Upload folder using huggingface_hub

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/appendix_chunks.jsonl

@@ -0,0 +1,117 @@
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0083", "section": "A. Algorithm Workflow", "page_start": 12, "page_end": 12, "type": "Text", "text": "We formalize the complete workflow of our proposed ColParse framework in two distinct algorithms. Algorithm 1 details the offline indexing process, where ColParse generates a highly compact set of document embeddings through its sequential three-stage process. Subsequently, Algorithm 2 illustrates the online retrieval phase, where the final relevance score is efficiently computed via a MaxSim operation using this compressed set of embeddings.", "source": "marker_v2", "marker_block_id": "/page/11/Text/2"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0084", "section": "Algorithm 1 The Offline Indexing Process of ColParse", "page_start": 12, "page_end": 12, "type": "Text", "text": "Input : A document image d \\in \\mathbb{R}^{H \\times W \\times 3} ;", "source": "marker_v2", "marker_block_id": "/page/11/Text/51"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0085", "section": "Algorithm 1 The Offline Indexing Process of ColParse", "page_start": 12, "page_end": 12, "type": "Code", "text": "A document parser model \\Psi_{\\text{parse}}; A \\text{ single-vector encoder } \\Phi_{\\text{enc}} : \\mathbb{R}^{H' \\times W' \\times 3} \\to \\mathbb{R}^D Output: A compact multi-vector representation \\mathbf{D}_{\\text{ColParse}} \\subset \\mathbb{R}^{k \\times D} /* Stage 1: Layout-Informed Document Parsing */ [\\{b_j,c_j\\}]_{j=1}^k \\leftarrow \\Psi_{\\text{parse}}(d) \\; ; \\; \\text{ // Get $k$ bboxes and content types} \\mathcal{S}_d \\leftarrow \\emptyset \\; \\text{ for } j \\leftarrow 1 \\; \\text{ to $k$ do}", "source": "marker_v2", "marker_block_id": "/page/11/Code/4"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0086", "section": "Algorithm 1 The Offline Indexing Process of ColParse", "page_start": 12, "page_end": 12, "type": "Text", "text": "\\begin{vmatrix} s_j \\leftarrow \\operatorname{Crop}(d, b_j) ; & \\text{Crop doc image } d \\text{ using bbox } b_j \\\\ S_d \\leftarrow S_d \\cup \\{s_j\\} \\end{vmatrix}", "source": "marker_v2", "marker_block_id": "/page/11/Text/5"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0087", "section": "Algorithm 1 The Offline Indexing Process of ColParse", "page_start": 12, "page_end": 12, "type": "Text", "text": "end", "source": "marker_v2", "marker_block_id": "/page/11/Text/6"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0088", "section": "Algorithm 1 The Offline Indexing Process of ColParse", "page_start": 12, "page_end": 12, "type": "Code", "text": "/* Stage 2: Dual-Stream Encoding */ \\mathbf{D}_{local} \\leftarrow \\emptyset \\text{ for } each \\ sub-image \\ s_j \\in \\mathcal{S}_d \\ \\mathbf{do} \\mid \\mathbf{v}_{local}^{(j)} \\leftarrow \\Phi_{enc}(s_j) \\ ; \\qquad \\text{// Encode local region} \\mid \\mathbf{D}_{local} \\leftarrow \\mathbf{D}_{local} \\cup \\{\\mathbf{v}_{local}^{(j)}\\}\\nend", "source": "marker_v2", "marker_block_id": "/page/11/Code/7"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0089", "section": "Algorithm 1 The Offline Indexing Process of ColParse", "page_start": 12, "page_end": 12, "type": "Text", "text": "\\mathbf{v}_{\\mathrm{global}} \\leftarrow \\Phi_{\\mathrm{enc}}(d) ; // Encode entire page for global context", "source": "marker_v2", "marker_block_id": "/page/11/Text/8"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0090", "section": "Algorithm 1 The Offline Indexing Process of ColParse", "page_start": 12, "page_end": 12, "type": "Text", "text": "/* Stage 3: Global-Local Fusion */", "source": "marker_v2", "marker_block_id": "/page/11/Text/9"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0091", "section": "Algorithm 1 The Offline Indexing Process of ColParse", "page_start": 12, "page_end": 12, "type": "Text", "text": "\\begin{aligned} \\mathbf{D}_{\\texttt{ColParse}} &\\leftarrow \\emptyset \\ \\ \\textbf{for} \\ \\textit{each local vector} \\ \\mathbf{v}_{local}^{(j)} \\in \\mathbf{D}_{local} \\ \\textbf{do} \\\\ & | \\ \\mathbf{d}_{\\texttt{fused}}^{(j)} \\leftarrow \\mathbf{v}_{local}^{(j)} + \\mathbf{v}_{\\texttt{global}} \\ ; \\qquad \\textit{//} \\ \\text{Fuse by element-wise} \\\\ & \\ \\ \\text{addition} \\\\ & | \\ \\ \\mathbf{D}_{\\texttt{ColParse}} \\leftarrow \\mathbf{D}_{\\texttt{ColParse}} \\cup \\{\\mathbf{d}_{\\texttt{fused}}^{(j)}\\} \\end{aligned}", "source": "marker_v2", "marker_block_id": "/page/11/Text/10"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0092", "section": "Algorithm 1 The Offline Indexing Process of ColParse", "page_start": 12, "page_end": 12, "type": "Text", "text": "end", "source": "marker_v2", "marker_block_id": "/page/11/Text/11"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0093", "section": "Algorithm 1 The Offline Indexing Process of ColParse", "page_start": 12, "page_end": 12, "type": "Text", "text": "return D_{\\texttt{ColParse}}", "source": "marker_v2", "marker_block_id": "/page/11/Text/12"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0094", "section": "B. More Theoretical Analysis", "page_start": 12, "page_end": 12, "type": "Text", "text": "This section provides a detailed theoretical exposition of the concepts introduced in Section 3, grounding the ColParse framework in fundamental principles of information theory.", "source": "marker_v2", "marker_block_id": "/page/11/Text/14"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0095", "section": "B.1. Information-Theoretic Preliminaries", "page_start": 12, "page_end": 12, "type": "Text", "text": "We begin by defining the core concepts used in our analysis. Definition B.1 (Mutual Information). The mutual information I(X;Y) between two random variables X and Y", "source": "marker_v2", "marker_block_id": "/page/11/Text/16"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0096", "section": "B.1. Information-Theoretic Preliminaries", "page_start": 12, "page_end": 12, "type": "Code", "text": "Algorithm 2 The Online Retrieval Process with ColParse", "source": "marker_v2", "marker_block_id": "/page/11/Code/17"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0097", "section": "B.1. Information-Theoretic Preliminaries", "page_start": 12, "page_end": 12, "type": "Text", "text": "/* Step 1: Encode Query", "source": "marker_v2", "marker_block_id": "/page/11/Text/19"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0098", "section": "B.1. Information-Theoretic Preliminaries", "page_start": 12, "page_end": 12, "type": "Text", "text": "\\mathbf{Q} \\leftarrow \\Phi_{\\mathrm{enc}}(q) ; // Encode q into N_q token vectors \\{\\mathbf{q}_i\\}", "source": "marker_v2", "marker_block_id": "/page/11/Text/20"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0099", "section": "B.1. Information-Theoretic Preliminaries", "page_start": 12, "page_end": 12, "type": "Text", "text": "/* Step 2: Late-Interaction Scoring (MaxSim) */ score \\leftarrow 0 for each query vector \\mathbf{q}_i \\in \\mathbf{Q} do", "source": "marker_v2", "marker_block_id": "/page/11/Text/21"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0100", "section": "B.1. Information-Theoretic Preliminaries", "page_start": 12, "page_end": 12, "type": "Text", "text": "*/", "source": "marker_v2", "marker_block_id": "/page/11/Text/44"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0101", "section": "B.1. Information-Theoretic Preliminaries", "page_start": 12, "page_end": 12, "type": "Text", "text": "max\\_sim \\leftarrow -\\infty for each fused document vector \\mathbf{d}_{fused}^{(j)} \\in \\mathbf{D}_{ColParse} do sim \\leftarrow \\mathbf{q}_i^{\\top} \\mathbf{d}_{fused}^{(j)} ; // Assuming L2-normalized vectors max\\_sim \\leftarrow \\max(max\\_sim, sim) end score \\leftarrow score + max\\_sim ; // Aggregate max", "source": "marker_v2", "marker_block_id": "/page/11/Text/22"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0102", "section": "B.1. Information-Theoretic Preliminaries", "page_start": 12, "page_end": 12, "type": "Text", "text": "similarity", "source": "marker_v2", "marker_block_id": "/page/11/Text/23"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0103", "section": "B.1. Information-Theoretic Preliminaries", "page_start": 12, "page_end": 12, "type": "Text", "text": "end", "source": "marker_v2", "marker_block_id": "/page/11/Text/24"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0104", "section": "return score", "page_start": 12, "page_end": 12, "type": "Text", "text": "measures their mutual dependence. It is defined as:", "source": "marker_v2", "marker_block_id": "/page/11/Text/26"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0105", "section": "return score", "page_start": 12, "page_end": 12, "type": "Equation", "text": "I(X;Y) = \\sum_{x \\in \\mathcal{X}} \\sum_{y \\in \\mathcal{Y}} p(x,y) \\log \\frac{p(x,y)}{p(x)p(y)}. (7)", "source": "marker_v2", "marker_block_id": "/page/11/Equation/27"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0106", "section": "return score", "page_start": 12, "page_end": 12, "type": "Text", "text": "where p(x,y) is the joint probability distribution, and p(x) and p(y) are the marginal distributions. I(X;Y)=0 if and only if X and Y are independent.", "source": "marker_v2", "marker_block_id": "/page/11/Text/28"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0107", "section": "return score", "page_start": 12, "page_end": 12, "type": "Text", "text": "Definition B.2 (Conditional Mutual Information). The conditional mutual information I(X;Y|Z) measures the mutual information between X and Y given that a third variable Z is known:", "source": "marker_v2", "marker_block_id": "/page/11/Text/29"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0108", "section": "return score", "page_start": 12, "page_end": 12, "type": "Equation", "text": "I(X;Y|Z) = \\mathbb{E}_{z \\sim p(z)}[I(X;Y|Z=z)]. \\tag{8}", "source": "marker_v2", "marker_block_id": "/page/11/Equation/30"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0109", "section": "return score", "page_start": 12, "page_end": 12, "type": "Text", "text": "Theorem B.3 (Chain Rule for Mutual Information). For a set of random variables \\{X_1, \\ldots, X_n\\} and another variable Y, the chain rule states:", "source": "marker_v2", "marker_block_id": "/page/11/Text/31"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0110", "section": "return score", "page_start": 12, "page_end": 12, "type": "Equation", "text": "I(X_1, \\dots, X_n; Y) = \\sum_{i=1}^n I(X_i; Y | X_1, \\dots, X_{i-1}). (9)", "source": "marker_v2", "marker_block_id": "/page/11/Equation/32"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0111", "section": "return score", "page_start": 12, "page_end": 12, "type": "Text", "text": "This rule is fundamental for decomposing the information content of a complex system.", "source": "marker_v2", "marker_block_id": "/page/11/Text/33"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0112", "section": "return score", "page_start": 12, "page_end": 12, "type": "Text", "text": "Theorem B.4 (Data Processing Inequality (DPI)). For any Markov chain of random variables X \\to Y \\to Z , where Z is conditionally independent of X given Y, the following inequality holds:", "source": "marker_v2", "marker_block_id": "/page/11/Text/34"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0113", "section": "return score", "page_start": 12, "page_end": 12, "type": "Equation", "text": "I(X;Z) < I(X;Y) \\text{ and } I(X;Z) < I(Y;Z). (10)", "source": "marker_v2", "marker_block_id": "/page/11/Equation/35"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0114", "section": "return score", "page_start": 12, "page_end": 12, "type": "Text", "text": "This theorem formalizes the notion that post-processing (the step from Y to Z) cannot increase information about the original source X.", "source": "marker_v2", "marker_block_id": "/page/11/Text/36"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0115", "section": "B.2. The Information Bottleneck (IB) Principle in VDR", "page_start": 13, "page_end": 13, "type": "Text", "text": "As stated in Section 3.3, the VDR compression task can be framed as an IB problem (Tishby et al., 2000). The objective is to find a compressed representation Z of a document D that maximizes information about a relevance variable R, while minimizing information about the source D itself.", "source": "marker_v2", "marker_block_id": "/page/12/Text/2"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0116", "section": "B.2. The Information Bottleneck (IB) Principle in VDR", "page_start": 13, "page_end": 13, "type": "Text", "text": "Proof of Intractability. The IB Lagrangian (Eq. 3 in the main text) requires computing an expectation over the distribution of all possible queries, P(Q).", "source": "marker_v2", "marker_block_id": "/page/12/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0117", "section": "B.2. The Information Bottleneck (IB) Principle in VDR", "page_start": 13, "page_end": 13, "type": "Equation", "text": "\\mathcal{L}(Z) = I(Z; D) - \\beta \\int_{q \\in \\mathcal{Q}} P(q)I(Z; R(D, q))dq. \\quad (11)", "source": "marker_v2", "marker_block_id": "/page/12/Equation/4"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0118", "section": "B.2. The Information Bottleneck (IB) Principle in VDR", "page_start": 13, "page_end": 13, "type": "Text", "text": "Since P(Q) is unknown and potentially infinite at the time of document indexing, this objective cannot be directly optimized. Therefore, practical methods must rely on principled approximations or surrogates for this ideal objective. ColParse provides such a surrogate.", "source": "marker_v2", "marker_block_id": "/page/12/Text/5"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0119", "section": "B.3. Justification for Structural Disentanglement", "page_start": 13, "page_end": 13, "type": "Text", "text": "ColParse's parsing stage, \\Psi_{\\text{parse}}(D) = \\{S_1, \\dots, S_k\\} , is justified by the Semantic Concentration Axiom. We now provide a more formal justification.", "source": "marker_v2", "marker_block_id": "/page/12/Text/7"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0120", "section": "B.3. Justification for Structural Disentanglement", "page_start": 13, "page_end": 13, "type": "Text", "text": "Axiom B.5 (Semantic Concentration). For a given query Q = q, there exists a primary semantic region S_{j^*} \\in \\{S_j\\} that contains almost all the information required to determine relevance. The remaining regions S_{\\neg j^*} = \\{S_j\\}_{j \\neq j^*} provide negligible additional information.", "source": "marker_v2", "marker_block_id": "/page/12/Text/8"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0121", "section": "B.3. Justification for Structural Disentanglement", "page_start": 13, "page_end": 13, "type": "Equation", "text": "I(S_{\\neg j^*}; R|S_{j^*}, Q = q) \\approx 0. (12)", "source": "marker_v2", "marker_block_id": "/page/12/Equation/9"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0122", "section": "B.3. Justification for Structural Disentanglement", "page_start": 13, "page_end": 13, "type": "Text", "text": "Justification. This axiom is an empirical assumption about the nature of user queries and documents. For a query \"What were the revenues in Q3 2023?\", the answer is almost certainly contained entirely within a single financial table. Information in other regions (e.g., the abstract, a methodology figure) is conditionally irrelevant once the correct table is identified.", "source": "marker_v2", "marker_block_id": "/page/12/Text/10"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0123", "section": "B.3. Justification for Structural Disentanglement", "page_start": 13, "page_end": 13, "type": "Text", "text": "Corollary B.6 (Information Equivalence of Decomposed Representation). Under the Semantic Concentration Axiom, the mutual information between the entire document and the relevance variable is approximately equal to the maximum information contained in any single semantic region.", "source": "marker_v2", "marker_block_id": "/page/12/Text/11"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0124", "section": "B.3. Justification for Structural Disentanglement", "page_start": 13, "page_end": 13, "type": "Equation", "text": "I(D;R) \\approx \\max_{j \\in \\{1,\\dots,k\\}} I(S_j;R). \\tag{13}", "source": "marker_v2", "marker_block_id": "/page/12/Equation/12"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0125", "section": "B.3. Justification for Structural Disentanglement", "page_start": 13, "page_end": 13, "type": "Text", "text": "Proof. From the chain rule, I(D;R) = I(S_1, ..., S_k; R) . For a specific query q, let j^* be the index of the primary region. We have:", "source": "marker_v2", "marker_block_id": "/page/12/Text/13"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0126", "section": "B.3. Justification for Structural Disentanglement", "page_start": 13, "page_end": 13, "type": "Equation", "text": "I(D; R|Q = q) = I(S_{j^*}; R|Q = q) + I(S_{\\neg j^*}; R|S_{j^*}, Q = q). (14)", "source": "marker_v2", "marker_block_id": "/page/12/Equation/14"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0127", "section": "B.3. Justification for Structural Disentanglement", "page_start": 13, "page_end": 13, "type": "Text", "text": "Applying Axiom B.5, the second term vanishes: I(D; R|Q=q) \\approx I(S_{j^*}; R|Q=q) . Taking the expectation over P(Q), and using the property that \\mathbb{E}[\\max(X_i)] \\geq \\max(\\mathbb{E}[X_i]) , we arrive at the approximation that the total information is well-represented by the information in the best possible channel, justifying the multi-vector approach. \\square", "source": "marker_v2", "marker_block_id": "/page/12/Text/15"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0128", "section": "B.4. Justification for Synergistic Fusion", "page_start": 13, "page_end": 13, "type": "Text", "text": "The fusion stage combines local vectors \\{V_j = \\Phi_{\\rm enc}(S_j)\\} with a global vector V_{\\rm global} = \\Phi_{\\rm enc}(D) to produce the final representation \\{Z_j = V_j + V_{\\rm global}\\} .", "source": "marker_v2", "marker_block_id": "/page/12/Text/17"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0129", "section": "B.4. Justification for Synergistic Fusion", "page_start": 13, "page_end": 13, "type": "Text", "text": "Definition B.7 (Contextual Information Gain). The contextual information gain for region j is the additional information about relevance R provided by the global context V_{\\text{global}} , given that the local information V_j is already known.", "source": "marker_v2", "marker_block_id": "/page/12/Text/18"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0130", "section": "B.4. Justification for Synergistic Fusion", "page_start": 13, "page_end": 13, "type": "Equation", "text": "G_j^{\\text{context}} \\triangleq I(V_{\\text{global}}; R|V_j). (15)", "source": "marker_v2", "marker_block_id": "/page/12/Equation/19"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0131", "section": "B.4. Justification for Synergistic Fusion", "page_start": 13, "page_end": 13, "type": "Text", "text": "Theorem B.8 (Information in the Fused Representation). The information contained in the fused vector Z_j = V_j + V_{global} is upper-bounded by the joint information of its components.", "source": "marker_v2", "marker_block_id": "/page/12/Text/20"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0132", "section": "B.4. Justification for Synergistic Fusion", "page_start": 13, "page_end": 13, "type": "Equation", "text": "I(Z_j; R) \\le I(V_j, V_{global}; R). \\tag{16}", "source": "marker_v2", "marker_block_id": "/page/12/Equation/21"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0133", "section": "B.4. Justification for Synergistic Fusion", "page_start": 13, "page_end": 13, "type": "Text", "text": "Proof. The fused vector Z_j is a deterministic function of V_j and V_{\\text{global}} . This forms the Markov chain (V_j, V_{\\text{global}}) \\to Z_j \\to R . Applying the Data Processing Inequality (Theorem B.4) to this chain directly yields the result.", "source": "marker_v2", "marker_block_id": "/page/12/Text/22"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0134", "section": "B.4. Justification for Synergistic Fusion", "page_start": 13, "page_end": 13, "type": "Text", "text": "Corollary B.9 (Condition for Information Improvement). The fusion step is beneficial (i.e., Z_j is more informative than V_j alone) if and only if the fusion function successfully captures a non-zero portion of the contextual information gain.", "source": "marker_v2", "marker_block_id": "/page/12/Text/23"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0135", "section": "B.4. Justification for Synergistic Fusion", "page_start": 13, "page_end": 13, "type": "Equation", "text": "\\Delta I_j = I(Z_j; R) - I(V_j; R) > 0 \\iff I(Z_j; R|V_j) > 0. (17)", "source": "marker_v2", "marker_block_id": "/page/12/Equation/24"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0136", "section": "B.4. Justification for Synergistic Fusion", "page_start": 13, "page_end": 13, "type": "Text", "text": "Proof. From the chain rule, I(Z_j,V_j;R) = I(V_j;R) + I(Z_j;R|V_j) . Since Z_j is a function of V_j and V_{\\text{global}} , knowing V_j does not make Z_j fully determined. The term I(Z_j;R|V_j) represents the information that the variation in Z_j (caused by V_{\\text{global}} ) provides about R, even when V_j is fixed. A positive net improvement \\Delta I_j > 0 directly requires this conditional term to be positive, which in turn means the fusion must have encoded some of the contextual gain G_j^{\\text{context}} . The vector addition V_j + V_{\\text{global}} is a simple, effective function for this purpose, as it non-linearly interacts