying Li, and Rui Zhang. 2022. MultiHiertt: Numerical reasoning over multi hierarchical tabular and textual data. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6588–6600, Dublin, Ireland. Association for Computational Linguistics. Yuanlei Zheng, Pei Fu, Hang Li, Ziyang Wang, Yuyi Zhang, Wenyu Ruan, Xiaojin Zhang, Zhongyu Wei, Zhenbo Luo, Jian Luan, Wei Chen, and Xiang Bai. 2026. Doc-V*: Coarse-to-fine interactive visual reasoning for multi-page document VQA. arXiv preprint arXiv:2604.13731. Zhaoqing Zhu, Chuwei Luo, Zirui Shao, Feiyu Gao, Hangdi Xing, Qi Zheng, and Ji Zhang. 2025. A simple yet effective layout token in large language models for document understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. A Detailed Related Work A.1 Document Visual Question Answering Recent advancements in DocVQA primar- ily diverge into OCR-based and OCR-free paradigms (Wang et al., 2022, 2023b, 2024a,b,c, 2023a). OCR-based methods serialize document pages into structured textual representations to serve as inputs for downstream reasoning. Representative architectures leverage multimodal pretraining or specialized attention mechanisms to fuse textual, visual, and layout features, as exemplified by UDOP (Tang et al., 2023), DocFormerv2 (Appalaraju et al., 2024), and Do- cLLM (Wang et al., 2024a). To enhance efficiency and structured generation, LayoutLLM (Luo et al., 2024) frames layout-sensitive reasoning as an instruction-tuning problem, whereas Lay- TokenLLM (Zhu et al., 2025) employs compact layout tokens to represent spatial bounding boxes without expanding lengthy coordinate sequences. For structured data, TaPERA (Zhao et al., 2024) incorporates content planning and execution-based reasoning to improve faithfulness in long-form table QA. Although highly effective when struc- tural parsing is reliable, the performance of these methods remains intrinsically upper-bounded by cascading parsing errors and the inevitable loss of visually-situated evidence during text serialization. OCR-free methods bypass external parsing pipelines by training VLMs to comprehend docu- ment images end-to-end. Early foundational archi- tectures prioritized enhancing reading capabilities under constrained token budgets; within this cohort, UReader (Ye et al., 2023a) introduces auxiliary reading and key-point generation tasks, KOSMOS- 2.5 (Lv et al., 2023) pretrains on text-intensive im- ages for structured text generation, DocPedia (Feng et al., 2023) processes visual inputs in the fre- quency domain, TextMonkey (Liu et al., 2024) incorporates high-resolution modeling with token filtering, and the mPLUG-DocOwl series (Ye et al., 2023b; Hu et al., 2024a,b) employs feature com- pression to scale from single-page to multi-page document understanding. More recently, frame- works have evolved toward specialized reasoning and optimization protocols. Specifically, GOT (Wei et al., 2024) frames vision-based reading as a uni- fied OCR-free generator, whereas DLaVA (Moham- madshirazi et al., 2024) and Marten (Wang et al., 2025b) implement visual answer localization and mask prediction to enhance model interpretabil- ity. To boost efficiency, DocVLM (Nacson et al., 2025) uses compact textual queries to lower visual token costs, while MACT (Yu et al., 2025b) orches- trates collaborative multi-agent reasoning at test time. To shift from passive reading to active pol- icy optimization, DocR1 (Xiong et al., 2025) and Doc-V ⋆ (Zheng et al., 2026) leverage evidence- guided GRPO to explicitly train visual reasoning trajectories for multi-page environments. A.2 Retrieval-Augmented Generation RAG methodologies have evolved from traditional text passage retrieval to multi-modal and structure- aware document discovery. Following the text- centric foundations laid by passage retrieval (Lewis et al., 2020) and token-level late interaction (Khat- tab and Zaharia, 2020), visual RAG systems have increasingly focused on page-level document in- terfaces. ColPali (Faysse et al., 2025) extends late interaction to rendered page images, while M3DocRAG (Cho et al., 