with the query vector during the dot product scoring: \\mathbf{q}^{\\top}(\\mathbf{v}_j + \\mathbf{v}_{\\text{global}}) , allowing the model to utilize both local and global signals.", "source": "marker_v2", "marker_block_id": "/page/12/Text/25"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0137", "section": "C.1. Benchmark Details", "page_start": 14, "page_end": 14, "type": "Text", "text": "To ensure a comprehensive and robust evaluation of our framework, we anchor our experiments on five mainstream benchmark suites for VDR, all of which are integrated within the visdoc section of the MMEB (Meng et al., 2025) . The following benchmarks collectively cover a diverse range of document types, query complexities, and retrieval scenarios, providing a multifaceted view of model performance.", "source": "marker_v2", "marker_block_id": "/page/13/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0138", "section": "C.1. Benchmark Details", "page_start": 14, "page_end": 14, "type": "ListGroup", "text": "▶ ViDoRe-V1 (Faysse et al., 2024) 3 : As a foundational benchmark for page-level VDR, ViDoRe-V1 was one of the first to systematically evaluate systems on visuallyrich documents. It combines repurposed academic VQA datasets with practical, topic-specific tasks, highlighting the inherent shortcomings of traditional text-only retrieval systems on documents containing complex layouts, tables, and figures. ▶ ViDoRe-V2 (Mace et al. ´ , 2025) 4 : As a successor to ViDoRe-V1, ViDoRe-V2 aims to raise the bar by introducing more challenging and realistic retrieval scenarios to address the performance saturation observed on the original. Its core contributions include the use of long-form, cross-document, and multilingual queries generated via a hybrid synthetic and human-in-the-loop process, which reduces extractive bias and more accurately reflects real-world user interactions. ▶ VisRAG (Yu et al., 2024) 5 : The VisRAG benchmark is constructed to specifically evaluate vision-based RAG pipelines by aggregating and refining multiple existing VQA datasets. Its primary contribution is the unification of a wide spectrum of document types—including scientific figures, charts, infographics, and presentation slides—under a single evaluation framework, coupled with a crucial filtering process to remove contextdependent questions and ensure its suitability for openretrieval tasks. ▶ ViDoSeek (Wang et al., 2025) 6 : ViDoSeek is a novel benchmark designed to evaluate end-to-end RAG systems on visually-rich documents that require complex reasoning. Its main contribution lies in providing a large document corpus where each query corresponds to a unique answer, which allows for a more realistic and rigorous evaluation of both the retrieval and subsequent reasoning stages in a large-scale setting.", "source": "marker_v2", "marker_block_id": "/page/13/ListGroup/572"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0139", "section": "C.1. Benchmark Details", "page_start": 14, "page_end": 14, "type": "Text", "text": "▶ MMLongBench (Ma et al., 2024b) 7 : MMLongBench is specifically designed to assess the long-context, multimodal understanding capabilities of LVLMs. It stands out by using lengthy documents (averaging 47.5 pages) and featuring a significant portion of cross-page questions that require multi-hop reasoning, as well as unanswerable questions to probe for model hallucination, thus rigorously testing a model's ability to locate and synthesize information from extensive contexts.", "source": "marker_v2", "marker_block_id": "/page/13/Text/11"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0140", "section": "C.2. Model Details", "page_start": 14, "page_end": 14, "type": "Text", "text": "We select ten representative single-vector multimodal retrieval models from recent literature to serve as the base models for our experiments. These models, built upon various architectures and pre-training paradigms, provide a comprehensive testbed for evaluating the versatility and effectiveness of our proposed framework.", "source": "marker_v2", "marker_block_id": "/page/13/Text/13"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0141", "section": "C.2. Model Details", "page_start": 14, "page_end": 14, "type": "ListGroup", "text": "▶ VLM2Vec-V1-2B/7B (Jiang et al., 2024) 8 : As a pioneering work in universal multimodal embeddings, VLM2Vec introduces a contrastive training framework to adapt any VLM for a wide range of tasks. Its core contribution is reformulating diverse multimodal tasks ( e.g., classification, VQA, retrieval) into a unified instructionfollowing ranking problem, enabling the model to learn general-purpose embeddings for both images and text. ▶ VLM2Vec-V2-2B (Meng et al., 2025) 9 : This model extends its predecessor by broadening the scope of multimodal embeddings to include videos and visual documents, in addition to images and text. Its primary contribution is the introduction of a more comprehensive benchmark and a unified training strategy that allows a single model to effectively learn representations across static, temporal, and structured visual data formats. ▶ LamRA-Ret-7B (Liu et al., 2025a) 10 : LamRA explores repurposing generative Large Multimodal Models for retrieval tasks, unifying diverse retrieval scenarios under a single instruction-following framework. Its key innovation is a two-stage training strategy that first pretrains the model on language-only tasks before multimodal instruction tuning, progressively adapting the generative model for retrieval. ▶ GME-2B/7B (Zhang et al., 2024b) 11 : The General Multimodal Embedder (GME) framework focuses on", "source": "marker_v2", "marker_block_id": "/page/13/ListGroup/573"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0142", "section": "C.2. Model Details", "page_start": 14, "page_end": 14, "type": "Code", "text": "7 yubo2333/MMLongBench-Doc 8 VLM2Vec-Full 9 VLM2Vec-V2.0 10 LamRA-Ret 11 Alibaba-NLP/gme-models", "source": "marker_v2", "marker_block_id": "/page/13/Code/18"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0143", "section": "C.2. Model Details", "page_start": 14, "page_end": 14, "type": "Footnote", "text": "3 vidore/vidore-benchmark 4", "source": "marker_v2", "marker_block_id": "/page/13/Footnote/8"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0144", "section": "C.2. Model Details", "page_start": 14, "page_end": 14, "type": "Footnote", "text": "vidore/vidore-benchmark-v2 5 openbmb/visrag", "source": "marker_v2", "marker_block_id": "/page/13/Footnote/9"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0145", "section": "C.2. Model Details", "page_start": 14, "page_end": 14, "type": "Footnote", "text": "6 Qiuchen-Wang/ViDoSeek", "source": "marker_v2", "marker_block_id": "/page/13/Footnote/10"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0146", "section": "C.2. Model Details", "page_start": 15, "page_end": 15, "type": "Text", "text": "improving universal multimodal retrieval by leveraging a more diverse mix of training data, including singlemodal, cross-modal, and fused-modal examples. Its core contribution is a novel data synthesis pipeline for creating large-scale, high-quality fused-modal data, which significantly enhances the model's ability to handle complex queries and retrieve visual documents.", "source": "marker_v2", "marker_block_id": "/page/14/Text/2"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0147", "section": "C.2. Model Details", "page_start": 15, "page_end": 15, "type": "ListGroup", "text": "▶ UniME-V2-2B/7B (Gu et al., 2025) 12 : UniME-V2 enhances representation learning by leveraging an MLLM as a \"judge\" to generate soft semantic matching scores for query-candidate pairs. This MLLM-as-a-Judge mechanism facilitates more effective hard negative mining and allows the embedding model to learn finergrained semantic distinctions, significantly improving its discriminative capacity. ▶ B3-2B/7B (Thirukovalluru et al., 2025) 13 : Breaking the Batch Barrier (B3) introduces a novel batch construction strategy for contrastive learning that curates high-quality batches rich in hard negatives. Instead of random sampling, it uses a teacher model and graphbased community detection to group mutually challenging examples together, thereby improving training efficiency and achieving state-of-the-art performance even with significantly smaller batch sizes.", "source": "marker_v2", "marker_block_id": "/page/14/ListGroup/805"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0148", "section": "C.3. MinerU2.5 Details", "page_start": 15, "page_end": 15, "type": "Text", "text": "To resolve the trade-off between the immense computational overhead (O(N 2 ) complexity) and information loss associated with directly processing high-resolution document images, MinerU2.5 innovatively employs a decoupled, coarse-to-fine two-stage strategy:", "source": "marker_v2", "marker_block_id": "/page/14/Text/6"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0149", "section": "C.3. MinerU2.5 Details", "page_start": 15, "page_end": 15, "type": "ListGroup", "text": "1. Stage I: Global Layout Analysis. In this stage, the model first resizes the input document image to a medium-resolution thumbnail ( e.g., 1036×1036 pixels). It then performs a fast, global layout analysis on this thumbnail to identify all structural elements (such as paragraphs, tables, formulas, and figures) and their positions at a low computational cost. 2. Stage II: Local Content Recognition. Guided by the layout information detected in the first stage, the model precisely crops the respective semantic regions from the original high-resolution image. Subsequently, it performs parallel, fine-grained content recognition ( e.g., text OCR, table structuring, formula transcription) on these native-resolution cropped patches. This preserves high recognition accuracy while avoiding redundant computations on the entire high-resolution image.", "source": "marker_v2", "marker_block_id": "/page/14/ListGroup/806"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0150", "section": "C.3. MinerU2.5 Details", "page_start": 15, "page_end": 15, "type": "Text", "text": "Algorithm 3 details the layout-informed image splitting", "source": "marker_v2", "marker_block_id": "/page/14/Text/9"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0151", "section": "C.3. MinerU2.5 Details", "page_start": 15, "page_end": 15, "type": "Text", "text": "process used in ColParse .", "source": "marker_v2", "marker_block_id": "/page/14/Text/12"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0152", "section": "C.3. MinerU2.5 Details", "page_start": 15, "page_end": 15, "type": "Code", "text": "Algorithm 3 Layout-Informed Image Splitting for ColParse Input : A document image d ∈ R H×W×3 ; A layout detector model Ψsplit (e.g., DocLayoutY- OLO); Minimum area ratio threshold τ (default 0.01); Maximum sub-images count Nmax (default 20); Grid fallback parameters: Rgrid, Cgrid Output : A list of cropped sub-images Sd; A list of content type labels Cd (optional) /* Step 1: Semantic Layout Detection */ TotalArea ← H × W B ← ∅, Cd ← ∅ if Ψsplit is available then R ← Ψsplit.predict(d) ; // Returns list of {bbox, category, score} if R is not empty then for each region r ∈ R do b ← (x1, y1, x2, y2) from r.poly c ← MapCategoryID(r.category id) B ← B ∪ {(b, c, centerY(b), centerX(b))} end end end /* Step 2: Fallback & Sorting Mechanism */ if B is empty then B ← GridBasedSplit(H, W, Rgrid, Cgrid) ; // Fallback to grid else B ← SortByReadingOrder(B) ; // Sort by vertical bands, then horizontal end /* Step 3: Filtering, Cropping and Output */ Sd ← ∅, count ← 0 for each (b, c) ∈ B do if count ≥ Nmax then break end Area ← width(b) × height(b) if Area/TotalArea ≥ τ then s ← Crop(d, b) ; // Extract region from original image Sd ← Sd∪{s} Cd ← Cd∪{c} count ← count+1 end end return Sd, Cd", "source": "marker_v2", "marker_block_id": "/page/14/Code/13"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0153", "section": "C.4. Main Results", "page_start": 15, "page_end": 15, "type": "Text", "text": "Refer to Table 2 and Table 3 for all results of ColParse and baselines across five benchmarks.", "source": "marker_v2", "marker_block_id": "/page/14/Text/15"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0154", "section": "C.4. Main Results", "page_start": 15, "page_end": 15, "type": "Footnote", "text": "12 TianchengGu/unime-v2", "source": "marker_v2", "marker_block_id": "/page/14/Footnote/10"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0155", "section": "C.4. Main Results", "page_start": 15, "page_end": 15, "type": "Footnote", "text": "13", "source": "marker_v2", "marker_block_id": "/page/14/Footnote/11"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0156", "section": "C.4. Main Results", "page_start": 16, "page_end": 16, "type": "Caption", "text": "Table 2. Performance comparison on MMLongBench and ViDoRe-V1 benchmarks. For each model block, we bold the best-performing optimization method in each column (except for the base result). The average scores for optimizations are shown with relative gains (↑/↓) compared to the base model.", "source": "marker_v2", "marker_block_id": "/page/15/Caption/1"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0157", "section": "C.4. Main Results", "page_start": 16, "page_end": 16, "type": "Table", "text": "MMLongBench ViDoRe-V1 Method Doc Page Avg. Arxiv DocV InfoV Shift TabF TatD S-AI S-En S-HC S-Gov Avg. VLM2Vec-V1-2B 25.62 26.23 25.93 17.80 13.98 39.41 9.18 36.32 10.56 16.39 15.96 23.56 24.11 20.73 s2m-add 21.54 15.08 18.31↓7.62 35.07 15.61 52.15 6.62 36.51 10.39 26.23 31.22 30.29 33.61 27.77↑7.04 s2m-mul 22.07 15.10 18.59↓7.34 34.91 16.12 52.61 6.57 36.65 10.34 23.90 30.89 28.88 33.60 27.45↑6.72 cl-t-c 16.57 10.59 13.58↓12.35 13.97 3.97 21.62 8.73 23.20 11.56 12.94 28.27 19.16 26.72 17.01↓3.72 cl-t-m 14.35 8.67 11.51↓14.42 16.47 2.98 26.97 16.27 18.48 8.59 14.40 23.03 18.71 13.10 15.90↓4.83 cl-s-c 18.29 11.54 14.92↓11.01 18.83 4.87 22.80 13.49 24.64 12.63 15.45 24.88 20.78 20.85 17.92↓2.81 cl-s-m 15.45 9.06 12.26↓13.67 15.73 2.54 27.46 15.68 18.99 9.76 13.52 24.73 22.59 20.72 17.17↓3.56 c-sem 18.76 13.88 16.32↓9.61 29.33 5.87 35.70 22.38 35.38 13.95 23.98 37.88 33.91 30.44 26.88↑6.15 multi-img 23.61 15.85 19.73↓6.20 37.78 13.96 54.20 11.35 40.50 9.10 20.36 32.15 36.75 32.04 28.82↑8.09 ColParse 34.31 29.83 32.07↑6.14 47.66 28.12 69.23 47.11 57.05 20.43 62.24 63.77 65.51 62.54 52.37↑31.64 VLM2Vec-V1-7B 23.85 37.63 30.74 28.07 17.93 44.47 2.06 16.78 5.86 17.93 25.04 28.90 14.59 20.16 s2m-add 34.57 28.11 31.34↑0.60 50.30 25.66 66.73 38.21 63.75 23.49 70.27 61.87 70.68 66.38 53.73↑33.57 s2m-mul 35.29 28.49 31.89↑1.15 50.46 26.34 67.28 36.45 64.13 23.84 69.43 61.94 68.35 66.99 53.52↑33.36 cl-t-c 18.75 14.31 16.53↓14.21 17.10 7.71 26.41 21.58 29.62 14.61 18.52 27.98 26.09 26.65 21.63↑1.47 cl-t-m 25.34 20.46 22.90↓7.84 28.06 7.00 47.43 29.48 45.92 19.42 27.66 47.56 44.96 52.94 35.04↑14.88 cl-s-c 22.25 14.48 18.37↓12.37 21.80 8.11 30.92 28.36 25.91 19.61 27.40 40.30 36.25 32.58 27.12↑6.96 cl-s-m 26.31 20.38 23.35↓7.39 28.58 8.85 49.12 32.67 46.07 20.85 31.96 50.86 50.85 49.99 36.98↑16.82 c-sem 31.36 26.16 28.76↓1.98 45.77 15.38 59.20 37.46 57.05 30.17 47.00 60.89 64.22 66.76 48.39↑28.23 multi-img 33.77 25.60 29.69↓1.05 49.40 19.55 62.09 28.19 66.34 17.19 41.89 51.44 60.84 48.89 44.58↑24.42 ColParse 43.34 40.58 41.96↑11.22 60.47 34.42 70.39 53.67 77.12 31.33 74.81 69.64 80.79 75.89 62.85↑42.69 VLM2Vec-V2-2B 48.55 50.34 49.45 78.98 38.51 82.21 64.57 87.64 44.68 85.06 82.99 89.89 87.08 74.16 s2m-add 43.33 39.00 41.17↓8.28 66.60 38.47 72.80 58.28 65.85 54.50 90.10 84.97 83.93 80.36 69.59↓4.57 s2m-mul 45.72 40.66 43.19↓6.26 68.04 39.80 75.45 58.87 69.90 54.31 90.93 84.53 84.56 82.40 70.88↓3.28 cl-t-c 20.08 18.48 19.28↓30.17 28.47 6.46 29.32 20.51 40.06 17.75 20.23 35.23 27.90 22.54 24.85↓49.31 cl-t-m 25.04 21.94 23.49↓25.96 44.82 7.28 42.98 31.23 38.96 21.19 26.38 47.14 45.34 37.44 34.28↓39.88 cl-s-c 23.25 18.43 20.84↓28.61 29.34 8.76 27.77 22.92 40.48 21.16 20.26 30.91 24.61 27.64 25.39↓48.77 cl-s-m 26.08 22.30 24.19↓25.26 44.95 7.16 45.42 19.48 39.29 25.44 31.57 47.18 48.84 40.13 34.95↓39.21 c-sem 29.94 27.99 28.97↓20.48 61.30 17.28 62.19 40.55 55.53 33.02 56.52 62.33 68.54 70.10 52.74↓21.42 multi-img 38.69 29.00 33.85↓15.60 65.39 25.80 70.54 27.92 71.56 33.79 55.93 62.93 72.27 53.14 53.93↓20.23 ColParse 49.49 50.53 50.01↑0.56 80.17 46.33 83.53 72.76 86.74 52.40 91.36 85.83 95.47 89.52 78.41↑4.25 LamRA-Ret 19.78 13.24 16.51 29.31 19.56 63.00 15.83 51.44 7.70 21.10 29.81 37.18 31.95 30.69 s2m-add 32.18 17.82 25.00↑8.49 9.80 14.37 46.06 19.49 28.13 19.16 22.79 30.98 37.98 24.81 25.36↓5.33 s2m-mul 30.52 16.37 23.45↑6.94 9.71 14.98 45.61 17.37 27.91 17.20 20.45 28.32 38.60 23.01 24.32↓6.37 cl-t-c 14.88 8.09 11.49↓5.02 5.57 2.22 19.76 13.94 17.55 6.59 14.23 20.69 17.83 20.52 13.89↓16.80 cl-t-m 20.43 13.28 16.86↑0.35 7.47 3.85 30.91 17.54 13.41 12.37 27.06 35.63 39.20 31.88 21.93↓8.76 cl-s-c 15.32 9.05 12.19↓4.32 6.78 2.62 20.53 13.79 18.30 9.11 17.05 27.03 25.60 22.25 16.31↓14.38 cl-s-m 19.84 12.91 16.38↓0.13 7.75 4.82 32.16 15.61 14.02 10.23 24.81 34.66 37.31 27.20 20.86↓9.83 c-sem 19.62 10.69 15.16↓1.35 13.51 5.91 34.43 9.13 32.30 15.94 26.32 36.02 39.82 37.01 25.04↓5.65 multi-img 23.32 7.71 15.52↓0.99 6.90 6.10 21.26 8.00 22.26 13.60 13.91 15.91 20.30 9.67 13.79↓16.90 ColParse 30.74 19.50 25.12↑8.61 17.27 20.61 58.35 21.39 39.69 13.34 25.57 35.27 43.21 26.29 30.10↓0.59 GME-2B 52.07 53.14 52.61 82.59 56.46 88.97 89.72 93.20 70.33 98.49 92.15 98.15 95.65 86.57 s2m-add 50.92 46.52 48.72↓3.89 71.85 43.00 83.58 72.17 77.65 69.45 93.29 89.80 89.92 90.50 78.12↓8.45 s2m-mul 53.82 50.06 51.94↓0.67 76.87 47.23 85.49 81.53 84.77 74.63 95.72 93.07 93.85 92.18 82.53↓4.04 cl-t-c 15.54 10.26 12.90↓39.71 14.96 4.70 25.32 16.93 24.97 12.01 12.02 19.02 19.26 19.34 16.85↓69.72 cl-t-m 17.95 13.61 15.78↓36.83 30.20 3.58 33.07 30.13 31.16 16.11 22.27 33.99 36.32 33.28 27.01↓59.56 cl-s-c 16.03 12.59 14.31↓38.30 15.90 5.88 22.23 26.42 26.16 15.36 20.67 18.52 25.26 22.00 19.84↓66.73", "source": "marker_v2", "marker_block_id": "/page/15/Table/2"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0158", "section": "C.4. Main Results", "page_start": 16, "page_end": 16, "type": "Text", "text": "Continued on next page", "source": "marker_v2", "marker_block_id": "/page/15/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0159", "section": "C.4. Main Results", "page_start": 17, "page_end": 17, "type": "Caption", "text": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "source": "marker_v2", "marker_block_id": "/page/16/Caption/0"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0160", "section": "C.4. Main Results", "page_start": 17, "page_end": 17, "type": "Table", "text": "Table 2 – Continued from previous page MMLongBench ViDoRe-V1 Method Doc Page Avg. Arxiv DocV InfoV Shift TabF TatD S-AI S-En S-HC S-Gov Avg. cl-s-m 16.39 12.39 14.39↓38.22 31.79 4.01 30.55 19.79 31.61 14.18 24.92 30.34 34.24 28.50 24.99↓61.58 c-sem 23.78 17.15 20.47↓32.14 45.19 11.49 51.57 25.34 42.78 27.26 51.26 44.73 49.30 53.58 40.25↓46.32 multi-img 45.10 32.87 38.99↓13.62 67.63 27.89 74.78 47.02 78.34 43.20 59.99 63.76 73.68 63.19 59.95↓26.62 ColParse 53.06 54.24 53.65↑1.04 82.39 54.11 88.93 88.59 92.33 70.65 97.75 92.30 97.91 96.10 86.11↓0.46 GME-7B 54.01 55.80 54.91 87.59 56.05 91.96 94.25 93.72 76.26 99.63 95.45 99.63 99.06 89.36 s2m-add 53.57 49.76 51.67↓3.24 75.91 46.41 85.01 80.64 83.47 74.66 95.72 92.93 94.35 92.17 82.13↓7.23 s2m-mul 50.73 45.95 48.34↓6.57 72.29 43.38 83.23 72.85 78.28 69.64 92.92 90.23 90.05 90.17 78.30↓11.06 cl-t-c 12.53 8.39 10.46↓44.45 7.32 5.16 17.36 18.86 16.81 12.83 13.28 16.18 20.30 13.97 14.21↓75.15 cl-t-m 14.64 9.03 11.84↓43.07 6.78 3.80 26.36 24.67 17.82 15.52 19.12 26.65 24.74 28.60 19.41↓69.95 cl-s-c 13.08 8.93 11.01↓43.90 7.25 3.85 16.83 17.08 18.68 13.51 19.17 20.53 19.15 21.64 15.77↓73.59 cl-s-m 13.62 7.93 10.78↓44.13 6.88 4.73 22.77 18.40 18.43 10.43 16.80 21.03 26.40 20.86 16.67↓72.69 c-sem 15.42 9.22 12.32↓42.59 15.02 6.13 27.52 18.27 36.39 22.53 20.40 28.91 30.81 26.13 23.21↓66.15 multi-img 47.50 36.01 41.76↓13.15 72.48 33.71 78.84 45.50 84.88 45.65 64.59 68.82 75.90 69.50 63.99↓25.37 ColParse 54.96 56.51 55.74↑0.83 87.35 57.91 90.76 95.35 95.44 75.92 99.63 94.67 99.63 98.89 89.56↑0.20 UniME-V2-2B 18.52 40.10 29.31 36.52 12.43 42.41 14.09 51.11 7.39 20.23 32.96 24.21 19.25 26.06 s2m-add 36.66 30.03 33.35↑4.04 50.78 25.04 58.61 37.71 54.90 36.67 68.00 68.50 69.21 68.39 53.78↑27.72 s2m-mul 36.06 28.97 32.52↑3.21 50.76 23.93 58.23 32.85 54.90 34.94 63.01 64.67 66.31 61.07 51.07↑25.01 cl-t-c 19.45 14.04 16.75↓12.56 19.03 7.04 25.79 22.55 30.85 14.57 13.25 21.82 32.30 26.67 21.39↓4.67 cl-t-m 16.70 12.63 14.67↓14.64 21.88 4.30 33.79 26.88 20.47 9.65 15.83 36.40 34.18 27.10 23.05↓3.01 cl-s-c 19.23 16.02 17.63↓11.68 19.53 6.89 23.75 21.96 28.29 15.97 17.34 31.06 33.58 27.73 22.61↓3.45 cl-s-m 17.77 13.19 15.48↓13.83 23.67 5.33 36.60 32.03 20.17 11.46 19.46 37.16 36.87 26.76 24.95↓1.11 c-sem 25.11 20.89 23.00↓6.31 38.24 13.31 51.27 38.76 40.04 21.25 33.42 51.86 60.03 52.80 40.10↑14.04 multi-img 33.99 24.39 29.19↓0.12 50.79 19.08 60.61 30.15 58.91 24.36 43.08 46.76 61.06 50.79 44.56↑18.50 ColParse 44.22 44.19 44.21↑14.90 62.39 37.69 73.33 72.35 77.45 38.83 82.50 75.80 89.35 85.84 69.55↑43.49 UniME-V2-7B 33.19 45.72 39.46 63.23 24.91 65.25 11.16 41.54 14.18 41.89 40.56 57.44 42.78 40.29 s2m-add 40.28 37.61 38.95↓0.51 55.17 32.79 69.82 58.54 65.47 45.46 84.23 81.72 87.38 89.16 66.97↑26.68 s2m-mul 40.57 38.14 39.36↓0.10 55.29 33.72 70.38 56.55 65.82 44.48 84.10 81.55 86.00 89.89 66.78↑26.49 cl-t-c 20.22 13.70 16.96↓22.50 24.63 5.79 29.62 27.36 21.81 15.81 17.79 34.01 27.97 28.61 23.34↓16.95 cl-t-m 23.20 19.90 21.55↓17.91 34.15 5.21 49.47 34.87 32.02 18.91 35.36 47.97 43.55 56.22 35.77↓4.52 cl-s-c 20.91 18.25 19.58↓19.88 23.98 8.28 28.30 26.59 28.51 21.08 25.63 32.59 30.66 43.46 26.91↓13.38 cl-s-m 24.81 19.71 22.26↓17.20 35.09 6.60 49.60 36.77 33.75 20.75 37.45 48.31 55.85 60.28 38.45↓1.84 c-sem 31.71 29.01 30.36↓9.10 59.37 19.84 64.17 42.06 50.69 34.28 70.63 66.71 72.14 75.85 55.57↑15.28 multi-img 34.79 28.05 31.42↓8.04 53.99 23.06 67.47 40.81 70.58 30.36 52.38 65.49 72.19 62.65 53.90↑13.61 ColParse 45.90 48.26 47.08↑7.62 64.78 37.43 78.51 76.74 81.47 43.69 89.32 82.68 92.74 88.13 73.55↑33.26 B3-2B 37.10 32.07 34.59 57.00 29.38 68.09 48.31 71.55 18.09 74.13 64.64 75.44 63.13 56.98 s2m-add 35.86 27.36 31.61↓2.98 47.93 24.57 64.85 39.99 50.16 33.16 66.24 63.90 69.75 65.45 52.60↓4.38 s2m-mul 35.66 26.97 31.32↓3.27 47.89 24.47 64.82 35.42 49.59 32.43 63.95 63.09 66.79 64.47 51.29↓5.69 cl-t-c 17.88 13.56 15.72↓18.87 19.64 4.53 32.73 10.60 20.15 9.10 12.68 26.80 30.88 18.64 18.58↓38.40 cl-t-m 18.36 14.95 16.66↓17.93 18.38 5.34 33.44 13.13 19.91 12.42 14.05 32.81 30.84 26.31 20.66↓36.32 cl-s-c 20.70 16.46 18.58↓16.01 16.92 5.67 34.77 16.79 22.56 13.04 22.41 34.47 37.31 36.75 24.07↓32.91 cl-s-m 20.41 17.63 19.02↓15.57 19.74 4.72 35.90 17.36 20.74 15.77 19.76 38.43 36.44 33.89 24.28↓32.70 c-sem 23.80 21.20 22.50↓12.09 34.98 11.15 51.26 21.98 37.45 19.41 29.16 49.03 54.62 46.73 35.58↓21.40 multi-img 28.93 16.01 22.47↓12.12 41.03 13.38 46.20 20.46 46.89 10.26 31.74 38.83 32.44 32.80 31.40↓25.58 ColParse 42.06 37.60 39.83↑5.24 56.47 30.91 66.69 67.42 69.33 29.42 79.88 72.67 83.24 71.41 62.74↑5.76 B3-7B 46.09 45.10 45.60 68.95 43.38 79.86 66.56 84.12 37.06 81.01 81.25 88.57 81.30 71.21 44.95 40.38 42.67↓2.93 59.11 38.45 75.42 69.63 70.95 51.71 88.55 81.68 86.10 86.34 70.79↓0.42 s2m-add s2m-mul cl-t-c 45.11 40.61 42.86↓2.74 23.96 19.73 21.85↓23.75 59.50 25.53 38.63 8.05 75.74 69.25 71.18 51.72 87.36 82.07 85.17 47.76 13.93 32.90 19.17 29.29 44.67 45.02 86.23 33.61 70.69↓0.52 29.99↓41.22", "source": "marker_v2", "marker_block_id": "/page/16/Table/1"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0161", "section": "C.4. Main Results", "page_start": 17, "page_end": 17, "type": "Text", "text": "Continued on next page", "source": "marker_v2", "marker_block_id": "/page/16/Text/2"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0162", "section": "C.4. Main Results", "page_start": 18, "page_end": 18, "type": "Caption", "text": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "source": "marker_v2", "marker_block_id": "/page/17/Caption/0"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0163", "section": "C.4. Main Results", "page_start": 18, "page_end": 18, "type": "Table", "text": "Table 2 – Continued from previous page Method MMLongBench ViDoRe-V1 Doc Page Avg. Arxiv DocV InfoV Shift TabF TatD S-AI S-En S-HC S-Gov Avg. cl-t-m 24.79 21.29 23.04↓22.56 31.72 9.21 52.63 18.92 28.84 22.25 31.48 47.83 51.71 43.89 33.85↓37.36 cl-s-c 25.66 21.29 23.48↓22.12 24.40 11.69 48.70 23.29 33.17 25.56 34.54 42.09 55.93 52.12 35.15↓36.06 cl-s-m 25.52 20.72 23.12↓22.48 24.78 11.31 49.68 26.51 32.71 26.20 37.63 46.29 51.53 44.47 35.11↓36.10 c-sem 29.88 25.98 27.93↓17.67 56.08 24.05 65.81 25.04 52.18 33.15 59.78 57.76 67.14 70.35 51.13↓20.08 multi-img 35.37 22.45 28.91↓16.69 52.60 22.38 56.80 27.96 65.24 16.05 53.57 47.96 43.69 51.39 43.76↓27.45 ColParse 49.11 48.39 48.75↑3.15 67.68 42.17 79.02 78.06 81.64 47.60 85.17 82.04 92.00 88.73 74.41↑3.20", "source": "marker_v2", "marker_block_id": "/page/17/Table/1"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0164", "section": "C.4. Main Results", "page_start": 19, "page_end": 19, "type": "TableGroup", "text": "Table 3. Performance comparison on ViDoRe-V2, ViDoSeek, and VisRAG benchmarks. For each model block, we bold the bestperforming optimization method in each column (except for the base result). The average scores for optimizations are shown with relative gains (↑/↓) compared to the base model. Method ViDoRe-V2 ViDoSeek VisRAG Bio-L Eco-R ESG-H ESG-M Avg. Doc Page Avg. Arxiv Chart InfoV MP-Doc Plot Slide Avg. VLM2Vec-V1-2B 6.88 14.15 12.25 20.54 13.46 56.40 67.73 62.07 41.68 58.21 70.79 42.74 23.83 74.07 51.89 s2m-add 10.40 8.59 4.68 3.46 6.78↓6.68 40.78 31.23 36.01↓26.06 28.94 44.79 59.76 30.86 9.67 59.59 38.94↓12.95 s2m-mul 10.00 8.40 5.09 3.51 6.75↓6.71 40.63 31.00 35.82↓26.25 28.85 44.64 60.48 30.55 9.83 59.52 38.98↓12.91 cl-t-c 14.86 14.49 4.14 4.48 9.49↓3.97 42.17 31.65 36.91↓25.16 13.56 18.52 25.66 13.11 1.41 31.91 17.36↓34.53 cl-t-m 13.34 4.60 8.78 5.15 7.97↓5.49 36.71 26.80 31.76↓30.31 8.29 18.27 37.74 11.85 1.13 32.46 18.29↓33.60 cl-s-c 16.38 10.03 4.93 4.64 9.00↓4.46 43.89 31.95 37.92↓24.15 13.03 27.01 28.61 18.82 1.54 31.31 20.05↓31.84 cl-s-m 13.46 6.50 6.89 3.74 7.65↓5.81 37.20 26.86 32.03↓30.04 8.21 22.04 37.37 12.65 1.15 33.31 19.12↓32.77 c-sem 20.32 11.01 11.86 10.44 13.41↓0.05 56.19 47.96 52.08↓9.99 16.53 44.39 43.88 24.77 3.86 54.32 31.29↓20.60 multi-img 15.59 30.33 14.90 29.55 5.70 33.21 5.95 38.33 10.54↓2.92 45.24 75.23 34.26 70.19 39.75↓22.32 30.20 38.18 41.53 60.09 60.40 69.44 29.53 48.29 8.44 18.83 60.10 76.95 38.37↓13.52 ColParse 32.86↑19.40 72.71↑10.64 51.96↑0.07 VLM2Vec-V1-7B s2m-add 4.93 13.74 6.82 11.27 9.19 54.26 77.39 65.83 52.58 69.83 71.43 52.86 34.24 73.22 59.03 s2m-mul 29.49 28.79 38.26 37.40 31.73 29.66 22.80 21.43 30.57↑21.38 29.32↑20.13 62.50 62.34 53.64 52.73 58.07↓7.76 57.54↓8.29 45.56 45.29 48.98 50.21 68.39 69.10 49.01 49.54 11.01 11.02 70.61 70.82 48.93↓10.10 49.33↓9.70 cl-t-c 17.52 18.22 11.68 8.82 14.06↑4.87 49.02 40.25 44.64↓21.19 15.12 21.70 36.61 19.07 2.22 37.89 22.10↓36.93 cl-t-m 22.08 14.23 24.74 15.61 19.17↑9.98 58.38 51.13 54.76↓11.07 18.75 30.49 58.40 22.24 3.90 52.32 31.02↓28.01 cl-s-c 19.74 22.03 14.36 12.74 17.22↑8.03 51.57 41.97 46.77↓19.06 15.24 17.35 36.59 24.24 1.79 35.59 21.80↓37.23 cl-s-m 23.08 16.08 20.53 14.47 18.54↑9.35 58.59 50.84 54.72↓11.11 18.76 31.65 58.91 24.90 3.10 52.63 31.66↓27.37 c-sem 29.88 31.92 30.24 20.34 28.10↑18.91 72.03 66.00 69.02↑3.19 39.77 57.94 63.94 41.41 14.39 69.66 47.85↓11.18 multi-img 30.91 39.93 23.80 17.07 27.93↑18.74 62.03 49.71 55.87↓9.96 45.32 44.84 64.95 40.71 10.55 67.95 45.72↓13.31 ColParse 42.63 42.89 50.55 42.86 44.73↑35.54 78.34 78.61 78.48↑12.65 54.43 70.30 69.27 58.49 33.46 77.98 60.66↑1.63 VLM2Vec-V2-2B 44.45 45.77 48.77 46.98 46.49 80.88 83.68 82.28 77.38 82.30 86.27 71.60 66.96 92.04 79.43 s2m-add 41.02 49.68 41.29 20.26 38.06↓8.43 69.01 64.83 66.92↓15.36 62.37 73.06 78.41 75.52 22.67 85.73 66.29↓13.14 s2m-mul 42.43 50.96 42.90 21.91 39.55↓6.94 71.66 66.41 69.04↓13.24 63.30 73.50 79.84 76.41 23.36 87.29 67.28↓12.15 cl-t-c 19.84 21.50 12.92 9.75 16.00↓30.49 48.48 49.99 49.24↓33.04 25.86 22.73 34.54 19.89 10.59 36.62 25.04↓54.39 cl-t-m 25.10 18.29 15.42 12.37 17.80↓28.69 57.69 59.96 58.83↓23.45 37.64 42.56 56.47 24.76 11.54 51.93 37.48↓41.95 cl-s-c 23.22 18.36 10.47 12.71 16.19↓30.30 49.73 51.97 50.85↓31.43 28.13 27.93 34.99 25.08 10.77 40.66 27.93↓51.50 cl-s-m 27.68 18.68 17.65 9.93 18.49↓28.00 57.31 61.50 59.41↓22.87 37.55 43.07 55.65 26.56 11.74 52.27 37.81↓41.62 c-sem multi-img 36.74 31.11 20.49 21.25 27.40↓19.09 71.91 71.47 71.69↓10.59 53.17 61.34 67.94 53.96 23.78 74.81 55.83↓23.60 ColParse 30.01 50.06 45.94 53.76 23.58 57.41 17.56 46.40 29.27↓17.22 63.17 80.94 50.95 83.87 57.06↓25.22 62.83 77.18 61.38 78.05 75.92 84.37 54.34 78.07 24.62 58.74 75.43 91.95 59.09↓20.34 51.91↑5.42 82.41↑0.13 78.06↓1.37 LamRA-Ret s2m-add 10.75 12.36 9.65 21.34 6.32 18.02 11.18 21.30 9.48 60.17 55.39 28.81 27.91 44.49 11.17 2.75 63.50 33.31 59.78 44.51 33.57 28.61 29.42 4.86 57.59 49.08 42.51 s2m-mul 9.91 17.36 21.00 20.88 18.26↑8.78 17.29↑7.81 52.53 25.95 41.65↓2.84 39.24↓5.25 2.75 33.53 43.10 28.28 5.19 49.10 27.19↓15.32 26.99↓15.52 cl-t-c 6.69 7.44 5.85 1.05 5.26↓4.22 39.52 30.63 35.08↓9.41 4.26 19.69 21.94 9.90 1.53 26.54 13.98↓28.53 cl-t-m 7.75 10.84 8.20 3.97 7.69↓1.79 44.87 38.69 41.78↓2.71 5.46 22.85 38.15 17.45 2.13 35.03 20.18↓22.33 cl-s-c 7.57 6.57 6.90 1.97 5.75↓3.73 38.41 30.98 34.70↓9.79 5.19 17.67 21.50 11.02 1.52 25.81 13.79↓28.72 cl-s-m 8.55 10.98 7.70 4.07 7.83↓1.65 45.50 38.86 42.18↓2.31 5.45 23.56 37.71 17.17 2.24 35.58 20.29↓22.22 c-sem 8.67 9.41 8.69 4.43 7.80↓1.68 53.12 38.51 45.82↑1.33 9.07 40.44 41.90 18.39 3.92 44.45 26.36↓16.15 multi-img 3.36 10.48 13.43 6.24 8.38↓1.10 28.18 10.96 19.57↓24.92 1.22 23.90 35.05 16.83 5.71 18.51 16.87↓25.64 ColParse 15.81 17.65 23.75 25.48 20.67↑11.19 58.55 39.99 49.27↑4.78 5.91 60.76 54.91 38.77 25.27 62.34 41.33↓1.18 GME-2B 54.25 50.65 59.44 49.15 53.37 81.44 79.62 80.53 81.37 81.70 91.31 85.03 63.81 93.60 82.80 s2m-add 48.33 46.46 40.98 30.02 41.45↓11.92 82.20 63.50 72.85↓7.68 70.43 77.10 85.04 80.94 20.83 89.96 70.72↓12.08 s2m-mul 51.15 49.41 46.69 28.65 43.98↓9.39 83.64 67.99 75.82↓4.71 74.25 77.91 88.07 84.75 22.40 92.99 73.40↓9.40 cl-t-c 17.91↓64.89 17.04 18.20 2.88 5.79 10.98↓42.39 34.72 28.03 31.38↓49.15 14.40 18.17 30.07 14.03 4.30 26.48 cl-t-m 18.60 12.72 12.39 7.72 12.86↓40.51 45.17 40.84 43.01↓37.52 25.41 31.73 44.53 16.56 2.77 40.18 cl-s-c 17.47 19.49 10.88 5.99 13.46↓39.91 34.42 29.29 31.86↓48.67 15.10 14.91 29.78 15.49 3.92 28.86 cl-s-m 21.66 11.82 14.37 5.63 13.37↓40.00 42.98 38.99 40.99↓39.54 24.76 30.32 45.48 17.41 2.60 39.35 c-sem 24.51 22.71 22.72 18.57 22.13↓31.24 65.58 53.29 59.44↓21.09 38.80 50.88 52.09 35.27 18.64 63.25 multi-img 32.00 55.14 41.67 52.08 25.86 56.24 19.83 51.36 29.84↓23.53 73.00 83.93 47.58 80.16 60.29↓20.24 67.87 81.65 67.77 83.87 81.98 91.06 58.54 85.89 22.85 63.17 77.89 93.76 26.86↓55.94 18.01↓64.79 26.65↓56.15 43.16↓39.64 62.82↓19.98 53.71↑0.34 82.05↑1.52 83.23↑0.43 53.66 54.34 65.38 54.32 56.93 83.21 84.18 83.70 87.20 82.32 92.92 88.89 63.36 94.81 84.92 50.40 47.49 47.63 29.02 43.64↓13.29 83.22 67.67 75.45↓8.25 73.94 79.30 87.84 84.63 22.35 92.79 73.48↓11.44 48.04 10.12 47.85 12.81 39.12 8.12 29.60 4.38 41.15↓15.78 82.30 29.66 63.00 23.82 72.65↓11.05 70.46 3.79 77.68 11.21 84.96 21.97 80.38 11.20 21.12 2.44 90.08 18.30 70.78↓14.14 9.38 10.36 6.42 4.63 8.86↓48.07 7.70↓49.23 35.86 32.84 26.74↓56.96 34.35↓49.35 3.56 21.05 35.14 12.97 1.45 24.43 11.49↓73.43 16.43↓68.49 8.58 16.36 5.41 5.20 8.89↓48.04 28.62 23.40 26.01↓57.69 5.46 11.14 23.14 11.35 1.54 21.57 10.84 9.96 6.62 2.80 7.56↓49.37 34.06 29.69 31.88↓51.82 3.93 22.24 33.74 13.14 1.42 23.76 12.37↓72.55 16.37↓68.55 16.30 13.79 8.34 8.42 11.71↓45.22 49.08 35.41 42.25↓41.45 7.40 29.57 32.02 15.54 5.76 35.26 29.32 39.47 30.32 16.76 28.97↓27.96 75.23 52.91 64.07↓19.63 72.68 70.18 85.20 63.27 23.75 81.21 62.40 59.73 68.43 57.90 62.12↑5.19 84.30 83.94 84.12↑0.42 87.12 83.94 92.84 89.62 62.53 94.97 9.50 15.78 15.51 19.03 14.96 54.24 77.98 66.11 61.19 65.48 76.76 59.65 45.25 77.50 20.93↓63.99 66.05↓18.87 85.17↑0.25 64.31 30.82 38.23 23.31 20.59 28.24↑13.28 58.58 50.31 54.45↓11.66 45.94 55.05 67.28 55.13 12.43 72.62 51.41↓12.90 29.67 34.92 21.62 19.02 26.31↑11.35 56.92 48.58 52.75↓13.36 45.21 55.60 66.84 53.32 12.49 71.71 50.86↓13.45 24.01 15.49 14.30 12.37 16.54↑1.58 46.27 38.93 42.60↓23.51 18.00 24.42 29.87 16.38 4.20 33.42 21.05↓43.26 21.67 10.06 13.15 9.43 13.58↓1.38 42.79 35.18 38.99↓27.12 17.18 27.33 42.30 14.99 5.25 44.29 25.22↓39.09 ColParse GME-7B s2m-add s2m-mul cl-t-c cl-t-m cl-s-c cl-s-m c-sem multi-img ColParse UniME-V2-2B s2m-add s2m-mul cl-t-c cl-t-m cl-s-c cl-s-m 23.56 22.75 15.34 13.02 12.00 12.48 12.62 10.46 15.88↑0.92 14.68↓0.28 48.96 43.51 39.41 37.60 44.19↓21.92 40.56↓25.55 19.80 17.20 24.93 29.53 32.56 42.92 21.55 16.67 4.33 5.00 38.90 47.36 23.68↓40.63 26.45↓37.86", "source": "marker_v2", "marker_block_id": "/page/18/TableGroup/2386"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0165", "section": "C.4. Main Results", "page_start": 19, "page_end": 19, "type": "Text", "text": "Continued on next page", "source": "marker_v2", "marker_block_id": "/page/18/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0166", "section": "C.4. Main Results", "page_start": 20, "page_end": 20, "type": "Caption", "text": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "source": "marker_v2", "marker_block_id": "/page/19/Caption/0"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0167", "section": "C.4. Main Results", "page_start": 20, "page_end": 20, "type": "Table", "text": "Table 3 – Continued from previous page Method ViDoRe-V2 ViDoSeek VisRAG Bio-L Eco-R ESG-H ESG-M Avg. Doc Page Avg. Arxiv Chart InfoV MP-Doc Plot Slide Avg. c-sem 35.10 27.35 21.50 28.81 28.19↑13.23 64.50 60.58 62.54↓3.57 32.65 51.23 57.19 40.32 14.91 64.66 43.49↓20.82 multi-img 31.08 38.51 16.09 18.02 25.93↑10.97 56.06 43.06 49.56↓16.55 46.47 46.80 65.91 42.99 11.82 63.81 46.30↓18.01 ColParse 44.52 45.86 48.15 51.48 47.50↑32.54 80.50 79.46 79.98↑13.87 58.51 63.17 74.52 66.58 35.57 80.21 63.09↓1.22 UniME-V2-7B 26.77 23.69 24.68 31.17 26.58 78.25 82.25 80.25 60.60 79.43 80.61 64.94 45.35 82.17 68.85 s2m-add 41.25 54.38 36.91 24.88 39.36↑12.78 71.49 64.32 67.91↓12.34 51.03 64.77 75.82 67.21 15.38 81.12 59.22↓9.63 s2m-mul 41.49 53.84 35.49 23.31 38.53↑11.95 71.69 64.24 67.97↓12.28 50.89 65.45 76.45 67.54 15.32 81.68 59.56↓9.29 cl-t-c 22.76 16.17 8.68 14.14 15.44↓11.14 49.38 45.31 47.35↓32.90 19.44 24.63 35.22 17.75 4.44 36.13 22.94↓45.91 cl-t-m 27.93 14.34 19.03 16.99 19.57↓7.01 55.94 52.84 54.39↓25.86 27.05 38.47 62.09 20.81 7.64 55.56 35.27↓33.58 cl-s-c 24.75 19.62 12.41 11.69 17.12↓9.46 52.10 48.38 50.24↓30.01 20.58 21.04 34.39 23.49 5.53 35.42 23.41↓45.44 cl-s-m 30.78 20.40 18.14 15.62 21.24↓5.34 56.69 53.63 55.16↓25.09 28.73 37.84 62.65 23.01 6.78 57.14 36.03↓32.82 c-sem 37.96 29.51 27.33 21.95 29.19↑2.61 71.01 68.82 69.92↓10.33 51.40 65.80 72.53 54.42 21.98 76.81 57.16↓11.69 multi-img 35.99 45.82 24.74 14.71 30.32↑3.74 65.89 53.21 59.55↓20.70 51.79 62.63 73.67 48.06 17.25 73.00 54.40↓14.45 ColParse 54.95 50.07 54.92 50.14 52.52↑25.94 81.16 83.32 82.24↑1.99 61.90 77.80 78.41 71.89 44.43 84.68 69.85↑1.00 B3-2B 38.41 31.80 45.23 45.10 40.14 78.56 74.87 76.72 51.75 66.86 70.43 45.73 36.69 77.81 58.21 s2m-add 32.14 40.97 22.62 19.74 28.87↓11.27 67.40 56.61 62.01↓14.71 43.01 51.27 68.01 50.92 13.40 73.78 50.07↓8.14 s2m-mul 30.87 39.31 21.36 19.69 27.81↓12.33 66.73 56.12 61.43↓15.29 42.74 49.75 68.21 51.27 13.31 73.79 49.85↓8.36 cl-t-c 11.90 13.53 9.61 5.70 10.19↓29.95 44.40 41.39 42.90↓33.82 14.13 37.31 41.30 17.72 3.32 44.50 26.38↓31.83 cl-t-m 15.16 14.20 13.12 6.61 12.27↓27.87 44.92 44.57 44.75↓31.97 14.12 32.08 42.61 19.91 1.06 43.03 25.47↓32.74 cl-s-c 16.03 20.23 10.10 8.92 13.82↓26.32 48.54 47.73 48.14↓28.58 13.09 36.05 44.30 24.60 2.76 48.77 28.26↓29.95 cl-s-m 16.43 16.68 14.29 9.21 14.15↓25.99 48.08 47.08 47.58↓29.14 13.51 35.91 44.17 25.47 1.20 45.45 27.62↓30.59 c-sem 25.26 25.23 18.89 18.29 21.92↓18.22 60.79 59.97 60.38↓16.34 32.30 54.90 61.44 39.77 10.80 61.17 43.40↓14.81 multi-img 21.83 23.38 7.29 11.37 15.97↓24.17 52.38 37.41 44.90↓31.82 38.52 41.26 53.67 24.33 7.97 37.56 33.89↓24.32 ColParse 45.03 39.23 48.20 49.30 45.44↑5.30 79.98 80.79 80.39↑3.67 51.00 62.94 67.06 53.86 32.02 80.01 57.82↓0.39 B3-7B 47.29 44.81 50.84 48.05 47.75 82.07 82.26 82.17 65.83 76.77 84.54 68.55 52.86 85.75 72.38 s2m-add 44.47 48.61 35.86 28.82 39.44↓8.31 75.93 65.24 70.59↓11.58 56.95 69.80 78.61 69.19 19.74 85.71 63.33↓9.05 s2m-mul 44.79 49.78 35.45 28.68 39.68↓8.07 76.17 64.83 70.50↓11.67 57.17 70.52 79.24 70.67 20.17 85.50 63.88↓8.50 cl-t-c 18.50 22.50 20.46 15.85 19.33↓28.42 66.30 63.26 64.78↓17.39 18.07 43.88 52.94 23.57 5.85 58.35 33.78↓38.60 cl-t-m 22.98 22.71 22.21 14.16 20.52↓27.23 67.41 67.43 67.42↓14.75 23.92 43.10 62.54 28.35 7.43 61.50 37.81↓34.57 cl-s-c 20.76 26.27 28.71 22.95 24.67↓23.08 68.09 65.59 66.84↓15.33 16.71 47.07 53.44 31.71 7.56 60.01 36.08↓36.30 cl-s-m 22.17 26.43 26.15 19.85 23.65↓24.10 67.85 64.05 65.95↓16.22 19.23 43.79 53.45 35.24 8.50 61.19 36.90↓35.48 c-sem 32.31 30.23 33.27 18.99 28.70↓19.05 75.30 71.83 73.57↓8.60 46.32 67.80 75.28 54.93 24.73 80.26 58.22↓14.16 multi-img 26.48 33.14 11.25 11.91 20.70↓27.05 64.84 47.83 56.34↓25.83 51.50 46.37 63.64 34.80 11.94 44.09 42.06↓30.32 ColParse 53.72 49.50 52.40 50.15 51.44↑3.69 83.15 83.60 83.38↑1.21 65.97 75.92 80.19 73.87 48.05 86.13 71.69↓0.69", "source": "marker_v2", "marker_block_id": "/page/19/Table/1"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0168", "section": "C.5.1. ALGORITHM WORKFLOW", "page_start": 21, "page_end": 21, "type": "Text", "text": "Algorithm 4 presents a unified offline indexing framework for the three variants. After shared layout parsing and dualstream encoding (Stages 1–2), Stage 3 diverges based on the specified variant type: single2multi retains raw subimage vectors; type_cluster aggregates vectors by semantic content types via averaging; global_inclusion appends the full-document global vector to the local set.", "source": "marker_v2", "marker_block_id": "/page/20/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0169", "section": "Algorithm 4 Integrated Offline Indexing for ColParse Vari", "page_start": 21, "page_end": 21, "type": "Code", "text": "Input: A document image d \\in \\mathbb{R}^{H \\times W \\times 3}; A document parser model \\Psi_{parse}; A single-vector encoder \\Phi_{\\mathrm{enc}}: \\mathbb{R}^{H' \\times W' \\times 3} \\rightarrow \\mathbb{R}^D: Mode M \\in \\{s2m, s2m-t-c, s2m-g-i\\} Output: A multi-vector representation D<sub>variant</sub>", "source": "marker_v2", "marker_block_id": "/page/20/Code/5"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0170", "section": "1119 1120 /* Stage 1: Layout-Informed Document Parsing", "page_start": 21, "page_end": 21, "type": "Code", "text": "[\\{b_j, c_j\\}]_{j=1}^k \\leftarrow \\Psi_{\\text{parse}}(d); 1121 // Get k bboxes and content 1122 types 1123 S_d \\leftarrow \\emptyset for j \\leftarrow 1 to k do 1124 s_j \\leftarrow \\text{Crop}(d, b_j); // Extract sub-image for each layout 1125", "source": "marker_v2", "marker_block_id": "/page/20/Code/7"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0171", "section": "1119 1120 /* Stage 1: Layout-Informed Document Parsing", "page_start": 21, "page_end": 21, "type": "Text", "text": "component 1126 \\mathcal{S}_d \\leftarrow \\mathcal{S}_d \\cup \\{s_i\\}", "source": "marker_v2", "marker_block_id": "/page/20/Text/8"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0172", "section": "1128 /* Stage 2: Regional Encoding 1129", "page_start": 21, "page_end": 21, "type": "Code", "text": "\\mathbf{D}_{\\text{local}} \\leftarrow \\emptyset for each sub-image s_j \\in \\mathcal{S}_d do 1130 \\mathbf{v}_{\\text{local}}^{(j)} \\leftarrow \\Phi_{\\text{enc}}(s_j); // Independent regional encoding 1131 \\mathbf{D}_{\\mathrm{local}} \\leftarrow \\mathbf{D}_{\\mathrm{local}} \\cup \\{\\mathbf{v}_{\\mathrm{local}}^{(j)}\\} 1132 1133 end", "source": "marker_v2", "marker_block_id": "/page/20/Code/11"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0173", "section": "1128 /* Stage 2: Regional Encoding 1129", "page_start": 21, "page_end": 21, "type": "Text", "text": "1134 /* Stage 3: Variant-specific Representation Construction */ 1135", "source": "marker_v2", "marker_block_id": "/page/20/Text/12"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0174", "section": "1128 /* Stage 2: Regional Encoding 1129", "page_start": 21, "page_end": 21, "type": "Code", "text": "if M = s2m then 1136 \\mathbf{D}_{\\text{variant}} \\leftarrow \\mathbf{D}_{\\text{local}}; // Standard layout-decomposed set", "source": "marker_v2", "marker_block_id": "/page/20/Code/13"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0175", "section": "1128 /* Stage 2: Regional Encoding 1129", "page_start": 21, "page_end": 21, "type": "Code", "text": "else if M = s2m-t-c then \\mathbf{D}_{\\text{variant}} \\leftarrow \\emptyset \\ \\mathcal{T} \\leftarrow \\text{Unique}(\\{c_1, \\dots, c_k\\}); \\ \\ // \\text{Identify}", "source": "marker_v2", "marker_block_id": "/page/20/Code/14"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0176", "section": "1128 /* Stage 2: Regional Encoding 1129", "page_start": 21, "page_end": 21, "type": "Code", "text": "1138 1139 unique content types 1140 for each type t \\in \\mathcal{T} do \\mathbf{v}_{\\text{avg}}^{(t)} \\leftarrow \\text{Mean}(\\{\\mathbf{v}_{\\text{local}}^{(j)} \\mid c_j = t\\}); average by type 1141 // Cluster and 1142 1143 \\mathbf{D}_{\\text{variant}} \\leftarrow \\mathbf{D}_{\\text{variant}} \\cup \\{\\mathbf{v}_{\\text{avg}}^{(t)}\\} 1144 end 1145", "source": "marker_v2", "marker_block_id": "/page/20/Code/15"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0177", "section": "1128 /* Stage 2: Regional Encoding 1129", "page_start": 21, "page_end": 21, "type": "Code", "text": "else if M = s2m-q-i then", "source": "marker_v2", "marker_block_id": "/page/20/Code/17"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0178", "section": "1128 /* Stage 2: Regional Encoding 1129", "page_start": 21, "page_end": 21, "type": "Code", "text": "\\mathbf{v}_{\\text{global}} \\leftarrow \\Phi_{\\text{enc}}(d); // Encode original page for context \\mathbf{D}_{variant} \\leftarrow \\mathbf{D}_{local} \\cup \\{\\mathbf{v}_{global}\\}\\;;\\;\\; // Append global vector to the set", "source": "marker_v2", "marker_block_id": "/page/20/Code/18"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0179", "section": "1128 /* Stage 2: Regional Encoding 1129", "page_start": 21, "page_end": 21, "type": "Text", "text": "1151 end", "source": "marker_v2", "marker_block_id": "/page/20/Text/19"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0180", "section": "1128 /* Stage 2: Regional Encoding 1129", "page_start": 21, "page_end": 21, "type": "Text", "text": "return D<sub>variant</sub>", "source": "marker_v2", "marker_block_id": "/page/20/Text/20"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0181", "section": "C.5.2. MORE ANALYSIS", "page_start": 21, "page_end": 21, "type": "Text", "text": "Due to the limited space of main text, we leave the radar plots of performance comparison between ColParse and its variants in Figure 9.", "source": "marker_v2", "marker_block_id": "/page/20/Text/22"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0182", "section": "C.5.2. MORE