2024) establishes this ap- proach as a core interface for multi-page QA. To decouple retrieval from reading, MDocAgent (Han et al., 2025) and SimpleDoc (Jain et al., 2025) uti- lize multi-module collaboration and iterative cas- cading, while RagVL (Chen et al., 2025) and M- R5 (Xu et al., 2025) employ instruction-tuning or reinforcement learning to train specialized vi- sual rerankers. Concurrently, efforts have been dedicated to structured or active navigation to handle complex layouts: MoLoRAG (Wu et al., 2025) leverages page topology graphs for logic- aware traversal, DocAgent (Sun et al., 2025) and ALDEN (Yang et al., 2026) introduce memory feedback and active exploration protocols, and DMAP (Fu et al., 2026) constructs human-aligned structural maps to guide global understanding. B Prompt Templates B.1 Prompt For Hierarchical Question Decomposition The system prompt for hierarchical question de- composition is detailed in Figure 3. Given the root question as input, the model is strictly constrained to output a maximum of five atomic child questions, each formatted strictly as an interrogative sentence. B.2 Prompt For EviAgent The system prompt used for both training and infer- ence of EVIAGENT is provided in Figure 4. Given a question and a set of input pages, EVIAGENT is required to produce structured reasoning, make a binary evidence decision for each page, and gener- ate the final answer. C Retrieval Metrics LetG q denote the gold evidence-page set for ques- tionq, and letR K q = (r 1 ,...,r K )denote the ranked top-Kretrieved pages. We report three standard retrieval metrics. Recall@K. Recall@K measures the fraction of gold evidence pages recovered in the top-Kresults: Recall@K(q) = |G q ∩ R K q | |G q | .(11) Higher Recall@K indicates better evidence cover- age. NDCG@K.Normalized Discounted Cumulative Gain emphasizes retrieving gold pages early in the ranked list. With binary relevance labelsrel i = I(r i ∈ G q ), we compute DCG@K(q) = K X i=1 2 rel i − 1 log 2 (i + 1) ,(12) and normalize it by the ideal ranking: NDCG@K(q) = DCG@K(q) IDCG@K(q) .(13) This metric rewards both correctness and ranking quality. MRR@K. Mean Reciprocal Rank focuses on how early the first relevant page appears: R@K(q) = ( 1 mini|r i ∈G q , if G q ∩ R K q ̸=∅, 0,otherwise. (14) MRR@K is the average reciprocal rank over all questions. D Training Datasets Table 6 details the statistics of the multi-page train- ing data utilized for EVIAGENT. To ensure a bal- anced and comprehensive optimization, we uni- formly sample 1,000 instances from each of the six constitutive datasets, with all selected samples inherently featuring multi-page visual contexts: • DUDE (Van Landeghem et al., 2023) and MP-DocVQA (Tito et al., 2022) supply rich structural information from real-world com- plex electronic layouts and scanned industry documents. SystemPrompt: You are a retrieval query decomposition assistant. Your task is to decompose a user question into multiple concise search queries that help retrieve documents containing the key information needed to answer the question. Guidelines: 1. First identify all key elements in the question, including entities (people, organizations, locations), topics or concepts, numbers or statistics, dates or time constraints, and conditions or qualifiers. 2. After identifying these key elements, construct retrieval queries by using one or two of the identified elements to form concise and searchable queries. 3. Each query should focus on retrieving documents that mention those specific elements or their relationships. 4. Preserve all quantitative constraints, temporal conditions, and logical relationships from the original question. 5. Queries should be concise and phrased in a way that could realistically appear in documents. 6. Avoid redundant queries that express the same concept. 7. Keep the total number of generated queries within MAX_QUERIES. 8. Return JSON only in the exact format: "queries": ["...", "..."] Do not include explanations, comments, or markdown. Example: Input: "What are the views of 20% of the people in California regarding autonomous vehicles in 2023?" Output: "queries": ["California residents views on autonomous vehicles", "Where 20% of California residents are mentioned?", "Public opinion on autonomous driving in 2023"] Figure 3: System prompt for hierarchical question decomposition. Input PageEvidence Page Dataset# SamplesMinMaxAvgMinMaxAvg DUDE10002206.93181.16 MP-DocVQA10002206.94111.00 MultiHiertt1000374.06131.48 pdfQA10004109.951101.42 SciEGQA100022011.42121.50 SlideVQA1000202020.00131.40 Table 6: Training statistics for EVIAGENT. •MultiHiertt (Zhao et al., 2022) and pdfQA (Schimanski et al., 2026) inject demanding financial reports and scientific literature requiring cross-page reasoning over hierarchical tables and textual narratives. •SciEGQA (Yu et al., 2025a) provides dense academic papers with tightly coupled illustra- tions, charts, and mathematical proofs. • SlideVQA (Tanaka et al., 2023) introduces sequential presentation slides characterized by sparse, highly stylized cross-page visual elements. E Case Study Figure 5 illustrates a concrete execution trace of our decomposition layer. The complex, multi-hop root query requires concurrent visual counting and cross-year mathematical synthesis. Directly retriev- ing pages using this intricate query typically mis- leads standard semantic retrievers. By partitioning the root query into three localized, atomic child questions, our framework effectively simplifies the retrieval objectives, ensuring all disjoint tabular and visual evidence pages are fully recalled with- out cascading omissions. SystemPrompt: Youareaspecializedmultimodalmodelforfine-grainedmulti-imagevisualunderstandingandevidenceimage localization.Youareanexpertatcarefullyinspectingmultipleimages,understandingdetailedvisualandtextual informationwithineachimage,andidentifyingwhichimagescontainevidencerelevanttoagivenquestion. Youwillbegivensomeimagesalongwithaquestion.Carefullyexamineeveryimage. Foreachimage,perform detailedvisualanalysis,includingtext,layout,objects,diagrams,tables,andanyotherpotentiallyinformative elements.Clearlydescribewhatinformationappearsineachimageandwhetheritmaycontributetoansweringthe question. Afteranalyzingallimages,performcross-imagereasoningtodeterminewhichimagescontainrelevantevidence relatedtothequestion.Someimagesmaynotdirectlyanswerthequestionbutmaystillprovidenecessary supportinginformation.However,oncetheexistingevidenceissufficienttoanswerthequestiondirectlyand completely,donotselectanyadditionalimages. Afteridentifyingtheevidenceimages,youmustuseonlytheinformationfromtheseevidenceimagestofurther reasonandderivethefinalanswertothequestion.Donotuseinformationfromnon-evidenceimages.Iftheanswer canbedetermined,outputaconciseanswerusingashortphraseorafewwordsonly.Iftheanswercannotbe determinedbasedontheevidenceimages,output"NotAnswerable". Youroutputmustcontainexactlythreeparts. First,provideyourreasoningprocesswrappedin<think></think>. Inside<think>,yourreasoningMUSTbestructuredintothreeclearlyseparatedparts: (1)Image-levelanalysis:analyzeeachimageonebyone. (2)Cross-imagereasoning:explainwhichimagesarerelevantandwhy. (3)Answerreasoning:basedonlyontheselectedevidenceimages,derivethefinalanswerstepbystep. Second,providetheevidence-imagejudgmentwrappedin<evidence_page></evidence_page>asacomma- separatedsequenceofTorFinimageorder.Thenumberofentriesmustexactlymatchthenumberofinputimages. Third,providethefinalanswerwrappedin<answer></answer>followingtherulesabove. Donotoutputanythingoutsidethesethreesections. Figure 4: System prompt used for EVIAGENT training and inference. Question:What is the sum of the total number of paid search's conversions in the year of 2007, 2008 and the number of green bars in the heroes happen here launch? Sub-question 1: What is the number of paid search conversions in 2007? Sub-question 2: What is the number of paid search conversions in 2008? Sub-question 3: How many green bars are shown in the “Heroes Happen Here” launch chart? LLM 9 Figure 5: Case study of hierarchical question decomposition in HIEVI-RAG.

Instant research discovery