ANALYSIS", "page_start": 21, "page_end": 21, "type": "Text", "text": "The introduction of global page-level context is indispensable for resolving semantic ambiguities within isolated layout components. Quantitative results in Table 4 demonstrate that adding global context—even via simple inclusion (s2m-g-i)—dramatically elevates performance over the local-only single2multi baseline, lifting the VLM2Vec-V1-2B score on ViDoRe-V1 from 34.39 to 49.93. Figure 5 highlights that this gap is most pronounced in benchmarks requiring holistic understanding, where local sub-images like tables or charts often lack the necessary contextual headers found elsewhere on the page. We hypothesize that the global vector acts as a \"semantic anchor\" that provides the overarching topic of the document, which is essential for the late-interaction mechanism to accurately align specific query tokens with relevant sub-regions.", "source": "marker_v2", "marker_block_id": "/page/20/Text/23"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0183", "section": "C.5.2. MORE ANALYSIS", "page_start": 21, "page_end": 21, "type": "Text", "text": "Maintaining the individual spatial and semantic integrity of layout components is superior to heuristic type-level clustering. As evidenced by the performance trends in Table 4 and the comparative bars in Figure 5, the s2m-t-c variant typically results in a performance regression compared to the standard single2multi, such as the 1.56point drop for VLM2Vec-V1-2B on ViDoRe-V1 benchmark. This trend is echoed across the radar charts in Figure 9, where the type-clustered variants consistently exhibit the narrowest performance profiles. This indicates that spatial locality is a vital semantic carrier in visual documents; by collapsing multiple distinct components into a single typelevel average, the model loses the fine-grained resolution required for the MaxSim operator to distinguish between specific relevant and irrelevant regions of the same type.", "source": "marker_v2", "marker_block_id": "/page/20/Text/24"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0184", "section": "C.5.2. MORE ANALYSIS", "page_start": 22, "page_end": 22, "type": "FigureGroup", "text": "Figure 9. The performance comparison (evaluated by nDCG@5) between ColParse and its variants on five VDR benchmarks across ten mainstream single-vector multimodal retrieval models. Refer to Table 4 and Table 5 for detailed result records due to the space limit.", "source": "marker_v2", "marker_block_id": "/page/21/FigureGroup/194"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0185", "section": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "page_start": 23, "page_end": 23, "type": "Caption", "text": "Table 4. Ablation study on MMLongBench and ViDoRe-V1 benchmarks. For each model block, we bold the best-performing method in each column (except for the base result). The average scores are shown with relative gains (↑/↓) compared to the base model.", "source": "marker_v2", "marker_block_id": "/page/22/Caption/1"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0186", "section": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "page_start": 23, "page_end": 23, "type": "Table", "text": "Method MMLongBench ViDoRe-V1 Doc Page Avg. Arxiv DocV InfoV Shift TabF TatD S-AI S-En S-HC S-Gov Avg. VLM2Vec-V1-2B 25.62 26.23 25.93 17.80 13.98 39.41 9.18 36.32 10.56 16.39 15.96 23.56 24.11 20.73 single2multi 25.19 19.59 22.39↓3.54 38.78 16.38 56.94 12.48 41.15 9.98 38.16 46.42 42.66 40.94 34.39↑13.66 s2m-t-c 25.37 18.43 21.90↓4.03 34.95 15.04 55.09 15.31 40.07 9.78 32.83 43.07 44.84 37.31 32.83↑12.10 s2m-g-i 31.57 26.55 29.06↑3.13 48.84 23.86 67.91 39.85 62.15 14.07 55.74 64.41 61.67 60.83 49.93↑29.20 ColParse 34.31 29.83 32.07↑6.14 47.66 28.12 69.23 47.11 57.05 20.43 62.24 63.77 65.51 62.54 52.37↑31.64 VLM2Vec-V1-7B 23.85 37.63 30.74 28.07 17.93 44.47 2.06 16.78 5.86 17.93 25.04 28.90 14.59 20.16 single2multi 35.75 29.40 32.58↑1.84 53.54 26.07 61.73 41.43 67.64 22.37 72.54 59.56 73.52 64.03 54.24↑34.08 s2m-t-c 35.62 28.64 32.13↑1.39 50.56 24.08 59.42 35.90 64.84 18.29 65.39 58.49 72.13 63.75 51.29↑31.13 s2m-g-i 37.95 34.08 36.02↑5.28 62.29 30.49 67.54 41.74 78.87 22.99 74.85 62.78 77.07 64.14 58.28↑38.12 ColParse 43.34 40.58 41.96↑11.22 60.47 34.42 70.39 53.67 77.12 31.33 74.81 69.64 80.79 75.89 62.85↑42.69 VLM2Vec-V2-2B 48.55 50.34 49.45 78.98 38.51 82.21 64.57 87.64 44.68 85.06 82.99 89.89 87.08 74.16 single2multi 45.89 41.04 43.47↓5.98 69.32 38.09 77.88 56.25 72.60 52.69 88.14 83.80 86.99 84.09 70.99↓3.17 s2m-t-c 44.99 40.97 42.98↓6.47 67.44 34.25 74.79 48.36 71.32 43.73 82.22 81.18 84.78 76.42 66.45↓7.71 s2m-g-i 48.03 46.30 47.17↓2.28 79.83 46.57 82.19 66.78 87.66 52.85 93.91 86.02 91.14 87.46 77.44↑3.28 ColParse 49.49 50.53 50.01↑0.56 80.17 46.33 83.53 72.76 86.74 52.40 91.36 85.83 95.47 89.52 78.41↑4.25 LamRA-Ret 19.78 13.24 16.51 29.31 19.56 63.00 15.83 51.44 7.70 21.10 29.81 37.18 31.95 30.69", "source": "marker_v2", "marker_block_id": "/page/22/Table/2"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0187", "section": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "page_start": 23, "page_end": 23, "type": "Text", "text": "Continued on next page", "source": "marker_v2", "marker_block_id": "/page/22/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0188", "section": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "page_start": 24, "page_end": 24, "type": "Caption", "text": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "source": "marker_v2", "marker_block_id": "/page/23/Caption/0"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0189", "section": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "page_start": 24, "page_end": 24, "type": "Table", "text": "Table 4 – Continued from previous page Method MMLongBench ViDoRe-V1 Doc Page Avg. Arxiv DocV InfoV Shift TabF TatD S-AI S-En S-HC S-Gov Avg. single2multi 33.04 18.56 25.80↑9.29 6.06 13.60 37.83 18.90 25.46 19.02 23.99 28.27 33.31 28.03 23.45↓7.24 s2m-t-c 31.74 17.49 24.62↑8.11 6.13 10.90 35.44 16.81 28.56 15.77 15.45 27.59 24.45 19.65 20.08↓10.61 s2m-g-i 33.16 18.48 25.82↑9.31 6.06 15.79 37.71 18.96 38.78 19.19 26.00 28.27 31.68 29.03 25.15↓5.54 ColParse 30.74 19.50 25.12↑8.61 17.27 20.61 58.35 21.39 39.69 13.34 25.57 35.27 43.21 26.29 30.10↓0.59 GME-2B 52.07 53.14 52.61 82.59 56.46 88.97 89.72 93.20 70.33 98.49 92.15 98.15 95.65 86.57 single2multi 50.56 45.68 48.12↓4.49 72.46 39.99 79.82 70.54 80.91 68.07 88.91 91.11 88.85 89.86 77.05↓9.52 s2m-t-c 49.51 44.73 47.12↓5.49 70.90 37.34 77.92 67.59 79.85 59.12 86.13 86.87 87.23 83.35 73.63↓12.94 s2m-g-i 51.81 48.49 50.15↓2.46 80.89 46.62 83.93 75.78 92.12 68.33 93.48 92.60 93.36 90.49 81.76↓4.81 ColParse 53.06 54.24 53.65↑1.04 82.39 54.11 88.93 88.51 92.33 70.65 97.75 92.30 97.91 96.10 86.10↓0.47 GME-7B 54.01 55.80 54.91 87.59 56.05 91.96 94.25 93.72 76.26 99.63 95.45 99.63 99.06 89.36 single2multi 53.55 48.17 50.86↓4.05 75.57 45.26 83.21 77.90 86.05 73.97 94.72 92.04 95.19 93.18 81.71↓7.65 s2m-t-c 52.97 48.05 50.51↓4.40 75.21 44.11 82.69 78.44 86.17 66.14 91.53 89.86 91.38 88.60 79.41↓9.95 s2m-g-i 54.32 51.14 52.73↓2.18 84.29 52.98 87.36 82.08 94.79 74.48 96.65 91.10 96.92 93.70 85.44↓3.92 ColParse 54.96 56.51 55.74↑0.83 87.35 57.91 90.76 95.35 95.44 75.92 99.63 94.67 99.63 98.89 89.56↑0.20 UniME-V2-2B 18.52 40.10 29.31 36.52 12.43 42.41 14.09 51.11 7.39 20.23 32.96 24.21 19.25 26.06 single2multi 38.47 33.34 35.91↑6.60 52.64 27.01 68.58 49.89 61.20 38.92 77.36 73.65 77.92 79.30 60.65↑34.59 s2m-t-c 35.13 30.33 32.73↑3.42 48.29 21.24 62.26 43.83 58.58 27.37 70.42 66.86 74.85 67.57 54.13↑28.07 s2m-g-i 41.90 39.97 40.94↑11.63 64.23 33.32 74.84 67.10 78.40 39.00 82.16 79.38 88.43 83.62 69.05↑42.99 ColParse 44.22 44.19 44.21↑14.90 62.39 37.69 73.33 71.19 77.45 38.83 82.50 75.80 89.35 85.84 69.44↑43.38 UniME-V2-7B 33.19 45.72 39.46 63.23 24.91 65.25 11.16 41.54 14.18 41.89 40.56 57.44 42.78 40.29 single2multi 39.83 39.39 39.61↑0.15 57.08 32.54 71.38 64.54 71.82 49.00 84.00 84.26 91.18 87.49 69.33↑29.04 s2m-t-c 38.66 36.94 37.80↓1.66 54.98 25.88 65.87 56.63 70.76 41.65 80.00 77.76 86.67 77.88 63.81↑23.52 s2m-g-i 41.35 43.08 42.22↑2.76 64.50 34.55 76.81 66.35 81.95 48.84 85.63 85.38 91.83 87.90 72.37↑32.08 ColParse 45.90 48.26 47.08↑7.62 64.78 37.43 78.51 73.56 81.47 43.69 89.32 82.68 92.74 88.13 73.23↑32.94 B3-2B 37.10 32.07 34.59 57.00 29.38 68.09 48.31 71.55 18.09 74.13 64.64 75.44 63.13 56.98 single2multi 36.47 29.20 32.84↓1.75 46.63 20.13 60.31 51.58 56.45 36.55 70.40 69.80 73.83 66.85 55.25↓1.73 s2m-t-c 35.15 28.33 31.74↓2.85 43.88 18.92 54.65 41.90 54.30 26.86 67.36 61.92 72.05 60.66 50.25↓6.73 s2m-g-i 38.87 33.51 36.19↑1.60 56.06 22.52 64.41 59.98 73.09 35.98 78.34 70.48 80.13 67.17 60.82↑3.84 ColParse 42.06 37.60 39.83↑5.24 56.47 30.91 66.69 67.42 69.33 29.42 79.88 72.67 83.24 71.41 62.74↑5.76 B3-7B 46.09 45.10 45.60 68.95 43.38 79.86 66.56 84.12 37.06 81.01 81.25 88.57 81.30 71.21 single2multi 44.43 40.79 42.61↓2.99 58.94 31.91 71.96 68.80 73.69 53.07 87.14 81.75 88.38 84.76 70.04↓1.17 s2m-t-c 43.32 39.84 41.58↓4.02 56.41 29.10 68.99 53.11 70.75 44.35 83.69 77.98 84.69 78.56 64.76↓6.45 s2m-g-i 45.62 43.88 44.75↓0.85 67.23 35.81 76.70 70.48 84.62 53.26 87.14 83.19 91.86 84.44 73.47↑2.26 ColParse 49.11 48.39 48.75↑3.15 67.68 42.17 79.02 78.06 81.64 47.60 85.17 82.04 92.00 88.73 74.41↑3.20", "source": "marker_v2", "marker_block_id": "/page/23/Table/1"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0190", "section": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "page_start": 25, "page_end": 25, "type": "Text", "text": "Table 5. Ablation study on ViDoRe-V2, ViDoSeek, and VisRAG benchmarks. For each model block, we bold the best-performing method in each column (except for the base result). The average scores are shown with relative gains (↑/↓) compared to the base model.", "source": "marker_v2", "marker_block_id": "/page/24/Text/1"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0191", "section": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations", "page_start": 25, "page_end": 25, "type": "Table", "text": "Method ViDoRe-V2 ViDoSeek VisRAG Bio-L Eco-R ESG-H ESG-M Avg. Doc Page Avg. Arxiv Chart InfoV MP-Doc Plot Slide Avg. VLM2Vec-V1-2B 6.88 14.15 12.25 20.54 13.46 56.40 67.73 62.07 41.68 58.21 70.79 42.74 23.83 74.07 51.89 single2multi 17.15 18.19 10.24 7.44 13.26↓0.20 50.51 41.52 46.02↓16.05 30.17 42.91 61.28 37.60 6.80 65.76 40.75↓11.14 s2m-t-c 13.46 18.31 11.58 8.85 13.05↓0.41 50.22 40.75 45.49↓16.58 27.87 41.64 59.37 34.82 6.94 62.76 38.90↓12.99 s2m-g-i 28.49 26.29 29.18 39.15 30.78↑17.32 73.38 68.14 70.76↑8.69 40.39 52.54 70.43 47.82 8.96 76.40 49.42↓2.47 ColParse 30.33 29.55 33.21 38.33 32.86↑19.40 75.23 70.19 72.71↑10.64 38.18 60.09 69.44 48.29 18.83 76.95 51.96↑0.07 VLM2Vec-V1-7B 4.93 13.74 6.82 11.27 9.19 54.26 77.39 65.83 52.58 69.83 71.43 52.86 34.24 73.22 59.03 single2multi 34.67 37.55 33.77 26.91 33.23↑24.04 66.88 58.17 62.53↓3.30 45.06 50.30 63.65 50.23 10.83 74.00 49.01↓10.02 s2m-t-c 31.17 41.18 26.15 21.07 29.89↑20.70 64.53 54.75 59.64↓6.19 43.98 52.75 61.47 45.75 10.49 71.33 47.63↓11.40 s2m-g-i 41.08 37.99 40.11 36.54 38.93↑29.74 75.46 75.28 75.37↑9.54 53.97 64.14 65.37 55.41 27.98 77.07 57.32↓1.71 ColParse 42.63 42.89 50.55 42.86 44.73↑35.54 78.34 78.61 78.48↑12.65 54.43 70.30 69.27 58.49 33.46 77.98 60.66↑1.63 VLM2Vec-V2-2B 44.45 45.77 48.77 46.98 46.49 80.88 83.68 82.28 77.38 82.30 86.27 71.60 66.96 92.04 79.43 single2multi 42.12 51.08 41.33 24.84 39.84↓6.65 73.81 67.71 70.76↓11.52 65.36 69.18 79.34 73.19 19.59 86.77 65.57↓13.86 s2m-t-c 41.81 51.83 37.07 25.78 39.12↓7.37 73.69 66.16 69.93↓12.35 64.18 66.49 77.61 64.32 19.58 85.70 62.98↓16.45 s2m-g-i 44.34 51.99 40.57 34.53 42.86↓3.63 79.23 81.01 80.12↓2.16 77.03 74.83 82.98 78.17 54.13 90.93 76.35↓3.08 ColParse 50.06 53.76 57.41 46.40 51.91↑5.42 80.94 83.87 82.41↑0.13 77.18 78.05 84.37 78.07 58.74 91.95 78.06↓1.37 LamRA-Ret 10.75 9.65 6.32 11.18 9.48 60.17 28.81 44.49 11.17 63.50 59.78 33.57 29.42 57.59 42.51 single2multi 11.11 26.32 20.54 23.76 20.43↑10.95 53.77 30.82 42.30↓2.19 1.94 25.59 30.56 27.49 3.95 44.28 22.30↓20.21 s2m-t-c 9.36 15.57 14.49 17.11 14.13↑4.65 49.02 29.38 39.20↓5.29 2.03 22.46 31.39 21.10 3.96 42.90 20.64↓21.87 s2m-g-i 11.11 26.56 18.78 21.76 19.55↑10.07 53.74 31.77 42.76↓1.73 1.91 28.55 31.15 32.12 25.09 44.22 27.17↓15.34 ColParse 15.81 17.65 23.75 25.48 20.67↑11.19 58.55 39.99 49.27↑4.78 5.91 60.76 54.91 38.77 25.27 62.34 41.33↓1.18 GME-2B 54.25 50.65 59.44 49.15 53.37 81.44 79.62 80.53 81.37 81.70 91.31 85.03 63.81 93.60 82.80 single2multi 47.25 43.18 42.80 35.67 42.23↓11.14 82.42 64.31 73.37↓7.16 69.46 75.14 82.71 77.65 19.43 87.64 68.67↓14.13 s2m-t-c 47.87 50.70 39.72 31.03 42.33↓11.04 82.06 63.82 72.94↓7.59 68.24 74.56 81.25 71.08 20.06 87.98 67.20↓15.60 s2m-g-i 49.19 43.73 47.90 40.90 45.43↓7.94 83.46 72.22 77.84↓2.69 78.82 77.90 85.79 81.37 51.04 90.55 77.58↓5.22 ColParse 55.14 52.08 56.24 51.36 53.71↑0.34 83.93 80.16 82.05↑1.52 81.65 83.87 91.06 85.89 63.17 93.76 83.23↑0.43 GME-7B 53.66 54.34 65.38 54.32 56.93 83.21 84.18 83.70 87.20 82.32 92.92 88.89 63.36 94.81 84.92 single2multi 44.39 42.94 50.47 35.10 43.23↓13.70 83.51 66.19 74.85↓8.85 75.24 77.08 84.57 82.41 20.95 90.67 71.82↓13.10 s2m-t-c 47.29 51.24 46.31 30.35 43.80↓13.13 83.03 65.61 74.32↓9.38 74.40 76.84 83.60 77.58 21.35 90.80 70.76↓14.16 s2m-g-i 45.99 42.98 56.63 39.79 46.35↓10.58 84.04 73.61 78.83↓4.87 84.45 80.35 88.21 87.03 53.24 92.27 80.93↓3.99 ColParse 62.40 59.73 68.43 57.90 62.12↑5.19 84.30 83.94 84.12↑0.42 87.12 83.94 92.84 89.62 62.53 94.97 85.17↑0.25 UniME-V2-2B 9.50 15.78 15.51 19.03 14.96 54.24 77.98 66.11 61.19 65.48 76.76 59.65 45.25 77.50 64.31 single2multi 37.44 47.15 27.88 26.68 34.79↑19.83 67.36 59.21 63.29↓2.82 47.81 54.17 71.61 61.52 9.44 76.72 53.55↓10.76 s2m-t-c 35.32 42.69 28.26 25.47 32.94↑17.98 63.55 55.59 59.57↓6.54 44.16 48.21 67.30 52.26 9.72 72.55 49.03↓15.28 s2m-g-i 44.64 48.14 46.83 44.53 46.04↑31.08 78.62 77.90 78.26↑12.15 60.30 60.76 77.30 68.15 30.63 82.23 63.23↓1.08 ColParse 44.52 45.86 48.15 51.48 47.50↑32.54 80.50 79.46 79.98↑13.87 58.51 63.17 74.52 66.58 35.57 80.21 63.09↓1.22 UniME-V2-7B 26.77 23.69 24.68 31.17 26.58 78.25 82.25 80.25 60.60 79.43 80.61 64.94 45.35 82.17 68.85 single2multi 45.06 53.38 44.29 27.79 42.63↑16.05 72.30 67.77 70.04↓10.21 52.99 64.35 74.06 69.25 16.48 82.14 59.88↓8.97 s2m-t-c 43.60 58.23 42.39 28.00 43.06↑16.48 70.92 65.12 68.02↓12.23 52.28 59.47 69.08 62.67 16.85 80.61 56.83↓12.02 s2m-g-i 50.51 53.70 46.85 37.13 47.05↑20.47 77.78 80.58 79.18↓1.07 60.99 68.25 76.65 72.59 34.78 85.05 66.39↓2.46 ColParse 54.95 50.07 54.92 50.14 52.52↑25.94 81.16 83.32 82.24↑1.99 61.90 77.80 78.41 71.89 44.43 84.68 69.85↑1.00 B3-2B 38.41 31.80 45.23 45.10 40.14 78.56 74.87 76.72 51.75 66.86 70.43 45.73 36.69 77.81 58.21 single2multi 36.98 45.67 23.09 18.57 31.08↓9.06 67.88 60.70 64.29↓12.43 42.15 56.31 60.69 51.81 11.43 74.38 49.46↓8.75 s2m-t-c 34.49 41.79 18.12 20.80 28.80↓11.34 65.79 58.75 62.27↓14.45 39.49 53.55 56.26 43.83 11.33 71.09 45.93↓12.28 s2m-g-i 40.64 44.71 33.32 38.76 39.36↓0.78 75.41 77.59 76.50↓0.22 51.45 64.51 62.71 55.60 30.12 78.04 57.07↓1.14 ColParse 45.03 39.23 48.20 49.30 45.44↑5.30 79.98 80.79 80.39↑3.67 51.00 62.94 67.06 53.86 32.02 80.01 57.82↓0.39 B3-7B 47.29 44.81 50.84 48.05 47.75 82.07 82.26 82.17 65.83 76.77 84.54 68.55 52.86 85.75 72.38 single2multi 45.33 52.80 39.17 30.33 41.91↓5.84 77.57 66.96 72.27↓9.90 54.91 67.44 72.97 68.97 17.68 82.51 60.75↓11.63 s2m-t-c 42.29 50.60 34.12 25.95 38.24↓9.51 76.48 66.12 71.30↓10.87 52.95 61.93 71.13 62.39 18.19 81.25 57.97↓14.41 s2m-g-i 49.68 53.53 48.27 38.41 47.47↓0.28 81.14 79.17 80.16↓2.01 64.55 71.83 77.07 72.23 39.18 84.84 68.28↓4.10 ColParse 53.72 49.50 52.40 50.15 51.44↑3.69 83.15 83.60 83.38↑1.21 65.97 75.92 80.19 73.87 48.05 86.13 71.69↓0.69", "source": "marker_v2", "marker_block_id": "/page/24/Table/2"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0192", "section": "C.6.1. EFFECT OF BALANCING FACTOR", "page_start": 26, "page_end": 26, "type": "FigureGroup", "text": "Figure 10. The comparison of the model-level performance using ColParse across difference balancing factors. The dash lines refer to the base results; and the star points refer to the best-performing balancing factors.", "source": "marker_v2", "marker_block_id": "/page/25/FigureGroup/373"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0193", "section": "C.6.2. EFFECT OF DOCUMENT PARSING MODEL", "page_start": 26, "page_end": 26, "type": "Text", "text": "The evaluation results demonstrate the superior balance of efficiency and accuracy achieved by MinerU2.5 compared to existing specialized vision-language models. Ta ble 6 quantifies inference efficiency on A100 (80G) hardware, utilizing Tokens/sec to measure generation speed and Pages/sec to evaluate end-to-end throughput. The findings show that while the 0.9B MinerU2-VLM maintains the highest processing speed, MinerU2.5 serves as the runnerup with 2422 tokens/s and 2.25 pages/s, both of which significantly outperform 3B-parameter baselines such as", "source": "marker_v2", "marker_block_id": "/page/25/Text/6"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0194", "section": "C.6.2. EFFECT OF DOCUMENT PARSING MODEL", "page_start": 26, "page_end": 26, "type": "TableGroup", "text": "Table 6. Inference efficiency comparison of MinerU2.5. The results for MinerU2.5 and baselines are tested on the A100(80G) machine. The best and runner-up results in each column are bolded and underlined, respectively. Model Para. Tokens/sec Pages/sec MinerU2-VLM 0.9B 3089 2.84 dots.ocr 3.0B 311 0.28 MonkeyOCR-pro-3B 3.7B 520 0.47 MonkeyOCR-pro-1.2B 1.9B 589 0.53 Nanonets-OCR-s 3.7B 605 0.55 MinerU2.5 1.2B 2422 2.25", "source": "marker_v2", "marker_block_id": "/page/25/TableGroup/374"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0195", "section": "C.6.2. EFFECT OF DOCUMENT PARSING MODEL", "page_start": 26, "page_end": 26, "type": "Text", "text": "dots.ocr and MonkeyOCR-pro. Simultaneously, Table Ta ble 7 benchmarks parsing accuracy across multiple categories on OmniDocBench, employing composite Overall scores, Edit Distances for text and reading order, and structural similarity metrics (CDM and TEDS) for formulas and tables. MinerU2.5 achieves state-of-the-art performance across all six indicators, recording a top Overall score of 90.67 and the lowest error rates in text and layout recognition. MonkeyOCR-pro-3B and dots.ocr alternate as runnerup models across structural and textual tasks, yet MinerU2.5 remains the only method to consistently lead in every evaluated dimension of document parsing quality.", "source": "marker_v2", "marker_block_id": "/page/25/Text/9"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0196", "section": "C.6.3. EFFICIENCY ANALYSIS", "page_start": 26, "page_end": 26, "type": "Text", "text": "Table 8 summarizes the average number of parsed vectors per document across 24 datasets in five VDR benchmarks.", "source": "marker_v2", "marker_block_id": "/page/25/Text/11"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0197", "section": "C.6.3. EFFICIENCY ANALYSIS", "page_start": 27, "page_end": 27, "type": "TableGroup", "text": "Table 7. Document parsing performance on OmniDocBench (Ouyang et al., 2025) across multiple tasks. The best and runner-up results are bolded and underlined, respectively. Μŧ Mi dot Mo Μc Na", "source": "marker_v2", "marker_block_id": "/page/26/TableGroup/574"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0198", "section": "C.6.3. EFFICIENCY ANALYSIS", "page_start": 27, "page_end": 27, "type": "TableGroup", "text": "Methods Para. Overall↑ TextEdit↓ FormulaCDM↑ Table TEDS↑ Table TEDS-S↑ Read OrderEdit↓ MinerU2-VLM 0.9B 85.56 0.078 80.95 83.54 87.66 0.086 dots.ocr 3B 88.41 0.048 83.22 86.78 90.62 0.053 MonkeyOCR-pro-3B 3.7B 88.85 0.075 87.25 86.78 90.63 0.128 MonkeyOCR-pro-1.2B 1.9B 86.96 0.084 85.02 84.24 89.02 0.130 Nanonets-OCR-s 3.7B 85.59 0.093 85.90 80.14 85.57 0.108 MinerU2.5 1.2B 90.67 0.047 88.46 88.22 92.38 0.044", "source": "marker_v2", "marker_block_id": "/page/26/TableGroup/579"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0199", "section": "C.6.3. EFFICIENCY ANALYSIS", "page_start": 27, "page_end": 27, "type": "TableGroup", "text": "Table 8. Summary of the average number of parsed vectors per document across 24 datasets in five VDR benchmarks. The values represent the number of layout-informed sub-image embeddings (k) generated by the document parser (MinerU2.5). Benchmark Dataset Avg. #Vectors (k) MMLongBench-doc 6.04 MMLongBench MMLongBench-page 6.04 ViDoSeek ViDoSeek-doc 5.60 ViDoSeek-page 5.77 VisRAG ArxivQA 1.97 VisRAG ChartQA 2.98 VisRAG InfoVQA 4.20 VisRAG VisRAG MP-DocVQA 5.82 VisRAG PlotQA 2.06 VisRAG SlideVQA 4.66 ViDoRe biomedical lectures v2 3.88 ViDoRe economics reports v2 5.89 ViDoRe-v2 ViDoRe esg reports human labeled v2 6.92 ViDoRe esg reports v2 multilingual 6.91 ViDoRe arxiviva 1.99 ViDoRe docvqa 5.64 ViDoRe infovqa 4.52 ViDoRe shiftproject 8.97 ViDoRe syntheticDocQA artificial intelligence 5.71 ViDoRe-V1 ViDoRe syntheticDocQA energy 5.07 ViDoRe syntheticDocQA government reports 5.98 ViDoRe syntheticDocQA healthcare industry 6.11 ViDoRe tabfquad 2.10 ViDoRe tatdqa 8.58", "source": "marker_v2", "marker_block_id": "/page/26/TableGroup/580"}

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/appendix_text_v3.txt

@@ -0,0 +1,350 @@
+[p. 12 | section: A. Algorithm Workflow | type: Text]
+We formalize the complete workflow of our proposed ColParse framework in two distinct algorithms. Algorithm 1 details the offline indexing process, where ColParse generates a highly compact set of document embeddings through its sequential three-stage process. Subsequently, Algorithm 2 illustrates the online retrieval phase, where the final relevance score is efficiently computed via a MaxSim operation using this compressed set of embeddings.
+
+[p. 12 | section: Algorithm 1 The Offline Indexing Process of ColParse | type: Text]
+Input : A document image d \in \mathbb{R}^{H \times W \times 3} ;
+
+[p. 12 | section: Algorithm 1 The Offline Indexing Process of ColParse | type: Code]
+A document parser model \Psi_{\text{parse}}; A \text{ single-vector encoder } \Phi_{\text{enc}} : \mathbb{R}^{H' \times W' \times 3} \to \mathbb{R}^D Output: A compact multi-vector representation \mathbf{D}_{\text{ColParse}} \subset \mathbb{R}^{k \times D} /* Stage 1: Layout-Informed Document Parsing */ [\{b_j,c_j\}]_{j=1}^k \leftarrow \Psi_{\text{parse}}(d) \; ; \; \text{ // Get $k$ bboxes and content types} \mathcal{S}_d \leftarrow \emptyset \; \text{ for } j \leftarrow 1 \; \text{ to $k$ do}
+
+[p. 12 | section: Algorithm 1 The Offline Indexing Process of ColParse | type: Text]
+\begin{vmatrix} s_j \leftarrow \operatorname{Crop}(d, b_j) ; & \text{Crop doc image } d \text{ using bbox } b_j \\ S_d \leftarrow S_d \cup \{s_j\} \end{vmatrix}
+
+[p. 12 | section: Algorithm 1 The Offline Indexing Process of ColParse | type: Text]
+end
+
+[p. 12 | section: Algorithm 1 The Offline Indexing Process of ColParse | type: Code]
+/* Stage 2: Dual-Stream Encoding */ \mathbf{D}_{local} \leftarrow \emptyset \text{ for } each \ sub-image \ s_j \in \mathcal{S}_d \ \mathbf{do} \mid \mathbf{v}_{local}^{(j)} \leftarrow \Phi_{enc}(s_j) \ ; \qquad \text{// Encode local region} \mid \mathbf{D}_{local} \leftarrow \mathbf{D}_{local} \cup \{\mathbf{v}_{local}^{(j)}\}\nend
+
+[p. 12 | section: Algorithm 1 The Offline Indexing Process of ColParse | type: Text]
+\mathbf{v}_{\mathrm{global}} \leftarrow \Phi_{\mathrm{enc}}(d) ; // Encode entire page for global context
+
+[p. 12 | section: Algorithm 1 The Offline Indexing Process of ColParse | type: Text]
+/* Stage 3: Global-Local Fusion */
+
+[p. 12 | section: Algorithm 1 The Offline Indexing Process of ColParse | type: Text]
+\begin{aligned} \mathbf{D}_{\texttt{ColParse}} &\leftarrow \emptyset \ \ \textbf{for} \ \textit{each local vector} \ \mathbf{v}_{local}^{(j)} \in \mathbf{D}_{local} \ \textbf{do} \\ & | \ \mathbf{d}_{\texttt{fused}}^{(j)} \leftarrow \mathbf{v}_{local}^{(j)} + \mathbf{v}_{\texttt{global}} \ ; \qquad \textit{//} \ \text{Fuse by element-wise} \\ & \ \ \text{addition} \\ & | \ \ \mathbf{D}_{\texttt{ColParse}} \leftarrow \mathbf{D}_{\texttt{ColParse}} \cup \{\mathbf{d}_{\texttt{fused}}^{(j)}\} \end{aligned}
+
+[p. 12 | section: Algorithm 1 The Offline Indexing Process of ColParse | type: Text]
+end
+
+[p. 12 | section: Algorithm 1 The Offline Indexing Process of ColParse | type: Text]
+return D_{\texttt{ColParse}}
+
+[p. 12 | section: B. More Theoretical Analysis | type: Text]
+This section provides a detailed theoretical exposition of the concepts introduced in Section 3, grounding the ColParse framework in fundamental principles of information theory.
+
+[p. 12 | section: B.1. Information-Theoretic Preliminaries | type: Text]
+We begin by defining the core concepts used in our analysis. Definition B.1 (Mutual Information). The mutual information I(X;Y) between two random variables X and Y
+
+[p. 12 | section: B.1. Information-Theoretic Preliminaries | type: Code]
+Algorithm 2 The Online Retrieval Process with ColParse
+
+[p. 12 | section: B.1. Information-Theoretic Preliminaries | type: Text]
+/* Step 1: Encode Query
+
+[p. 12 | section: B.1. Information-Theoretic Preliminaries | type: Text]
+\mathbf{Q} \leftarrow \Phi_{\mathrm{enc}}(q) ; // Encode q into N_q token vectors \{\mathbf{q}_i\}
+
+[p. 12 | section: B.1. Information-Theoretic Preliminaries | type: Text]
+/* Step 2: Late-Interaction Scoring (MaxSim) */ score \leftarrow 0 for each query vector \mathbf{q}_i \in \mathbf{Q} do
+
+[p. 12 | section: B.1. Information-Theoretic Preliminaries | type: Text]
+*/
+
+[p. 12 | section: B.1. Information-Theoretic Preliminaries | type: Text]
+max\_sim \leftarrow -\infty for each fused document vector \mathbf{d}_{fused}^{(j)} \in \mathbf{D}_{ColParse} do sim \leftarrow \mathbf{q}_i^{\top} \mathbf{d}_{fused}^{(j)} ; // Assuming L2-normalized vectors max\_sim \leftarrow \max(max\_sim, sim) end score \leftarrow score + max\_sim ; // Aggregate max
+
+[p. 12 | section: B.1. Information-Theoretic Preliminaries | type: Text]
+similarity
+
+[p. 12 | section: B.1. Information-Theoretic Preliminaries | type: Text]
+end
+
+[p. 12 | section: return score | type: Text]
+measures their mutual dependence. It is defined as:
+
+[p. 12 | section: return score | type: Equation]
+I(X;Y) = \sum_{x \in \mathcal{X}} \sum_{y \in \mathcal{Y}} p(x,y) \log \frac{p(x,y)}{p(x)p(y)}. (7)
+
+[p. 12 | section: return score | type: Text]
+where p(x,y) is the joint probability distribution, and p(x) and p(y) are the marginal distributions. I(X;Y)=0 if and only if X and Y are independent.
+
+[p. 12 | section: return score | type: Text]
+Definition B.2 (Conditional Mutual Information). The conditional mutual information I(X;Y|Z) measures the mutual information between X and Y given that a third variable Z is known:
+
+[p. 12 | section: return score | type: Equation]
+I(X;Y|Z) = \mathbb{E}_{z \sim p(z)}[I(X;Y|Z=z)]. \tag{8}
+
+[p. 12 | section: return score | type: Text]
+Theorem B.3 (Chain Rule for Mutual Information). For a set of random variables \{X_1, \ldots, X_n\} and another variable Y, the chain rule states:
+
+[p. 12 | section: return score | type: Equation]
+I(X_1, \dots, X_n; Y) = \sum_{i=1}^n I(X_i; Y | X_1, \dots, X_{i-1}). (9)
+
+[p. 12 | section: return score | type: Text]
+This rule is fundamental for decomposing the information content of a complex system.
+
+[p. 12 | section: return score | type: Text]
+Theorem B.4 (Data Processing Inequality (DPI)). For any Markov chain of random variables X \to Y \to Z , where Z is conditionally independent of X given Y, the following inequality holds:
+
+[p. 12 | section: return score | type: Equation]
+I(X;Z) < I(X;Y) \text{ and } I(X;Z) < I(Y;Z). (10)
+
+[p. 12 | section: return score | type: Text]
+This theorem formalizes the notion that post-processing (the step from Y to Z) cannot increase information about the original source X.
+
+[p. 13 | section: B.2. The Information Bottleneck (IB) Principle in VDR | type: Text]
+As stated in Section 3.3, the VDR compression task can be framed as an IB problem (Tishby et al., 2000). The objective is to find a compressed representation Z of a document D that maximizes information about a relevance variable R, while minimizing information about the source D itself.
+
+[p. 13 | section: B.2. The Information Bottleneck (IB) Principle in VDR | type: Text]
+Proof of Intractability. The IB Lagrangian (Eq. 3 in the main text) requires computing an expectation over the distribution of all possible queries, P(Q).
+
+[p. 13 | section: B.2. The Information Bottleneck (IB) Principle in VDR | type: Equation]
+\mathcal{L}(Z) = I(Z; D) - \beta \int_{q \in \mathcal{Q}} P(q)I(Z; R(D, q))dq. \quad (11)
+
+[p. 13 | section: B.2. The Information Bottleneck (IB) Principle in VDR | type: Text]
+Since P(Q) is unknown and potentially infinite at the time of document indexing, this objective cannot be directly optimized. Therefore, practical methods must rely on principled approximations or surrogates for this ideal objective. ColParse provides such a surrogate.
+
+[p. 13 | section: B.3. Justification for Structural Disentanglement | type: Text]
+ColParse's parsing stage, \Psi_{\text{parse}}(D) = \{S_1, \dots, S_k\} , is justified by the Semantic Concentration Axiom. We now provide a more formal justification.
+
+[p. 13 | section: B.3. Justification for Structural Disentanglement | type: Text]
+Axiom B.5 (Semantic Concentration). For a given query Q = q, there exists a primary semantic region S_{j^*} \in \{S_j\} that contains almost all the information required to determine relevance. The remaining regions S_{\neg j^*} = \{S_j\}_{j \neq j^*} provide negligible additional information.
+
+[p. 13 | section: B.3. Justification for Structural Disentanglement | type: Equation]
+I(S_{\neg j^*}; R|S_{j^*}, Q = q) \approx 0. (12)
+
+[p. 13 | section: B.3. Justification for Structural Disentanglement | type: Text]
+Justification. This axiom is an empirical assumption about the nature of user queries and documents. For a query "What were the revenues in Q3 2023?", the answer is almost certainly contained entirely within a single financial table. Information in other regions (e.g., the abstract, a methodology figure) is conditionally irrelevant once the correct table is identified.
+
+[p. 13 | section: B.3. Justification for Structural Disentanglement | type: Text]
+Corollary B.6 (Information Equivalence of Decomposed Representation). Under the Semantic Concentration Axiom, the mutual information between the entire document and the relevance variable is approximately equal to the maximum information contained in any single semantic region.
+
+[p. 13 | section: B.3. Justification for Structural Disentanglement | type: Equation]
+I(D;R) \approx \max_{j \in \{1,\dots,k\}} I(S_j;R). \tag{13}
+
+[p. 13 | section: B.3. Justification for Structural Disentanglement | type: Text]
+Proof. From the chain rule, I(D;R) = I(S_1, ..., S_k; R) . For a specific query q, let j^* be the index of the primary region. We have:
+
+[p. 13 | section: B.3. Justification for Structural Disentanglement | type: Equation]
+I(D; R|Q = q) = I(S_{j^*}; R|Q = q) + I(S_{\neg j^*}; R|S_{j^*}, Q = q). (14)
+
+[p. 13 | section: B.3. Justification for Structural Disentanglement | type: Text]
+Applying Axiom B.5, the second term vanishes: I(D; R|Q=q) \approx I(S_{j^*}; R|Q=q) . Taking the expectation over P(Q), and using the property that \mathbb{E}[\max(X_i)] \geq \max(\mathbb{E}[X_i]) , we arrive at the approximation that the total information is well-represented by the information in the best possible channel, justifying the multi-vector approach. \square
+
+[p. 13 | section: B.4. Justification for Synergistic Fusion | type: Text]
+The fusion stage combines local vectors \{V_j = \Phi_{\rm enc}(S_j)\} with a global vector V_{\rm global} = \Phi_{\rm enc}(D) to produce the final representation \{Z_j = V_j + V_{\rm global}\} .
+
+[p. 13 | section: B.4. Justification for Synergistic Fusion | type: Text]
+Definition B.7 (Contextual Information Gain). The contextual information gain for region j is the additional information about relevance R provided by the global context V_{\text{global}} , given that the local information V_j is already known.
+
+[p. 13 | section: B.4. Justification for Synergistic Fusion | type: Equation]
+G_j^{\text{context}} \triangleq I(V_{\text{global}}; R|V_j). (15)
+
+[p. 13 | section: B.4. Justification for Synergistic Fusion | type: Text]
+Theorem B.8 (Information in the Fused Representation). The information contained in the fused vector Z_j = V_j + V_{global} is upper-bounded by the joint information of its components.
+
+[p. 13 | section: B.4. Justification for Synergistic Fusion | type: Equation]
+I(Z_j; R) \le I(V_j, V_{global}; R). \tag{16}
+
+[p. 13 | section: B.4. Justification for Synergistic Fusion | type: Text]
+Proof. The fused vector Z_j is a deterministic function of V_j and V_{\text{global}} . This forms the Markov chain (V_j, V_{\text{global}}) \to Z_j \to R . Applying the Data Processing Inequality (Theorem B.4) to this chain directly yields the result.
+
+[p. 13 | section: B.4. Justification for Synergistic Fusion | type: Text]
+Corollary B.9 (Condition for Information Improvement). The fusion step is beneficial (i.e., Z_j is more informative than V_j alone) if and only if the fusion function successfully captures a non-zero portion of the contextual information gain.
+
+[p. 13 | section: B.4. Justification for Synergistic Fusion | type: Equation]
+\Delta I_j = I(Z_j; R) - I(V_j; R) > 0 \iff I(Z_j; R|V_j) > 0. (17)
+
+[p. 13 | section: B.4. Justification for Synergistic Fusion | type: Text]
+Proof. From the chain rule, I(Z_j,V_j;R) = I(V_j;R) + I(Z_j;R|V_j) . Since Z_j is a function of V_j and V_{\text{global}} , knowing V_j does not make Z_j fully determined. The term I(Z_j;R|V_j) represents the information that the variation in Z_j (caused by V_{\text{global}} ) provides about R, even when V_j is fixed. A positive net improvement \Delta I_j > 0 directly requires this conditional term to be positive, which in turn means the fusion must have encoded some of the contextual gain G_j^{\text{context}} . The vector addition V_j + V_{\text{global}} is a simple, effective function for this purpose, as it non-linearly interacts with the query vector during the dot product scoring: \mathbf{q}^{\top}(\mathbf{v}_j + \mathbf{v}_{\text{global}}) , allowing the model to utilize both local and global signals.
+
+[p. 14 | section: C.1. Benchmark Details | type: Text]
+To ensure a comprehensive and robust evaluation of our framework, we anchor our experiments on five mainstream benchmark suites for VDR, all of which are integrated within the visdoc section of the MMEB (Meng et al., 2025) . The following benchmarks collectively cover a diverse range of document types, query complexities, and retrieval scenarios, providing a multifaceted view of model performance.
+
+[p. 14 | section: C.1. Benchmark Details | type: ListGroup]
+▶ ViDoRe-V1 (Faysse et al., 2024) 3 : As a foundational benchmark for page-level VDR, ViDoRe-V1 was one of the first to systematically evaluate systems on visuallyrich documents. It combines repurposed academic VQA datasets with practical, topic-specific tasks, highlighting the inherent shortcomings of traditional text-only retrieval systems on documents containing complex layouts, tables, and figures. ▶ ViDoRe-V2 (Mace et al. ´ , 2025) 4 : As a successor to ViDoRe-V1, ViDoRe-V2 aims to raise the bar by introducing more challenging and realistic retrieval scenarios to address the performance saturation observed on the original. Its core contributions include the use of long-form, cross-document, and multilingual queries generated via a hybrid synthetic and human-in-the-loop process, which reduces extractive bias and more accurately reflects real-world user interactions. ▶ VisRAG (Yu et al., 2024) 5 : The VisRAG benchmark is constructed to specifically evaluate vision-based RAG pipelines by aggregating and refining multiple existing VQA datasets. Its primary contribution is the unification of a wide spectrum of document types—including scientific figures, charts, infographics, and presentation slides—under a single evaluation framework, coupled with a crucial filtering process to remove contextdependent questions and ensure its suitability for openretrieval tasks. ▶ ViDoSeek (Wang et al., 2025) 6 : ViDoSeek is a novel benchmark designed to evaluate end-to-end RAG systems on visually-rich documents that require complex reasoning. Its main contribution lies in providing a large document corpus where each query corresponds to a unique answer, which allows for a more realistic and rigorous evaluation of both the retrieval and subsequent reasoning stages in a large-scale setting.
+
+[p. 14 | section: C.1. Benchmark Details | type: Text]
+▶ MMLongBench (Ma et al., 2024b) 7 : MMLongBench is specifically designed to assess the long-context, multimodal understanding capabilities of LVLMs. It stands out by using lengthy documents (averaging 47.5 pages) and featuring a significant portion of cross-page questions that require multi-hop reasoning, as well as unanswerable questions to probe for model hallucination, thus rigorously testing a model's ability to locate and synthesize information from extensive contexts.
+
+[p. 14 | section: C.2. Model Details | type: Text]
+We select ten representative single-vector multimodal retrieval models from recent literature to serve as the base models for our experiments. These models, built upon various architectures and pre-training paradigms, provide a comprehensive testbed for evaluating the versatility and effectiveness of our proposed framework.
+
+[p. 14 | section: C.2. Model Details | type: ListGroup]
+▶ VLM2Vec-V1-2B/7B (Jiang et al., 2024) 8 : As a pioneering work in universal multimodal embeddings, VLM2Vec introduces a contrastive training framework to adapt any VLM for a wide range of tasks. Its core contribution is reformulating diverse multimodal tasks ( e.g., classification, VQA, retrieval) into a unified instructionfollowing ranking problem, enabling the model to learn general-purpose embeddings for both images and text. ▶ VLM2Vec-V2-2B (Meng et al., 2025) 9 : This model extends its predecessor by broadening the scope of multimodal embeddings to include videos and visual documents, in addition to images and text. Its primary contribution is the introduction of a more comprehensive benchmark and a unified training strategy that allows a single model to effectively learn representations across static, temporal, and structured visual data formats. ▶ LamRA-Ret-7B (Liu et al., 2025a) 10 : LamRA explores repurposing generative Large Multimodal Models for retrieval tasks, unifying diverse retrieval scenarios under a single instruction-following framework. Its key innovation is a two-stage training strategy that first pretrains the model on language-only tasks before multimodal instruction tuning, progressively adapting the generative model for retrieval. ▶ GME-2B/7B (Zhang et al., 2024b) 11 : The General Multimodal Embedder (GME) framework focuses on
+
+[p. 14 | section: C.2. Model Details | type: Code]
+7 yubo2333/MMLongBench-Doc 8 VLM2Vec-Full 9 VLM2Vec-V2.0 10 LamRA-Ret 11 Alibaba-NLP/gme-models
+
+[p. 14 | section: C.2. Model Details | type: Footnote]
+3 vidore/vidore-benchmark 4
+
+[p. 14 | section: C.2. Model Details | type: Footnote]
+vidore/vidore-benchmark-v2 5 openbmb/visrag
+
+[p. 14 | section: C.2. Model Details | type: Footnote]
+6 Qiuchen-Wang/ViDoSeek
+
+[p. 15 | section: C.2. Model Details | type: Text]
+improving universal multimodal retrieval by leveraging a more diverse mix of training data, including singlemodal, cross-modal, and fused-modal examples. Its core contribution is a novel data synthesis pipeline for creating large-scale, high-quality fused-modal data, which significantly enhances the model's ability to handle complex queries and retrieve visual documents.
+
+[p. 15 | section: C.2. Model Details | type: ListGroup]
+▶ UniME-V2-2B/7B (Gu et al., 2025) 12 : UniME-V2 enhances representation learning by leveraging an MLLM as a "judge" to generate soft semantic matching scores for query-candidate pairs. This MLLM-as-a-Judge mechanism facilitates more effective hard negative mining and allows the embedding model to learn finergrained semantic distinctions, significantly improving its discriminative capacity. ▶ B3-2B/7B (Thirukovalluru et al., 2025) 13 : Breaking the Batch Barrier (B3) introduces a novel batch construction strategy for contrastive learning that curates high-quality batches rich in hard negatives. Instead of random sampling, it uses a teacher model and graphbased community detection to group mutually challenging examples together, thereby improving training efficiency and achieving state-of-the-art performance even with significantly smaller batch sizes.
+
+[p. 15 | section: C.3. MinerU2.5 Details | type: Text]
+To resolve the trade-off between the immense computational overhead (O(N 2 ) complexity) and information loss associated with directly processing high-resolution document images, MinerU2.5 innovatively employs a decoupled, coarse-to-fine two-stage strategy:
+
+[p. 15 | section: C.3. MinerU2.5 Details | type: ListGroup]
+1. Stage I: Global Layout Analysis. In this stage, the model first resizes the input document image to a medium-resolution thumbnail ( e.g., 1036×1036 pixels). It then performs a fast, global layout analysis on this thumbnail to identify all structural elements (such as paragraphs, tables, formulas, and figures) and their positions at a low computational cost. 2. Stage II: Local Content Recognition. Guided by the layout information detected in the first stage, the model precisely crops the respective semantic regions from the original high-resolution image. Subsequently, it performs parallel, fine-grained content recognition ( e.g., text OCR, table structuring, formula transcription) on these native-resolution cropped patches. This preserves high recognition accuracy while avoiding redundant computations on the entire high-resolution image.
+
+[p. 15 | section: C.3. MinerU2.5 Details | type: Text]
+Algorithm 3 details the layout-informed image splitting
+
+[p. 15 | section: C.3. MinerU2.5 Details | type: Text]
+process used in ColParse .
+
+[p. 15 | section: C.3. MinerU2.5 Details | type: Code]
+Algorithm 3 Layout-Informed Image Splitting for ColParse Input : A document image d ∈ R H×W×3 ; A layout detector model Ψsplit (e.g., DocLayoutY- OLO); Minimum area ratio threshold τ (default 0.01); Maximum sub-images count Nmax (default 20); Grid fallback parameters: Rgrid, Cgrid Output : A list of cropped sub-images Sd; A list of content type labels Cd (optional) /* Step 1: Semantic Layout Detection */ TotalArea ← H × W B ← ∅, Cd ← ∅ if Ψsplit is available then R ← Ψsplit.predict(d) ; // Returns list of {bbox, category, score} if R is not empty then for each region r ∈ R do b ← (x1, y1, x2, y2) from r.poly c ← MapCategoryID(r.category id) B ← B ∪ {(b, c, centerY(b), centerX(b))} end end end /* Step 2: Fallback & Sorting Mechanism */ if B is empty then B ← GridBasedSplit(H, W, Rgrid, Cgrid) ; // Fallback to grid else B ← SortByReadingOrder(B) ; // Sort by vertical bands, then horizontal end /* Step 3: Filtering, Cropping and Output */ Sd ← ∅, count ← 0 for each (b, c) ∈ B do if count ≥ Nmax then break end Area ← width(b) × height(b) if Area/TotalArea ≥ τ then s ← Crop(d, b) ; // Extract region from original image Sd ← Sd∪{s} Cd ← Cd∪{c} count ← count+1 end end return Sd, Cd
+
+[p. 15 | section: C.4. Main Results | type: Text]
+Refer to Table 2 and Table 3 for all results of ColParse and baselines across five benchmarks.
+
+[p. 15 | section: C.4. Main Results | type: Footnote]
+12 TianchengGu/unime-v2
+
+[p. 15 | section: C.4. Main Results | type: Footnote]
+13
+
+[p. 16 | section: C.4. Main Results | type: Caption]
+Table 2. Performance comparison on MMLongBench and ViDoRe-V1 benchmarks. For each model block, we bold the best-performing optimization method in each column (except for the base result). The average scores for optimizations are shown with relative gains (↑/↓) compared to the base model.
+
+[p. 16 | section: C.4. Main Results | type: Table]
+MMLongBench ViDoRe-V1 Method Doc Page Avg. Arxiv DocV InfoV Shift TabF TatD S-AI S-En S-HC S-Gov Avg. VLM2Vec-V1-2B 25.62 26.23 25.93 17.80 13.98 39.41 9.18 36.32 10.56 16.39 15.96 23.56 24.11 20.73 s2m-add 21.54 15.08 18.31↓7.62 35.07 15.61 52.15 6.62 36.51 10.39 26.23 31.22 30.29 33.61 27.77↑7.04 s2m-mul 22.07 15.10 18.59↓7.34 34.91 16.12 52.61 6.57 36.65 10.34 23.90 30.89 28.88 33.60 27.45↑6.72 cl-t-c 16.57 10.59 13.58↓12.35 13.97 3.97 21.62 8.73 23.20 11.56 12.94 28.27 19.16 26.72 17.01↓3.72 cl-t-m 14.35 8.67 11.51↓14.42 16.47 2.98 26.97 16.27 18.48 8.59 14.40 23.03 18.71 13.10 15.90↓4.83 cl-s-c 18.29 11.54 14.92↓11.01 18.83 4.87 22.80 13.49 24.64 12.63 15.45 24.88 20.78 20.85 17.92↓2.81 cl-s-m 15.45 9.06 12.26↓13.67 15.73 2.54 27.46 15.68 18.99 9.76 13.52 24.73 22.59 20.72 17.17↓3.56 c-sem 18.76 13.88 16.32↓9.61 29.33 5.87 35.70 22.38 35.38 13.95 23.98 37.88 33.91 30.44 26.88↑6.15 multi-img 23.61 15.85 19.73↓6.20 37.78 13.96 54.20 11.35 40.50 9.10 20.36 32.15 36.75 32.04 28.82↑8.09 ColParse 34.31 29.83 32.07↑6.14 47.66 28.12 69.23 47.11 57.05 20.43 62.24 63.77 65.51 62.54 52.37↑31.64 VLM2Vec-V1-7B 23.85 37.63 30.74 28.07 17.93 44.47 2.06 16.78 5.86 17.93 25.04 28.90 14.59 20.16 s2m-add 34.57 28.11 31.34↑0.60 50.30 25.66 66.73 38.21 63.75 23.49 70.27 61.87 70.68 66.38 53.73↑33.57 s2m-mul 35.29 28.49 31.89↑1.15 50.46 26.34 67.28 36.45 64.13 23.84 69.43 61.94 68.35 66.99 53.52↑33.36 cl-t-c 18.75 14.31 16.53↓14.21 17.10 7.71 26.41 21.58 29.62 14.61 18.52 27.98 26.09 26.65 21.63↑1.47 cl-t-m 25.34 20.46 22.90↓7.84 28.06 7.00 47.43 29.48 45.92 19.42 27.66 47.56 44.96 52.94 35.04↑14.88 cl-s-c 22.25 14.48 18.37↓12.37 21.80 8.11 30.92 28.36 25.91 19.61 27.40 40.30 36.25 32.58 27.12↑6.96 cl-s-m 26.31 20.38 23.35↓7.39 28.58 8.85 49.12 32.67 46.07 20.85 31.96 50.86 50.85 49.99 36.98↑16.82 c-sem 31.36 26.16 28.76↓1.98 45.77 15.38 59.20 37.46 57.05 30.17 47.00 60.89 64.22 66.76 48.39↑28.23 multi-img 33.77 25.60 29.69↓1.05 49.40 19.55 62.09 28.19 66.34 17.19 41.89 51.44 60.84 48.89 44.58↑24.42 ColParse 43.34 40.58 41.96↑11.22 60.47 34.42 70.39 53.67 77.12 31.33 74.81 69.64 80.79 75.89 62.85↑42.69 VLM2Vec-V2-2B 48.55 50.34 49.45 78.98 38.51 82.21 64.57 87.64 44.68 85.06 82.99 89.89 87.08 74.16 s2m-add 43.33 39.00 41.17↓8.28 66.60 38.47 72.80 58.28 65.85 54.50 90.10 84.97 83.93 80.36 69.59↓4.57 s2m-mul 45.72 40.66 43.19↓6.26 68.04 39.80 75.45 58.87 69.90 54.31 90.93 84.53 84.56 82.40 70.88↓3.28 cl-t-c 20.08 18.48 19.28↓30.17 28.47 6.46 29.32 20.51 40.06 17.75 20.23 35.23 27.90 22.54 24.85↓49.31 cl-t-m 25.04 21.94 23.49↓25.96 44.82 7.28 42.98 31.23 38.96 21.19 26.38 47.14 45.34 37.44 34.28↓39.88 cl-s-c 23.25 18.43 20.84↓28.61 29.34 8.76 27.77 22.92 40.48 21.16 20.26 30.91 24.61 27.64 25.39↓48.77 cl-s-m 26.08 22.30 24.19↓25.26 44.95 7.16 45.42 19.48 39.29 25.44 31.57 47.18 48.84 40.13 34.95↓39.21 c-sem 29.94 27.99 28.97↓20.48 61.30 17.28 62.19 40.55 55.53 33.02 56.52 62.33 68.54 70.10 52.74↓21.42 multi-img 38.69 29.00 33.85↓15.60 65.39 25.80 70.54 27.92 71.56 33.79 55.93 62.93 72.27 53.14 53.93↓20.23 ColParse 49.49 50.53 50.01↑0.56 80.17 46.33 83.53 72.76 86.74 52.40 91.36 85.83 95.47 89.52 78.41↑4.25 LamRA-Ret 19.78 13.24 16.51 29.31 19.56 63.00 15.83 51.44 7.70 21.10 29.81 37.18 31.95 30.69 s2m-add 32.18 17.82 25.00↑8.49 9.80 14.37 46.06 19.49 28.13 19.16 22.79 30.98 37.98 24.81 25.36↓5.33 s2m-mul 30.52 16.37 23.45↑6.94 9.71 14.98 45.61 17.37 27.91 17.20 20.45 28.32 38.60 23.01 24.32↓6.37 cl-t-c 14.88 8.09 11.49↓5.02 5.57 2.22 19.76 13.94 17.55 6.59 14.23 20.69 17.83 20.52 13.89↓16.80 cl-t-m 20.43 13.28 16.86↑0.35 7.47 3.85 30.91 17.54 13.41 12.37 27.06 35.63 39.20 31.88 21.93↓8.76 cl-s-c 15.32 9.05 12.19↓4.32 6.78 2.62 20.53 13.79 18.30 9.11 17.05 27.03 25.60 22.25 16.31↓14.38 cl-s-m 19.84 12.91 16.38↓0.13 7.75 4.82 32.16 15.61 14.02 10.23 24.81 34.66 37.31 27.20 20.86↓9.83 c-sem 19.62 10.69 15.16↓1.35 13.51 5.91 34.43 9.13 32.30 15.94 26.32 36.02 39.82 37.01 25.04↓5.65 multi-img 23.32 7.71 15.52↓0.99 6.90 6.10 21.26 8.00 22.26 13.60 13.91 15.91 20.30 9.67 13.79↓16.90 ColParse 30.74 19.50 25.12↑8.61 17.27 20.61 58.35 21.39 39.69 13.34 25.57 35.27 43.21 26.29 30.10↓0.59 GME-2B 52.07 53.14 52.61 82.59 56.46 88.97 89.72 93.20 70.33 98.49 92.15 98.15 95.65 86.57 s2m-add 50.92 46.52 48.72↓3.89 71.85 43.00 83.58 72.17 77.65 69.45 93.29 89.80 89.92 90.50 78.12↓8.45 s2m-mul 53.82 50.06 51.94↓0.67 76.87 47.23 85.49 81.53 84.77 74.63 95.72 93.07 93.85 92.18 82.53↓4.04 cl-t-c 15.54 10.26 12.90↓39.71 14.96 4.70 25.32 16.93 24.97 12.01 12.02 19.02 19.26 19.34 16.85↓69.72 cl-t-m 17.95 13.61 15.78↓36.83 30.20 3.58 33.07 30.13 31.16 16.11 22.27 33.99 36.32 33.28 27.01↓59.56 cl-s-c 16.03 12.59 14.31↓38.30 15.90 5.88 22.23 26.42 26.16 15.36 20.67 18.52 25.26 22.00 19.84↓66.73
+
+[p. 16 | section: C.4. Main Results | type: Text]
+Continued on next page
+
+[p. 17 | section: C.4. Main Results | type: Caption]
+Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations
+
+[p. 17 | section: C.4. Main Results | type: Table]
+Table 2 – Continued from previous page MMLongBench ViDoRe-V1 Method Doc Page Avg. Arxiv DocV InfoV Shift TabF TatD S-AI S-En S-HC S-Gov Avg. cl-s-m 16.39 12.39 14.39↓38.22 31.79 4.01 30.55 19.79 31.61 14.18 24.92 30.34 34.24 28.50 24.99↓61.58 c-sem 23.78 17.15 20.47↓32.14 45.19 11.49 51.57 25.34 42.78 27.26 51.26 44.73 49.30 53.58 40.25↓46.32 multi-img 45.10 32.87 38.99↓13.62 67.63 27.89 74.78 47.02 78.34 43.20 59.99 63.76 73.68 63.19 59.95↓26.62 ColParse 53.06 54.24 53.65↑1.04 82.39 54.11 88.93 88.59 92.33 70.65 97.75 92.30 97.91 96.10 86.11↓0.46 GME-7B 54.01 55.80 54.91 87.59 56.05 91.96 94.25 93.72 76.26 99.63 95.45 99.63 99.06 89.36 s2m-add 53.57 49.76 51.67↓3.24 75.91 46.41 85.01 80.64 83.47 74.66 95.72 92.93 94.35 92.17 82.13↓7.23 s2m-mul 50.73 45.95 48.34↓6.57 72.29 43.38 83.23 72.85 78.28 69.64 92.92 90.23 90.05 90.17 78.30↓11.06 cl-t-c 12.53 8.39 10.46↓44.45 7.32 5.16 17.36 18.86 16.81 12.83 13.28 16.18 20.30 13.97 14.21↓75.15 cl-t-m 14.64 9.03 11.84↓43.07 6.78 3.80 26.36 24.67 17.82 15.52 19.12 26.65 24.74 28.60 19.41↓69.95 cl-s-c 13.08 8.93 11.01↓43.90 7.25 3.85 16.83 17.08 18.68 13.51 19.17 20.53 19.15 21.64 15.77↓73.59 cl-s-m 13.62 7.93 10.78↓44.13 6.88 4.73 22.77 18.40 18.43 10.43 16.80 21.03 26.40 20.86 16.67↓72.69 c-sem 15.42 9.22 12.32↓42.59 15.02 6.13 27.52 18.27 36.39 22.53 20.40 28.91 30.81 26.13 23.21↓66.15 multi-img 47.50 36.01 41.76↓13.15 72.48 33.71 78.84 45.50 84.88 45.65 64.59 68.82 75.90 69.50 63.99↓25.37 ColParse 54.96 56.51 55.74↑0.83 87.35 57.91 90.76 95.35 95.44 75.92 99.63 94.67 99.63 98.89 89.56↑0.20 UniME-V2-2B 18.52 40.10 29.31 36.52 12.43 42.41 14.09 51.11 7.39 20.23 32.96 24.21 19.25 26.06 s2m-add 36.66 30.03 33.35↑4.04 50.78 25.04 58.61 37.71 54.90 36.67 68.00 68.50 69.21 68.39 53.78↑27.72 s2m-mul 36.06 28.97 32.52↑3.21 50.76 23.93 58.23 32.85 54.90 34.94 63.01 64.67 66.31 61.07 51.07↑25.01 cl-t-c 19.45 14.04 16.75↓12.56 19.03 7.04 25.79 22.55 30.85 14.57 13.25 21.82 32.30 26.67 21.39↓4.67 cl-t-m 16.70 12.63 14.67↓14.64 21.88 4.30 33.79 26.88 20.47 9.65 15.83 36.40 34.18 27.10 23.05↓3.01 cl-s-c 19.23 16.02 17.63↓11.68 19.53 6.89 23.75 21.96 28.29 15.97 17.34 31.06 33.58 27.73 22.61↓3.45 cl-s-m 17.77 13.19 15.48↓13.83 23.67 5.33 36.60 32.03 20.17 11.46 19.46 37.16 36.87 26.76 24.95��1.11 c-sem 25.11 20.89 23.00↓6.31 38.24 13.31 51.27 38.76 40.04 21.25 33.42 51.86 60.03 52.80 40.10↑14.04 multi-img 33.99 24.39 29.19↓0.12 50.79 19.08 60.61 30.15 58.91 24.36 43.08 46.76 61.06 50.79 44.56↑18.50 ColParse 44.22 44.19 44.21↑14.90 62.39 37.69 73.33 72.35 77.45 38.83 82.50 75.80 89.35 85.84 69.55↑43.49 UniME-V2-7B 33.19 45.72 39.46 63.23 24.91 65.25 11.16 41.54 14.18 41.89 40.56 57.44 42.78 40.29 s2m-add 40.28 37.61 38.95↓0.51 55.17 32.79 69.82 58.54 65.47 45.46 84.23 81.72 87.38 89.16 66.97↑26.68 s2m-mul 40.57 38.14 39.36↓0.10 55.29 33.72 70.38 56.55 65.82 44.48 84.10 81.55 86.00 89.89 66.78↑26.49 cl-t-c 20.22 13.70 16.96↓22.50 24.63 5.79 29.62 27.36 21.81 15.81 17.79 34.01 27.97 28.61 23.34↓16.95 cl-t-m 23.20 19.90 21.55↓17.91 34.15 5.21 49.47 34.87 32.02 18.91 35.36 47.97 43.55 56.22 35.77↓4.52 cl-s-c 20.91 18.25 19.58↓19.88 23.98 8.28 28.30 26.59 28.51 21.08 25.63 32.59 30.66 43.46 26.91↓13.38 cl-s-m 24.81 19.71 22.26↓17.20 35.09 6.60 49.60 36.77 33.75 20.75 37.45 48.31 55.85 60.28 38.45↓1.84 c-sem 31.71 29.01 30.36↓9.10 59.37 19.84 64.17 42.06 50.69 34.28 70.63 66.71 72.14 75.85 55.57↑15.28 multi-img 34.79 28.05 31.42↓8.04 53.99 23.06 67.47 40.81 70.58 30.36 52.38 65.49 72.19 62.65 53.90↑13.61 ColParse 45.90 48.26 47.08↑7.62 64.78 37.43 78.51 76.74 81.47 43.69 89.32 82.68 92.74 88.13 73.55↑33.26 B3-2B 37.10 32.07 34.59 57.00 29.38 68.09 48.31 71.55 18.09 74.13 64.64 75.44 63.13 56.98 s2m-add 35.86 27.36 31.61↓2.98 47.93 24.57 64.85 39.99 50.16 33.16 66.24 63.90 69.75 65.45 52.60↓4.38 s2m-mul 35.66 26.97 31.32↓3.27 47.89 24.47 64.82 35.42 49.59 32.43 63.95 63.09 66.79 64.47 51.29↓5.69 cl-t-c 17.88 13.56 15.72↓18.87 19.64 4.53 32.73 10.60 20.15 9.10 12.68 26.80 30.88 18.64 18.58↓38.40 cl-t-m 18.36 14.95 16.66↓17.93 18.38 5.34 33.44 13.13 19.91 12.42 14.05 32.81 30.84 26.31 20.66↓36.32 cl-s-c 20.70 16.46 18.58↓16.01 16.92 5.67 34.77 16.79 22.56 13.04 22.41 34.47 37.31 36.75 24.07↓32.91 cl-s-m 20.41 17.63 19.02↓15.57 19.74 4.72 35.90 17.36 20.74 15.77 19.76 38.43 36.44 33.89 24.28↓32.70 c-sem 23.80 21.20 22.50↓12.09 34.98 11.15 51.26 21.98 37.45 19.41 29.16 49.03 54.62 46.73 35.58↓21.40 multi-img 28.93 16.01 22.47↓12.12 41.03 13.38 46.20 20.46 46.89 10.26 31.74 38.83 32.44 32.80 31.40↓25.58 ColParse 42.06 37.60 39.83↑5.24 56.47 30.91 66.69 67.42 69.33 29.42 79.88 72.67 83.24 71.41 62.74↑5.76 B3-7B 46.09 45.10 45.60 68.95 43.38 79.86 66.56 84.12 37.06 81.01 81.25 88.57 81.30 71.21 44.95 40.38 42.67↓2.93 59.11 38.45 75.42 69.63 70.95 51.71 88.55 81.68 86.10 86.34 70.79↓0.42 s2m-add s2m-mul cl-t-c 45.11 40.61 42.86↓2.74 23.96 19.73 21.85↓23.75 59.50 25.53 38.63 8.05 75.74 69.25 71.18 51.72 87.36 82.07 85.17 47.76 13.93 32.90 19.17 29.29 44.67 45.02 86.23 33.61 70.69↓0.52 29.99↓41.22
+
+[p. 17 | section: C.4. Main Results | type: Text]
+Continued on next page
+
+[p. 18 | section: C.4. Main Results | type: Caption]
+Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations
+
+[p. 18 | section: C.4. Main Results | type: Table]
+Table 2 – Continued from previous page Method MMLongBench ViDoRe-V1 Doc Page Avg. Arxiv DocV InfoV Shift TabF TatD S-AI S-En S-HC S-Gov Avg. cl-t-m 24.79 21.29 23.04↓22.56 31.72 9.21 52.63 18.92 28.84 22.25 31.48 47.83 51.71 43.89 33.85↓37.36 cl-s-c 25.66 21.29 23.48↓22.12 24.40 11.69 48.70 23.29 33.17 25.56 34.54 42.09 55.93 52.12 35.15↓36.06 cl-s-m 25.52 20.72 23.12↓22.48 24.78 11.31 49.68 26.51 32.71 26.20 37.63 46.29 51.53 44.47 35.11↓36.10 c-sem 29.88 25.98 27.93↓17.67 56.08 24.05 65.81 25.04 52.18 33.15 59.78 57.76 67.14 70.35 51.13↓20.08 multi-img 35.37 22.45 28.91↓16.69 52.60 22.38 56.80 27.96 65.24 16.05 53.57 47.96 43.69 51.39 43.76↓27.45 ColParse 49.11 48.39 48.75↑3.15 67.68 42.17 79.02 78.06 81.64 47.60 85.17 82.04 92.00 88.73 74.41↑3.20
+
+[p. 19 | section: C.4. Main Results | type: TableGroup]
+Table 3. Performance comparison on ViDoRe-V2, ViDoSeek, and VisRAG benchmarks. For each model block, we bold the bestperforming optimization method in each column (except for the base result). The average scores for optimizations are shown with relative gains (↑/↓) compared to the base model. Method ViDoRe-V2 ViDoSeek VisRAG Bio-L Eco-R ESG-H ESG-M Avg. Doc Page Avg. Arxiv Chart InfoV MP-Doc Plot Slide Avg. VLM2Vec-V1-2B 6.88 14.15 12.25 20.54 13.46 56.40 67.73 62.07 41.68 58.21 70.79 42.74 23.83 74.07 51.89 s2m-add 10.40 8.59 4.68 3.46 6.78↓6.68 40.78 31.23 36.01↓26.06 28.94 44.79 59.76 30.86 9.67 59.59 38.94↓12.95 s2m-mul 10.00 8.40 5.09 3.51 6.75↓6.71 40.63 31.00 35.82↓26.25 28.85 44.64 60.48 30.55 9.83 59.52 38.98↓12.91 cl-t-c 14.86 14.49 4.14 4.48 9.49↓3.97 42.17 31.65 36.91↓25.16 13.56 18.52 25.66 13.11 1.41 31.91 17.36↓34.53 cl-t-m 13.34 4.60 8.78 5.15 7.97↓5.49 36.71 26.80 31.76↓30.31 8.29 18.27 37.74 11.85 1.13 32.46 18.29↓33.60 cl-s-c 16.38 10.03 4.93 4.64 9.00↓4.46 43.89 31.95 37.92↓24.15 13.03 27.01 28.61 18.82 1.54 31.31 20.05↓31.84 cl-s-m 13.46 6.50 6.89 3.74 7.65↓5.81 37.20 26.86 32.03↓30.04 8.21 22.04 37.37 12.65 1.15 33.31 19.12↓32.77 c-sem 20.32 11.01 11.86 10.44 13.41↓0.05 56.19 47.96 52.08↓9.99 16.53 44.39 43.88 24.77 3.86 54.32 31.29↓20.60 multi-img 15.59 30.33 14.90 29.55 5.70 33.21 5.95 38.33 10.54↓2.92 45.24 75.23 34.26 70.19 39.75↓22.32 30.20 38.18 41.53 60.09 60.40 69.44 29.53 48.29 8.44 18.83 60.10 76.95 38.37↓13.52 ColParse 32.86↑19.40 72.71↑10.64 51.96↑0.07 VLM2Vec-V1-7B s2m-add 4.93 13.74 6.82 11.27 9.19 54.26 77.39 65.83 52.58 69.83 71.43 52.86 34.24 73.22 59.03 s2m-mul 29.49 28.79 38.26 37.40 31.73 29.66 22.80 21.43 30.57↑21.38 29.32↑20.13 62.50 62.34 53.64 52.73 58.07↓7.76 57.54↓8.29 45.56 45.29 48.98 50.21 68.39 69.10 49.01 49.54 11.01 11.02 70.61 70.82 48.93↓10.10 49.33↓9.70 cl-t-c 17.52 18.22 11.68 8.82 14.06↑4.87 49.02 40.25 44.64↓21.19 15.12 21.70 36.61 19.07 2.22 37.89 22.10↓36.93 cl-t-m 22.08 14.23 24.74 15.61 19.17↑9.98 58.38 51.13 54.76↓11.07 18.75 30.49 58.40 22.24 3.90 52.32 31.02↓28.01 cl-s-c 19.74 22.03 14.36 12.74 17.22↑8.03 51.57 41.97 46.77↓19.06 15.24 17.35 36.59 24.24 1.79 35.59 21.80↓37.23 cl-s-m 23.08 16.08 20.53 14.47 18.54↑9.35 58.59 50.84 54.72↓11.11 18.76 31.65 58.91 24.90 3.10 52.63 31.66↓27.37 c-sem 29.88 31.92 30.24 20.34 28.10↑18.91 72.03 66.00 69.02↑3.19 39.77 57.94 63.94 41.41 14.39 69.66 47.85↓11.18 multi-img 30.91 39.93 23.80 17.07 27.93↑18.74 62.03 49.71 55.87↓9.96 45.32 44.84 64.95 40.71 10.55 67.95 45.72↓13.31 ColParse 42.63 42.89 50.55 42.86 44.73↑35.54 78.34 78.61 78.48↑12.65 54.43 70.30 69.27 58.49 33.46 77.98 60.66↑1.63 VLM2Vec-V2-2B 44.45 45.77 48.77 46.98 46.49 80.88 83.68 82.28 77.38 82.30 86.27 71.60 66.96 92.04 79.43 s2m-add 41.02 49.68 41.29 20.26 38.06↓8.43 69.01 64.83 66.92↓15.36 62.37 73.06 78.41 75.52 22.67 85.73 66.29↓13.14 s2m-mul 42.43 50.96 42.90 21.91 39.55↓6.94 71.66 66.41 69.04↓13.24 63.30 73.50 79.84 76.41 23.36 87.29 67.28↓12.15 cl-t-c 19.84 21.50 12.92 9.75 16.00↓30.49 48.48 49.99 49.24↓33.04 25.86 22.73 34.54 19.89 10.59 36.62 25.04↓54.39 cl-t-m 25.10 18.29 15.42 12.37 17.80↓28.69 57.69 59.96 58.83↓23.45 37.64 42.56 56.47 24.76 11.54 51.93 37.48↓41.95 cl-s-c 23.22 18.36 10.47 12.71 16.19↓30.30 49.73 51.97 50.85↓31.43 28.13 27.93 34.99 25.08 10.77 40.66 27.93↓51.50 cl-s-m 27.68 18.68 17.65 9.93 18.49↓28.00 57.31 61.50 59.41↓22.87 37.55 43.07 55.65 26.56 11.74 52.27 37.81↓41.62 c-sem multi-img 36.74 31.11 20.49 21.25 27.40↓19.09 71.91 71.47 71.69↓10.59 53.17 61.34 67.94 53.96 23.78 74.81 55.83↓23.60 ColParse 30.01 50.06 45.94 53.76 23.58 57.41 17.56 46.40 29.27↓17.22 63.17 80.94 50.95 83.87 57.06↓25.22 62.83 77.18 61.38 78.05 75.92 84.37 54.34 78.07 24.62 58.74 75.43 91.95 59.09↓20.34 51.91↑5.42 82.41↑0.13 78.06↓1.37 LamRA-Ret s2m-add 10.75 12.36 9.65 21.34 6.32 18.02 11.18 21.30 9.48 60.17 55.39 28.81 27.91 44.49 11.17 2.75 63.50 33.31 59.78 44.51 33.57 28.61 29.42 4.86 57.59 49.08 42.51 s2m-mul 9.91 17.36 21.00 20.88 18.26↑8.78 17.29↑7.81 52.53 25.95 41.65↓2.84 39.24↓5.25 2.75 33.53 43.10 28.28 5.19 49.10 27.19↓15.32 26.99↓15.52 cl-t-c 6.69 7.44 5.85 1.05 5.26↓4.22 39.52 30.63 35.08↓9.41 4.26 19.69 21.94 9.90 1.53 26.54 13.98↓28.53 cl-t-m 7.75 10.84 8.20 3.97 7.69↓1.79 44.87 38.69 41.78↓2.71 5.46 22.85 38.15 17.45 2.13 35.03 20.18↓22.33 cl-s-c 7.57 6.57 6.90 1.97 5.75↓3.73 38.41 30.98 34.70↓9.79 5.19 17.67 21.50 11.02 1.52 25.81 13.79↓28.72 cl-s-m 8.55 10.98 7.70 4.07 7.83↓1.65 45.50 38.86 42.18↓2.31 5.45 23.56 37.71 17.17 2.24 35.58 20.29↓22.22 c-sem 8.67 9.41 8.69 4.43 7.80↓1.68 53.12 38.51 45.82↑1.33 9.07 40.44 41.90 18.39 3.92 44.45 26.36↓16.15 multi-img 3.36 10.48 13.43 6.24 8.38↓1.10 28.18 10.96 19.57↓24.92 1.22 23.90 35.05 16.83 5.71 18.51 16.87↓25.64 ColParse 15.81 17.65 23.75 25.48 20.67↑11.19 58.55 39.99 49.27↑4.78 5.91 60.76 54.91 38.77 25.27 62.34 41.33↓1.18 GME-2B 54.25 50.65 59.44 49.15 53.37 81.44 79.62 80.53 81.37 81.70 91.31 85.03 63.81 93.60 82.80 s2m-add 48.33 46.46 40.98 30.02 41.45↓11.92 82.20 63.50 72.85↓7.68 70.43 77.10 85.04 80.94 20.83 89.96 70.72↓12.08 s2m-mul 51.15 49.41 46.69 28.65 43.98↓9.39 83.64 67.99 75.82↓4.71 74.25 77.91 88.07 84.75 22.40 92.99 73.40↓9.40 cl-t-c 17.91↓64.89 17.04 18.20 2.88 5.79 10.98↓42.39 34.72 28.03 31.38↓49.15 14.40 18.17 30.07 14.03 4.30 26.48 cl-t-m 18.60 12.72 12.39 7.72 12.86↓40.51 45.17 40.84 43.01↓37.52 25.41 31.73 44.53 16.56 2.77 40.18 cl-s-c 17.47 19.49 10.88 5.99 13.46↓39.91 34.42 29.29 31.86↓48.67 15.10 14.91 29.78 15.49 3.92 28.86 cl-s-m 21.66 11.82 14.37 5.63 13.37↓40.00 42.98 38.99 40.99↓39.54 24.76 30.32 45.48 17.41 2.60 39.35 c-sem 24.51 22.71 22.72 18.57 22.13↓31.24 65.58 53.29 59.44↓21.09 38.80 50.88 52.09 35.27 18.64 63.25 multi-img 32.00 55.14 41.67 52.08 25.86 56.24 19.83 51.36 29.84↓23.53 73.00 83.93 47.58 80.16 60.29↓20.24 67.87 81.65 67.77 83.87 81.98 91.06 58.54 85.89 22.85 63.17 77.89 93.76 26.86↓55.94 18.01↓64.79 26.65↓56.15 43.16↓39.64 62.82↓19.98 53.71↑0.34 82.05↑1.52 83.23↑0.43 53.66 54.34 65.38 54.32 56.93 83.21 84.18 83.70 87.20 82.32 92.92 88.89 63.36 94.81 84.92 50.40 47.49 47.63 29.02 43.64↓13.29 83.22 67.67 75.45↓8.25 73.94 79.30 87.84 84.63 22.35 92.79 73.48↓11.44 48.04 10.12 47.85 12.81 39.12 8.12 29.60 4.38 41.15↓15.78 82.30 29.66 63.00 23.82 72.65↓11.05 70.46 3.79 77.68 11.21 84.96 21.97 80.38 11.20 21.12 2.44 90.08 18.30 70.78↓14.14 9.38 10.36 6.42 4.63 8.86↓48.07 7.70↓49.23 35.86 32.84 26.74↓56.96 34.35↓49.35 3.56 21.05 35.14 12.97 1.45 24.43 11.49↓73.43 16.43↓68.49 8.58 16.36 5.41 5.20 8.89↓48.04 28.62 23.40 26.01↓57.69 5.46 11.14 23.14 11.35 1.54 21.57 10.84 9.96 6.62 2.80 7.56↓49.37 34.06 29.69 31.88↓51.82 3.93 22.24 33.74 13.14 1.42 23.76 12.37↓72.55 16.37↓68.55 16.30 13.79 8.34 8.42 11.71↓45.22 49.08 35.41 42.25↓41.45 7.40 29.57 32.02 15.54 5.76 35.26 29.32 39.47 30.32 16.76 28.97↓27.96 75.23 52.91 64.07↓19.63 72.68 70.18 85.20 63.27 23.75 81.21 62.40 59.73 68.43 57.90 62.12↑5.19 84.30 83.94 84.12↑0.42 87.12 83.94 92.84 89.62 62.53 94.97 9.50 15.78 15.51 19.03 14.96 54.24 77.98 66.11 61.19 65.48 76.76 59.65 45.25 77.50 20.93↓63.99 66.05↓18.87 85.17↑0.25 64.31 30.82 38.23 23.31 20.59 28.24↑13.28 58.58 50.31 54.45↓11.66 45.94 55.05 67.28 55.13 12.43 72.62 51.41↓12.90 29.67 34.92 21.62 19.02 26.31↑11.35 56.92 48.58 52.75↓13.36 45.21 55.60 66.84 53.32 12.49 71.71 50.86↓13.45 24.01 15.49 14.30 12.37 16.54↑1.58 46.27 38.93 42.60↓23.51 18.00 24.42 29.87 16.38 4.20 33.42 21.05↓43.26 21.67 10.06 13.15 9.43 13.58↓1.38 42.79 35.18 38.99↓27.12 17.18 27.33 42.30 14.99 5.25 44.29 25.22↓39.09 ColParse GME-7B s2m-add s2m-mul cl-t-c cl-t-m cl-s-c cl-s-m c-sem multi-img ColParse UniME-V2-2B s2m-add s2m-mul cl-t-c cl-t-m cl-s-c cl-s-m 23.56 22.75 15.34 13.02 12.00 12.48 12.62 10.46 15.88↑0.92 14.68↓0.28 48.96 43.51 39.41 37.60 44.19↓21.92 40.56↓25.55 19.80 17.20 24.93 29.53 32.56 42.92 21.55 16.67 4.33 5.00 38.90 47.36 23.68↓40.63 26.45↓37.86
+
+[p. 19 | section: C.4. Main Results | type: Text]
+Continued on next page
+
+[p. 20 | section: C.4. Main Results | type: Caption]
+Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations
+
+[p. 20 | section: C.4. Main Results | type: Table]
+Table 3 – Continued from previous page Method ViDoRe-V2 ViDoSeek VisRAG Bio-L Eco-R ESG-H ESG-M Avg. Doc Page Avg. Arxiv Chart InfoV MP-Doc Plot Slide Avg. c-sem 35.10 27.35 21.50 28.81 28.19↑13.23 64.50 60.58 62.54↓3.57 32.65 51.23 57.19 40.32 14.91 64.66 43.49↓20.82 multi-img 31.08 38.51 16.09 18.02 25.93↑10.97 56.06 43.06 49.56↓16.55 46.47 46.80 65.91 42.99 11.82 63.81 46.30↓18.01 ColParse 44.52 45.86 48.15 51.48 47.50↑32.54 80.50 79.46 79.98↑13.87 58.51 63.17 74.52 66.58 35.57 80.21 63.09↓1.22 UniME-V2-7B 26.77 23.69 24.68 31.17 26.58 78.25 82.25 80.25 60.60 79.43 80.61 64.94 45.35 82.17 68.85 s2m-add 41.25 54.38 36.91 24.88 39.36↑12.78 71.49 64.32 67.91↓12.34 51.03 64.77 75.82 67.21 15.38 81.12 59.22↓9.63 s2m-mul 41.49 53.84 35.49 23.31 38.53↑11.95 71.69 64.24 67.97↓12.28 50.89 65.45 76.45 67.54 15.32 81.68 59.56↓9.29 cl-t-c 22.76 16.17 8.68 14.14 15.44↓11.14 49.38 45.31 47.35↓32.90 19.44 24.63 35.22 17.75 4.44 36.13 22.94↓45.91 cl-t-m 27.93 14.34 19.03 16.99 19.57↓7.01 55.94 52.84 54.39↓25.86 27.05 38.47 62.09 20.81 7.64 55.56 35.27↓33.58 cl-s-c 24.75 19.62 12.41 11.69 17.12↓9.46 52.10 48.38 50.24↓30.01 20.58 21.04 34.39 23.49 5.53 35.42 23.41↓45.44 cl-s-m 30.78 20.40 18.14 15.62 21.24↓5.34 56.69 53.63 55.16↓25.09 28.73 37.84 62.65 23.01 6.78 57.14 36.03↓32.82 c-sem 37.96 29.51 27.33 21.95 29.19↑2.61 71.01 68.82 69.92↓10.33 51.40 65.80 72.53 54.42 21.98 76.81 57.16↓11.69 multi-img 35.99 45.82 24.74 14.71 30.32↑3.74 65.89 53.21 59.55↓20.70 51.79 62.63 73.67 48.06 17.25 73.00 54.40↓14.45 ColParse 54.95 50.07 54.92 50.14 52.52↑25.94 81.16 83.32 82.24↑1.99 61.90 77.80 78.41 71.89 44.43 84.68 69.85↑1.00 B3-2B 38.41 31.80 45.23 45.10 40.14 78.56 74.87 76.72 51.75 66.86 70.43 45.73 36.69 77.81 58.21 s2m-add 32.14 40.97 22.62 19.74 28.87↓11.27 67.40 56.61 62.01↓14.71 43.01 51.27 68.01 50.92 13.40 73.78 50.07↓8.14 s2m-mul 30.87 39.31 21.36 19.69 27.81↓12.33 66.73 56.12 61.43↓15.29 42.74 49.75 68.21 51.27 13.31 73.79 49.85↓8.36 cl-t-c 11.90 13.53 9.61 5.70 10.19↓29.95 44.40 41.39 42.90↓33.82 14.13 37.31 41.30 17.72 3.32 44.50 26.38↓31.83 cl-t-m 15.16 14.20 13.12 6.61 12.27↓27.87 44.92 44.57 44.75↓31.97 14.12 32.08 42.61 19.91 1.06 43.03 25.47↓32.74 cl-s-c 16.03 20.23 10.10 8.92 13.82↓26.32 48.54 47.73 48.14↓28.58 13.09 36.05 44.30 24.60 2.76 48.77 28.26↓29.95 cl-s-m 16.43 16.68 14.29 9.21 14.15↓25.99 48.08 47.08 47.58↓29.14 13.51 35.91 44.17 25.47 1.20 45.45 27.62↓30.59 c-sem 25.26 25.23 18.89 18.29 21.92↓18.22 60.79 59.97 60.38↓16.34 32.30 54.90 61.44 39.77 10.80 61.17 43.40↓14.81 multi-img 21.83 23.38 7.29 11.37 15.97↓24.17 52.38 37.41 44.90↓31.82 38.52 41.26 53.67 24.33 7.97 37.56 33.89↓24.32 ColParse 45.03 39.23 48.20 49.30 45.44↑5.30 79.98 80.79 80.39↑3.67 51.00 62.94 67.06 53.86 32.02 80.01 57.82↓0.39 B3-7B 47.29 44.81 50.84 48.05 47.75 82.07 82.26 82.17 65.83 76.77 84.54 68.55 52.86 85.75 72.38 s2m-add 44.47 48.61 35.86 28.82 39.44↓8.31 75.93 65.24 70.59↓11.58 56.95 69.80 78.61 69.19 19.74 85.71 63.33↓9.05 s2m-mul 44.79 49.78 35.45 28.68 39.68↓8.07 76.17 64.83 70.50↓11.67 57.17 70.52 79.24 70.67 20.17 85.50 63.88↓8.50 cl-t-c 18.50 22.50 20.46 15.85 19.33↓28.42 66.30 63.26 64.78↓17.39 18.07 43.88 52.94 23.57 5.85 58.35 33.78↓38.60 cl-t-m 22.98 22.71 22.21 14.16 20.52↓27.23 67.41 67.43 67.42↓14.75 23.92 43.10 62.54 28.35 7.43 61.50 37.81↓34.57 cl-s-c 20.76 26.27 28.71 22.95 24.67↓23.08 68.09 65.59 66.84↓15.33 16.71 47.07 53.44 31.71 7.56 60.01 36.08↓36.30 cl-s-m 22.17 26.43 26.15 19.85 23.65↓24.10 67.85 64.05 65.95↓16.22 19.23 43.79 53.45 35.24 8.50 61.19 36.90↓35.48 c-sem 32.31 30.23 33.27 18.99 28.70↓19.05 75.30 71.83 73.57↓8.60 46.32 67.80 75.28 54.93 24.73 80.26 58.22↓14.16 multi-img 26.48 33.14 11.25 11.91 20.70↓27.05 64.84 47.83 56.34↓25.83 51.50 46.37 63.64 34.80 11.94 44.09 42.06↓30.32 ColParse 53.72 49.50 52.40 50.15 51.44↑3.69 83.15 83.60 83.38↑1.21 65.97 75.92 80.19 73.87 48.05 86.13 71.69↓0.69
+
+[p. 21 | section: C.5.1. ALGORITHM WORKFLOW | type: Text]
+Algorithm 4 presents a unified offline indexing framework for the three variants. After shared layout parsing and dualstream encoding (Stages 1–2), Stage 3 diverges based on the specified variant type: single2multi retains raw subimage vectors; type_cluster aggregates vectors by semantic content types via averaging; global_inclusion appends the full-document global vector to the local set.
+
+[p. 21 | section: Algorithm 4 Integrated Offline Indexing for ColParse Vari | type: Code]
+Input: A document image d \in \mathbb{R}^{H \times W \times 3}; A document parser model \Psi_{parse}; A single-vector encoder \Phi_{\mathrm{enc}}: \mathbb{R}^{H' \times W' \times 3} \rightarrow \mathbb{R}^D: Mode M \in \{s2m, s2m-t-c, s2m-g-i\} Output: A multi-vector representation D<sub>variant</sub>
+
+[p. 21 | section: 1119 1120 /* Stage 1: Layout-Informed Document Parsing | type: Code]
+[\{b_j, c_j\}]_{j=1}^k \leftarrow \Psi_{\text{parse}}(d); 1121 // Get k bboxes and content 1122 types 1123 S_d \leftarrow \emptyset for j \leftarrow 1 to k do 1124 s_j \leftarrow \text{Crop}(d, b_j); // Extract sub-image for each layout 1125
+
+[p. 21 | section: 1119 1120 /* Stage 1: Layout-Informed Document Parsing | type: Text]
+component 1126 \mathcal{S}_d \leftarrow \mathcal{S}_d \cup \{s_i\}
+
+[p. 21 | section: 1128 /* Stage 2: Regional Encoding 1129 | type: Code]
+\mathbf{D}_{\text{local}} \leftarrow \emptyset for each sub-image s_j \in \mathcal{S}_d do 1130 \mathbf{v}_{\text{local}}^{(j)} \leftarrow \Phi_{\text{enc}}(s_j); // Independent regional encoding 1131 \mathbf{D}_{\mathrm{local}} \leftarrow \mathbf{D}_{\mathrm{local}} \cup \{\mathbf{v}_{\mathrm{local}}^{(j)}\} 1132 1133 end
+
+[p. 21 | section: 1128 /* Stage 2: Regional Encoding 1129 | type: Text]
+1134 /* Stage 3: Variant-specific Representation Construction */ 1135
+
+[p. 21 | section: 1128 /* Stage 2: Regional Encoding 1129 | type: Code]
+if M = s2m then 1136 \mathbf{D}_{\text{variant}} \leftarrow \mathbf{D}_{\text{local}}; // Standard layout-decomposed set
+
+[p. 21 | section: 1128 /* Stage 2: Regional Encoding 1129 | type: Code]
+else if M = s2m-t-c then \mathbf{D}_{\text{variant}} \leftarrow \emptyset \ \mathcal{T} \leftarrow \text{Unique}(\{c_1, \dots, c_k\}); \ \ // \text{Identify}
+
+[p. 21 | section: 1128 /* Stage 2: Regional Encoding 1129 | type: Code]
+1138 1139 unique content types 1140 for each type t \in \mathcal{T} do \mathbf{v}_{\text{avg}}^{(t)} \leftarrow \text{Mean}(\{\mathbf{v}_{\text{local}}^{(j)} \mid c_j = t\}); average by type 1141 // Cluster and 1142 1143 \mathbf{D}_{\text{variant}} \leftarrow \mathbf{D}_{\text{variant}} \cup \{\mathbf{v}_{\text{avg}}^{(t)}\} 1144 end 1145
+
+[p. 21 | section: 1128 /* Stage 2: Regional Encoding 1129 | type: Code]
+else if M = s2m-q-i then
+
+[p. 21 | section: 1128 /* Stage 2: Regional Encoding 1129 | type: Code]
+\mathbf{v}_{\text{global}} \leftarrow \Phi_{\text{enc}}(d); // Encode original page for context \mathbf{D}_{variant} \leftarrow \mathbf{D}_{local} \cup \{\mathbf{v}_{global}\}\;;\;\; // Append global vector to the set
+
+[p. 21 | section: 1128 /* Stage 2: Regional Encoding 1129 | type: Text]
+1151 end
+
+[p. 21 | section: 1128 /* Stage 2: Regional Encoding 1129 | type: Text]
+return D<sub>variant</sub>
+
+[p. 21 | section: C.5.2. MORE ANALYSIS | type: Text]
+Due to the limited space of main text, we leave the radar plots of performance comparison between ColParse and its variants in Figure 9.
+
+[p. 21 | section: C.5.2. MORE ANALYSIS | type: Text]
+The introduction of global page-level context is indispensable for resolving semantic ambiguities within isolated layout components. Quantitative results in Table 4 demonstrate that adding global context—even via simple inclusion (s2m-g-i)—dramatically elevates performance over the local-only single2multi baseline, lifting the VLM2Vec-V1-2B score on ViDoRe-V1 from 34.39 to 49.93. Figure 5 highlights that this gap is most pronounced in benchmarks requiring holistic understanding, where local sub-images like tables or charts often lack the necessary contextual headers found elsewhere on the page. We hypothesize that the global vector acts as a "semantic anchor" that provides the overarching topic of the document, which is essential for the late-interaction mechanism to accurately align specific query tokens with relevant sub-regions.
+
+[p. 21 | section: C.5.2. MORE ANALYSIS | type: Text]
+Maintaining the individual spatial and semantic integrity of layout components is superior to heuristic type-level clustering. As evidenced by the performance trends in Table 4 and the comparative bars in Figure 5, the s2m-t-c variant typically results in a performance regression compared to the standard single2multi, such as the 1.56point drop for VLM2Vec-V1-2B on ViDoRe-V1 benchmark. This trend is echoed across the radar charts in Figure 9, where the type-clustered variants consistently exhibit the narrowest performance profiles. This indicates that spatial locality is a vital semantic carrier in visual documents; by collapsing multiple distinct components into a single typelevel average, the model loses the fine-grained resolution required for the MaxSim operator to distinguish between specific relevant and irrelevant regions of the same type.
+
+[p. 22 | section: C.5.2. MORE ANALYSIS | type: FigureGroup]
+Figure 9. The performance comparison (evaluated by nDCG@5) between ColParse and its variants on five VDR benchmarks across ten mainstream single-vector multimodal retrieval models. Refer to Table 4 and Table 5 for detailed result records due to the space limit.
+
+[p. 23 | section: Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations | type: Caption]
+Table 4. Ablation study on MMLongBench and ViDoRe-V1 benchmarks. For each model block, we bold the best-performing method in each column (except for the base result). The average scores are shown with relative gains (↑/↓) compared to the base model.
+
+[p. 23 | section: Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations | type: Table]
+Method MMLongBench ViDoRe-V1 Doc Page Avg. Arxiv DocV InfoV Shift TabF TatD S-AI S-En S-HC S-Gov Avg. VLM2Vec-V1-2B 25.62 26.23 25.93 17.80 13.98 39.41 9.18 36.32 10.56 16.39 15.96 23.56 24.11 20.73 single2multi 25.19 19.59 22.39↓3.54 38.78 16.38 56.94 12.48 41.15 9.98 38.16 46.42 42.66 40.94 34.39↑13.66 s2m-t-c 25.37 18.43 21.90↓4.03 34.95 15.04 55.09 15.31 40.07 9.78 32.83 43.07 44.84 37.31 32.83↑12.10 s2m-g-i 31.57 26.55 29.06↑3.13 48.84 23.86 67.91 39.85 62.15 14.07 55.74 64.41 61.67 60.83 49.93↑29.20 ColParse 34.31 29.83 32.07↑6.14 47.66 28.12 69.23 47.11 57.05 20.43 62.24 63.77 65.51 62.54 52.37↑31.64 VLM2Vec-V1-7B 23.85 37.63 30.74 28.07 17.93 44.47 2.06 16.78 5.86 17.93 25.04 28.90 14.59 20.16 single2multi 35.75 29.40 32.58↑1.84 53.54 26.07 61.73 41.43 67.64 22.37 72.54 59.56 73.52 64.03 54.24↑34.08 s2m-t-c 35.62 28.64 32.13↑1.39 50.56 24.08 59.42 35.90 64.84 18.29 65.39 58.49 72.13 63.75 51.29↑31.13 s2m-g-i 37.95 34.08 36.02↑5.28 62.29 30.49 67.54 41.74 78.87 22.99 74.85 62.78 77.07 64.14 58.28↑38.12 ColParse 43.34 40.58 41.96↑11.22 60.47 34.42 70.39 53.67 77.12 31.33 74.81 69.64 80.79 75.89 62.85↑42.69 VLM2Vec-V2-2B 48.55 50.34 49.45 78.98 38.51 82.21 64.57 87.64 44.68 85.06 82.99 89.89 87.08 74.16 single2multi 45.89 41.04 43.47↓5.98 69.32 38.09 77.88 56.25 72.60 52.69 88.14 83.80 86.99 84.09 70.99↓3.17 s2m-t-c 44.99 40.97 42.98↓6.47 67.44 34.25 74.79 48.36 71.32 43.73 82.22 81.18 84.78 76.42 66.45↓7.71 s2m-g-i 48.03 46.30 47.17↓2.28 79.83 46.57 82.19 66.78 87.66 52.85 93.91 86.02 91.14 87.46 77.44↑3.28 ColParse 49.49 50.53 50.01↑0.56 80.17 46.33 83.53 72.76 86.74 52.40 91.36 85.83 95.47 89.52 78.41↑4.25 LamRA-Ret 19.78 13.24 16.51 29.31 19.56 63.00 15.83 51.44 7.70 21.10 29.81 37.18 31.95 30.69
+
+[p. 23 | section: Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations | type: Text]
+Continued on next page
+
+[p. 24 | section: Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations | type: Caption]
+Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations
+
+[p. 24 | section: Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations | type: Table]
+Table 4 – Continued from previous page Method MMLongBench ViDoRe-V1 Doc Page Avg. Arxiv DocV InfoV Shift TabF TatD S-AI S-En S-HC S-Gov Avg. single2multi 33.04 18.56 25.80↑9.29 6.06 13.60 37.83 18.90 25.46 19.02 23.99 28.27 33.31 28.03 23.45↓7.24 s2m-t-c 31.74 17.49 24.62↑8.11 6.13 10.90 35.44 16.81 28.56 15.77 15.45 27.59 24.45 19.65 20.08↓10.61 s2m-g-i 33.16 18.48 25.82↑9.31 6.06 15.79 37.71 18.96 38.78 19.19 26.00 28.27 31.68 29.03 25.15↓5.54 ColParse 30.74 19.50 25.12↑8.61 17.27 20.61 58.35 21.39 39.69 13.34 25.57 35.27 43.21 26.29 30.10↓0.59 GME-2B 52.07 53.14 52.61 82.59 56.46 88.97 89.72 93.20 70.33 98.49 92.15 98.15 95.65 86.57 single2multi 50.56 45.68 48.12↓4.49 72.46 39.99 79.82 70.54 80.91 68.07 88.91 91.11 88.85 89.86 77.05↓9.52 s2m-t-c 49.51 44.73 47.12↓5.49 70.90 37.34 77.92 67.59 79.85 59.12 86.13 86.87 87.23 83.35 73.63↓12.94 s2m-g-i 51.81 48.49 50.15↓2.46 80.89 46.62 83.93 75.78 92.12 68.33 93.48 92.60 93.36 90.49 81.76↓4.81 ColParse 53.06 54.24 53.65↑1.04 82.39 54.11 88.93 88.51 92.33 70.65 97.75 92.30 97.91 96.10 86.10↓0.47 GME-7B 54.01 55.80 54.91 87.59 56.05 91.96 94.25 93.72 76.26 99.63 95.45 99.63 99.06 89.36 single2multi 53.55 48.17 50.86↓4.05 75.57 45.26 83.21 77.90 86.05 73.97 94.72 92.04 95.19 93.18 81.71↓7.65 s2m-t-c 52.97 48.05 50.51↓4.40 75.21 44.11 82.69 78.44 86.17 66.14 91.53 89.86 91.38 88.60 79.41↓9.95 s2m-g-i 54.32 51.14 52.73↓2.18 84.29 52.98 87.36 82.08 94.79 74.48 96.65 91.10 96.92 93.70 85.44↓3.92 ColParse 54.96 56.51 55.74↑0.83 87.35 57.91 90.76 95.35 95.44 75.92 99.63 94.67 99.63 98.89 89.56↑0.20 UniME-V2-2B 18.52 40.10 29.31 36.52 12.43 42.41 14.09 51.11 7.39 20.23 32.96 24.21 19.25 26.06 single2multi 38.47 33.34 35.91↑6.60 52.64 27.01 68.58 49.89 61.20 38.92 77.36 73.65 77.92 79.30 60.65↑34.59 s2m-t-c 35.13 30.33 32.73↑3.42 48.29 21.24 62.26 43.83 58.58 27.37 70.42 66.86 74.85 67.57 54.13↑28.07 s2m-g-i 41.90 39.97 40.94↑11.63 64.23 33.32 74.84 67.10 78.40 39.00 82.16 79.38 88.43 83.62 69.05↑42.99 ColParse 44.22 44.19 44.21↑14.90 62.39 37.69 73.33 71.19 77.45 38.83 82.50 75.80 89.35 85.84 69.44↑43.38 UniME-V2-7B 33.19 45.72 39.46 63.23 24.91 65.25 11.16 41.54 14.18 41.89 40.56 57.44 42.78 40.29 single2multi 39.83 39.39 39.61↑0.15 57.08 32.54 71.38 64.54 71.82 49.00 84.00 84.26 91.18 87.49 69.33↑29.04 s2m-t-c 38.66 36.94 37.80↓1.66 54.98 25.88 65.87 56.63 70.76 41.65 80.00 77.76 86.67 77.88 63.81↑23.52 s2m-g-i 41.35 43.08 42.22↑2.76 64.50 34.55 76.81 66.35 81.95 48.84 85.63 85.38 91.83 87.90 72.37↑32.08 ColParse 45.90 48.26 47.08↑7.62 64.78 37.43 78.51 73.56 81.47 43.69 89.32 82.68 92.74 88.13 73.23↑32.94 B3-2B 37.10 32.07 34.59 57.00 29.38 68.09 48.31 71.55 18.09 74.13 64.64 75.44 63.13 56.98 single2multi 36.47 29.20 32.84↓1.75 46.63 20.13 60.31 51.58 56.45 36.55 70.40 69.80 73.83 66.85 55.25↓1.73 s2m-t-c 35.15 28.33 31.74↓2.85 43.88 18.92 54.65 41.90 54.30 26.86 67.36 61.92 72.05 60.66 50.25↓6.73 s2m-g-i 38.87 33.51 36.19↑1.60 56.06 22.52 64.41 59.98 73.09 35.98 78.34 70.48 80.13 67.17 60.82↑3.84 ColParse 42.06 37.60 39.83↑5.24 56.47 30.91 66.69 67.42 69.33 29.42 79.88 72.67 83.24 71.41 62.74↑5.76 B3-7B 46.09 45.10 45.60 68.95 43.38 79.86 66.56 84.12 37.06 81.01 81.25 88.57 81.30 71.21 single2multi 44.43 40.79 42.61↓2.99 58.94 31.91 71.96 68.80 73.69 53.07 87.14 81.75 88.38 84.76 70.04↓1.17 s2m-t-c 43.32 39.84 41.58↓4.02 56.41 29.10 68.99 53.11 70.75 44.35 83.69 77.98 84.69 78.56 64.76↓6.45 s2m-g-i 45.62 43.88 44.75↓0.85 67.23 35.81 76.70 70.48 84.62 53.26 87.14 83.19 91.86 84.44 73.47↑2.26 ColParse 49.11 48.39 48.75↑3.15 67.68 42.17 79.02 78.06 81.64 47.60 85.17 82.04 92.00 88.73 74.41↑3.20
+
+[p. 25 | section: Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations | type: Text]
+Table 5. Ablation study on ViDoRe-V2, ViDoSeek, and VisRAG benchmarks. For each model block, we bold the best-performing method in each column (except for the base result). The average scores are shown with relative gains (↑/↓) compared to the base model.
+
+[p. 25 | section: Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations | type: Table]
+Method ViDoRe-V2 ViDoSeek VisRAG Bio-L Eco-R ESG-H ESG-M Avg. Doc Page Avg. Arxiv Chart InfoV MP-Doc Plot Slide Avg. VLM2Vec-V1-2B 6.88 14.15 12.25 20.54 13.46 56.40 67.73 62.07 41.68 58.21 70.79 42.74 23.83 74.07 51.89 single2multi 17.15 18.19 10.24 7.44 13.26↓0.20 50.51 41.52 46.02↓16.05 30.17 42.91 61.28 37.60 6.80 65.76 40.75↓11.14 s2m-t-c 13.46 18.31 11.58 8.85 13.05↓0.41 50.22 40.75 45.49↓16.58 27.87 41.64 59.37 34.82 6.94 62.76 38.90↓12.99 s2m-g-i 28.49 26.29 29.18 39.15 30.78↑17.32 73.38 68.14 70.76↑8.69 40.39 52.54 70.43 47.82 8.96 76.40 49.42↓2.47 ColParse 30.33 29.55 33.21 38.33 32.86↑19.40 75.23 70.19 72.71↑10.64 38.18 60.09 69.44 48.29 18.83 76.95 51.96↑0.07 VLM2Vec-V1-7B 4.93 13.74 6.82 11.27 9.19 54.26 77.39 65.83 52.58 69.83 71.43 52.86 34.24 73.22 59.03 single2multi 34.67 37.55 33.77 26.91 33.23↑24.04 66.88 58.17 62.53↓3.30 45.06 50.30 63.65 50.23 10.83 74.00 49.01↓10.02 s2m-t-c 31.17 41.18 26.15 21.07 29.89↑20.70 64.53 54.75 59.64↓6.19 43.98 52.75 61.47 45.75 10.49 71.33 47.63↓11.40 s2m-g-i 41.08 37.99 40.11 36.54 38.93↑29.74 75.46 75.28 75.37↑9.54 53.97 64.14 65.37 55.41 27.98 77.07 57.32↓1.71 ColParse 42.63 42.89 50.55 42.86 44.73↑35.54 78.34 78.61 78.48↑12.65 54.43 70.30 69.27 58.49 33.46 77.98 60.66↑1.63 VLM2Vec-V2-2B 44.45 45.77 48.77 46.98 46.49 80.88 83.68 82.28 77.38 82.30 86.27 71.60 66.96 92.04 79.43 single2multi 42.12 51.08 41.33 24.84 39.84↓6.65 73.81 67.71 70.76↓11.52 65.36 69.18 79.34 73.19 19.59 86.77 65.57↓13.86 s2m-t-c 41.81 51.83 37.07 25.78 39.12↓7.37 73.69 66.16 69.93↓12.35 64.18 66.49 77.61 64.32 19.58 85.70 62.98↓16.45 s2m-g-i 44.34 51.99 40.57 34.53 42.86↓3.63 79.23 81.01 80.12↓2.16 77.03 74.83 82.98 78.17 54.13 90.93 76.35↓3.08 ColParse 50.06 53.76 57.41 46.40 51.91↑5.42 80.94 83.87 82.41↑0.13 77.18 78.05 84.37 78.07 58.74 91.95 78.06↓1.37 LamRA-Ret 10.75 9.65 6.32 11.18 9.48 60.17 28.81 44.49 11.17 63.50 59.78 33.57 29.42 57.59 42.51 single2multi 11.11 26.32 20.54 23.76 20.43↑10.95 53.77 30.82 42.30↓2.19 1.94 25.59 30.56 27.49 3.95 44.28 22.30↓20.21 s2m-t-c 9.36 15.57 14.49 17.11 14.13↑4.65 49.02 29.38 39.20↓5.29 2.03 22.46 31.39 21.10 3.96 42.90 20.64↓21.87 s2m-g-i 11.11 26.56 18.78 21.76 19.55↑10.07 53.74 31.77 42.76↓1.73 1.91 28.55 31.15 32.12 25.09 44.22 27.17↓15.34 ColParse 15.81 17.65 23.75 25.48 20.67↑11.19 58.55 39.99 49.27↑4.78 5.91 60.76 54.91 38.77 25.27 62.34 41.33↓1.18 GME-2B 54.25 50.65 59.44 49.15 53.37 81.44 79.62 80.53 81.37 81.70 91.31 85.03 63.81 93.60 82.80 single2multi 47.25 43.18 42.80 35.67 42.23↓11.14 82.42 64.31 73.37↓7.16 69.46 75.14 82.71 77.65 19.43 87.64 68.67↓14.13 s2m-t-c 47.87 50.70 39.72 31.03 42.33↓11.04 82.06 63.82 72.94↓7.59 68.24 74.56 81.25 71.08 20.06 87.98 67.20↓15.60 s2m-g-i 49.19 43.73 47.90 40.90 45.43↓7.94 83.46 72.22 77.84↓2.69 78.82 77.90 85.79 81.37 51.04 90.55 77.58↓5.22 ColParse 55.14 52.08 56.24 51.36 53.71↑0.34 83.93 80.16 82.05↑1.52 81.65 83.87 91.06 85.89 63.17 93.76 83.23↑0.43 GME-7B 53.66 54.34 65.38 54.32 56.93 83.21 84.18 83.70 87.20 82.32 92.92 88.89 63.36 94.81 84.92 single2multi 44.39 42.94 50.47 35.10 43.23↓13.70 83.51 66.19 74.85↓8.85 75.24 77.08 84.57 82.41 20.95 90.67 71.82↓13.10 s2m-t-c 47.29 51.24 46.31 30.35 43.80↓13.13 83.03 65.61 74.32↓9.38 74.40 76.84 83.60 77.58 21.35 90.80 70.76↓14.16 s2m-g-i 45.99 42.98 56.63 39.79 46.35↓10.58 84.04 73.61 78.83↓4.87 84.45 80.35 88.21 87.03 53.24 92.27 80.93↓3.99 ColParse 62.40 59.73 68.43 57.90 62.12↑5.19 84.30 83.94 84.12↑0.42 87.12 83.94 92.84 89.62 62.53 94.97 85.17↑0.25 UniME-V2-2B 9.50 15.78 15.51 19.03 14.96 54.24 77.98 66.11 61.19 65.48 76.76 59.65 45.25 77.50 64.31 single2multi 37.44 47.15 27.88 26.68 34.79↑19.83 67.36 59.21 63.29↓2.82 47.81 54.17 71.61 61.52 9.44 76.72 53.55↓10.76 s2m-t-c 35.32 42.69 28.26 25.47 32.94↑17.98 63.55 55.59 59.57↓6.54 44.16 48.21 67.30 52.26 9.72 72.55 49.03↓15.28 s2m-g-i 44.64 48.14 46.83 44.53 46.04↑31.08 78.62 77.90 78.26↑12.15 60.30 60.76 77.30 68.15 30.63 82.23 63.23↓1.08 ColParse 44.52 45.86 48.15 51.48 47.50↑32.54 80.50 79.46 79.98↑13.87 58.51 63.17 74.52 66.58 35.57 80.21 63.09↓1.22 UniME-V2-7B 26.77 23.69 24.68 31.17 26.58 78.25 82.25 80.25 60.60 79.43 80.61 64.94 45.35 82.17 68.85 single2multi 45.06 53.38 44.29 27.79 42.63↑16.05 72.30 67.77 70.04↓10.21 52.99 64.35 74.06 69.25 16.48 82.14 59.88↓8.97 s2m-t-c 43.60 58.23 42.39 28.00 43.06↑16.48 70.92 65.12 68.02↓12.23 52.28 59.47 69.08 62.67 16.85 80.61 56.83↓12.02 s2m-g-i 50.51 53.70 46.85 37.13 47.05↑20.47 77.78 80.58 79.18↓1.07 60.99 68.25 76.65 72.59 34.78 85.05 66.39↓2.46 ColParse 54.95 50.07 54.92 50.14 52.52↑25.94 81.16 83.32 82.24↑1.99 61.90 77.80 78.41 71.89 44.43 84.68 69.85↑1.00 B3-2B 38.41 31.80 45.23 45.10 40.14 78.56 74.87 76.72 51.75 66.86 70.43 45.73 36.69 77.81 58.21 single2multi 36.98 45.67 23.09 18.57 31.08↓9.06 67.88 60.70 64.29↓12.43 42.15 56.31 60.69 51.81 11.43 74.38 49.46↓8.75 s2m-t-c 34.49 41.79 18.12 20.80 28.80↓11.34 65.79 58.75 62.27↓14.45 39.49 53.55 56.26 43.83 11.33 71.09 45.93↓12.28 s2m-g-i 40.64 44.71 33.32 38.76 39.36↓0.78 75.41 77.59 76.50↓0.22 51.45 64.51 62.71 55.60 30.12 78.04 57.07↓1.14 ColParse 45.03 39.23 48.20 49.30 45.44↑5.30 79.98 80.79 80.39↑3.67 51.00 62.94 67.06 53.86 32.02 80.01 57.82↓0.39 B3-7B 47.29 44.81 50.84 48.05 47.75 82.07 82.26 82.17 65.83 76.77 84.54 68.55 52.86 85.75 72.38 single2multi 45.33 52.80 39.17 30.33 41.91↓5.84 77.57 66.96 72.27↓9.90 54.91 67.44 72.97 68.97 17.68 82.51 60.75↓11.63 s2m-t-c 42.29 50.60 34.12 25.95 38.24↓9.51 76.48 66.12 71.30↓10.87 52.95 61.93 71.13 62.39 18.19 81.25 57.97↓14.41 s2m-g-i 49.68 53.53 48.27 38.41 47.47↓0.28 81.14 79.17 80.16↓2.01 64.55 71.83 77.07 72.23 39.18 84.84 68.28↓4.10 ColParse 53.72 49.50 52.40 50.15 51.44↑3.69 83.15 83.60 83.38↑1.21 65.97 75.92 80.19 73.87 48.05 86.13 71.69↓0.69
+
+[p. 26 | section: C.6.1. EFFECT OF BALANCING FACTOR | type: FigureGroup]
+Figure 10. The comparison of the model-level performance using ColParse across difference balancing factors. The dash lines refer to the base results; and the star points refer to the best-performing balancing factors.
+
+[p. 26 | section: C.6.2. EFFECT OF DOCUMENT PARSING MODEL | type: Text]
+The evaluation results demonstrate the superior balance of efficiency and accuracy achieved by MinerU2.5 compared to existing specialized vision-language models. Ta ble 6 quantifies inference efficiency on A100 (80G) hardware, utilizing Tokens/sec to measure generation speed and Pages/sec to evaluate end-to-end throughput. The findings show that while the 0.9B MinerU2-VLM maintains the highest processing speed, MinerU2.5 serves as the runnerup with 2422 tokens/s and 2.25 pages/s, both of which significantly outperform 3B-parameter baselines such as
+
+[p. 26 | section: C.6.2. EFFECT OF DOCUMENT PARSING MODEL | type: TableGroup]
+Table 6. Inference efficiency comparison of MinerU2.5. The results for MinerU2.5 and baselines are tested on the A100(80G) machine. The best and runner-up results in each column are bolded and underlined, respectively. Model Para. Tokens/sec Pages/sec MinerU2-VLM 0.9B 3089 2.84 dots.ocr 3.0B 311 0.28 MonkeyOCR-pro-3B 3.7B 520 0.47 MonkeyOCR-pro-1.2B 1.9B 589 0.53 Nanonets-OCR-s 3.7B 605 0.55 MinerU2.5 1.2B 2422 2.25
+
+[p. 26 | section: C.6.2. EFFECT OF DOCUMENT PARSING MODEL | type: Text]
+dots.ocr and MonkeyOCR-pro. Simultaneously, Table Ta ble 7 benchmarks parsing accuracy across multiple categories on OmniDocBench, employing composite Overall scores, Edit Distances for text and reading order, and structural similarity metrics (CDM and TEDS) for formulas and tables. MinerU2.5 achieves state-of-the-art performance across all six indicators, recording a top Overall score of 90.67 and the lowest error rates in text and layout recognition. MonkeyOCR-pro-3B and dots.ocr alternate as runnerup models across structural and textual tasks, yet MinerU2.5 remains the only method to consistently lead in every evaluated dimension of document parsing quality.
+
+[p. 26 | section: C.6.3. EFFICIENCY ANALYSIS | type: Text]
+Table 8 summarizes the average number of parsed vectors per document across 24 datasets in five VDR benchmarks.
+
+[p. 27 | section: C.6.3. EFFICIENCY ANALYSIS | type: TableGroup]
+Table 7. Document parsing performance on OmniDocBench (Ouyang et al., 2025) across multiple tasks. The best and runner-up results are bolded and underlined, respectively. Μŧ Mi dot Mo Μc Na
+
+[p. 27 | section: C.6.3. EFFICIENCY ANALYSIS | type: TableGroup]
+Methods Para. Overall↑ TextEdit↓ FormulaCDM↑ Table TEDS↑ Table TEDS-S↑ Read OrderEdit↓ MinerU2-VLM 0.9B 85.56 0.078 80.95 83.54 87.66 0.086 dots.ocr 3B 88.41 0.048 83.22 86.78 90.62 0.053 MonkeyOCR-pro-3B 3.7B 88.85 0.075 87.25 86.78 90.63 0.128 MonkeyOCR-pro-1.2B 1.9B 86.96 0.084 85.02 84.24 89.02 0.130 Nanonets-OCR-s 3.7B 85.59 0.093 85.90 80.14 85.57 0.108 MinerU2.5 1.2B 90.67 0.047 88.46 88.22 92.38 0.044
+
+[p. 27 | section: C.6.3. EFFICIENCY ANALYSIS | type: TableGroup]
+Table 8. Summary of the average number of parsed vectors per document across 24 datasets in five VDR benchmarks. The values represent the number of layout-informed sub-image embeddings (k) generated by the document parser (MinerU2.5). Benchmark Dataset Avg. #Vectors (k) MMLongBench-doc 6.04 MMLongBench MMLongBench-page 6.04 ViDoSeek ViDoSeek-doc 5.60 ViDoSeek-page 5.77 VisRAG ArxivQA 1.97 VisRAG ChartQA 2.98 VisRAG InfoVQA 4.20 VisRAG VisRAG MP-DocVQA 5.82 VisRAG PlotQA 2.06 VisRAG SlideVQA 4.66 ViDoRe biomedical lectures v2 3.88 ViDoRe economics reports v2 5.89 ViDoRe-v2 ViDoRe esg reports human labeled v2 6.92 ViDoRe esg reports v2 multilingual 6.91 ViDoRe arxiviva 1.99 ViDoRe docvqa 5.64 ViDoRe infovqa 4.52 ViDoRe shiftproject 8.97 ViDoRe syntheticDocQA artificial intelligence 5.71 ViDoRe-V1 ViDoRe syntheticDocQA energy 5.07 ViDoRe syntheticDocQA government reports 5.98 ViDoRe syntheticDocQA healthcare industry 6.11 ViDoRe tabfquad 2.10 ViDoRe tatdqa 8.58

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/assets.json

@@ -0,0 +1,156 @@
+[
+  {
+    "filename": "_page_0_Figure_21.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_0_Figure_21.jpeg",
+    "bytes": 23995,
+    "width": 496,
+    "height": 192,
+    "aspect_ratio": 2.5833333333333335,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_0_Figure_21.jpeg"
+  },
+  {
+    "filename": "_page_21_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_21_Figure_1.jpeg",
+    "bytes": 83888,
+    "width": 1287,
+    "height": 573,
+    "aspect_ratio": 2.2460732984293195,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_21_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_21_Figure_2.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_21_Figure_2.jpeg",
+    "bytes": 63643,
+    "width": 1287,
+    "height": 377,
+    "aspect_ratio": 3.413793103448276,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_21_Figure_2.jpeg"
+  },
+  {
+    "filename": "_page_25_Figure_3.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_25_Figure_3.jpeg",
+    "bytes": 138666,
+    "width": 738,
+    "height": 1115,
+    "aspect_ratio": 0.6618834080717488,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_25_Figure_3.jpeg"
+  },
+  {
+    "filename": "_page_26_Picture_14.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_26_Picture_14.jpeg",
+    "bytes": 1671,
+    "width": 61,
+    "height": 104,
+    "aspect_ratio": 0.5865384615384616,
+    "keep": false,
+    "reject_reason": "too_small",
+    "model_path": null
+  },
+  {
+    "filename": "_page_26_Picture_7.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_26_Picture_7.jpeg",
+    "bytes": 1442,
+    "width": 53,
+    "height": 80,
+    "aspect_ratio": 0.6625,
+    "keep": false,
+    "reject_reason": "too_small",
+    "model_path": null
+  },
+  {
+    "filename": "_page_26_Picture_8.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_26_Picture_8.jpeg",
+    "bytes": 1348,
+    "width": 59,
+    "height": 68,
+    "aspect_ratio": 0.8676470588235294,
+    "keep": false,
+    "reject_reason": "too_small",
+    "model_path": null
+  },
+  {
+    "filename": "_page_26_Picture_9.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_26_Picture_9.jpeg",
+    "bytes": 1502,
+    "width": 59,
+    "height": 84,
+    "aspect_ratio": 0.7023809523809523,
+    "keep": false,
+    "reject_reason": "too_small",
+    "model_path": null
+  },
+  {
+    "filename": "_page_2_Picture_8.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_2_Picture_8.jpeg",
+    "bytes": 38185,
+    "width": 673,
+    "height": 297,
+    "aspect_ratio": 2.265993265993266,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_2_Picture_8.jpeg"
+  },
+  {
+    "filename": "_page_5_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_5_Figure_1.jpeg",
+    "bytes": 111677,
+    "width": 1056,
+    "height": 724,
+    "aspect_ratio": 1.4585635359116023,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_5_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_6_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_6_Figure_1.jpeg",
+    "bytes": 20860,
+    "width": 188,
+    "height": 796,
+    "aspect_ratio": 0.23618090452261306,
+    "keep": false,
+    "reject_reason": "too_small",
+    "model_path": null
+  },
+  {
+    "filename": "_page_6_Figure_7.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_6_Figure_7.jpeg",
+    "bytes": 26329,
+    "width": 541,
+    "height": 415,
+    "aspect_ratio": 1.3036144578313253,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_6_Figure_7.jpeg"
+  },
+  {
+    "filename": "_page_7_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_7_Figure_1.jpeg",
+    "bytes": 23559,
+    "width": 517,
+    "height": 429,
+    "aspect_ratio": 1.205128205128205,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_7_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_7_Figure_10.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/3250cb92-2f69-4e16-9df9-f569224173f0/marker_markdown/3250cb92-2f69-4e16-9df9-f569224173f0/_page_7_Figure_10.jpeg",
+    "bytes": 33936,
+    "width": 567,
+    "height": 266,
+    "aspect_ratio": 2.1315789473684212,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_7_Figure_10.jpeg"
+  }
+]

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/dataset_meta.json

@@ -0,0 +1,65 @@
+{
+  "paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 27,
+  "chunk_count": 200,
+  "main_body_chunk_count": 75,
+  "appendix_chunk_count": 117,
+  "reference_chunk_count": 8,
+  "model_text_chars": 42730,
+  "raw_markdown_chars": 191958,
+  "sanitized_chars": 130384,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 27,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 12,
+    "Code": 13,
+    "Equation": 17,
+    "Figure": 3,
+    "FigureGroup": 6,
+    "Footnote": 7,
+    "ListGroup": 10,
+    "PageFooter": 27,
+    "PageHeader": 24,
+    "Picture": 1,
+    "PictureGroup": 4,
+    "SectionHeader": 68,
+    "Table": 7,
+    "TableGroup": 7,
+    "Text": 573
+  },
+  "asset_count_raw": 14,
+  "asset_count_model_kept": 9,
+  "asset_count_rejected": 5,
+  "asset_reject_reasons": {
+    "kept": 9,
+    "too_small": 5
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt",
+  "source": "koala_icml26_due_queue",
+  "run_name": "icml26_20260429_1952_duequeue"
+}

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/main_body_chunks.jsonl

@@ -0,0 +1,75 @@
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0000", "section": "Abstract", "page_start": 1, "page_end": 1, "type": "Text", "text": "Harnessing the full potential of visually-rich documents requires retrieval systems that understand not just text, but intricate layouts, a core challenge in Visual Document Retrieval (VDR). The prevailing multi-vector architectures, while powerful, face a crucial storage bottleneck that current optimization strategies, such as embedding merging, pruning, or introducing abstract tokens, fail to resolve without compromising performance or ignoring vital layout cues. To address this, we introduce ColParse, a novel paradigm that leverages a document parsing model to generate a small set of layout-informed sub-image embeddings, which are then fused with a global pagelevel vector to create a compact and structurallyaware multi-vector representation. Extensive experiments demonstrate that ColParse reduces storage requirements by over 95% while simultaneously yielding significant performance gains across numerous benchmarks and base models. ColParse thus bridges the critical gap between the fine-grained accuracy of multi-vector retrieval and the practical demands of large-scale deployment, offering a new path towards efficient and interpretable multimodal information systems.", "source": "marker_v2", "marker_block_id": "/page/0/Text/16"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0001", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Visual Document Retrieval (VDR), the task of retrieving relevant document pages from a large-scale corpus, has become a cornerstone of modern information retrieval (Mei et al., 2025; Yan et al., 2026a). Unlike natural image retrieval, visual documents, such as academic papers, financial reports, and invoices, are defined by a dense interplay of textual content, intricate layouts, and graphical elements, as illustrated in Figure 1. To effectively capture this fine-", "source": "marker_v2", "marker_block_id": "/page/0/Text/18"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0002", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "FigureGroup", "text": "Figure 1. Comparison of natural image retrieval versus VDR.", "source": "marker_v2", "marker_block_id": "/page/0/FigureGroup/274"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0003", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "grained detail, the field has predominantly converged on multi-vector retrieval architectures (Faysse et al., 2024; Günther et al., 2025; Team, 2025). These models represent each document page as a set of patch-level embeddings and employ a late-interaction mechanism, such as MaxSim, to compute relevance (Khattab & Zaharia, 2020; Santhanam et al., 2022). This paradigm excels at aligning specific query phrases with corresponding visual or textual regions within a document, a capability essential for the high-precision information-seeking tasks inherent to VDR.", "source": "marker_v2", "marker_block_id": "/page/0/Text/23"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0004", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Despite their superior performance, the widespread adoption of multi-vector VDR models is hindered by a critical bottleneck: prohibitive storage overhead (Jayaram et al., 2024; Shrestha et al., 2024; Liu & Mao, 2023). Storing hundreds or even thousands of embedding vectors for every page makes large-scale deployment practically challenging. To address this, the research community has explored several optimization strategies, as illustrated in Figure 2. • One line of work involves merging patch embeddings, where methods like Light-ColPali (Ma et al., 2025) use clustering techniques to aggregate similar vectors. However, this approach often leads to a dilution of fine-grained information, resulting in unstable performance. 2 Another direction is pruning , where frameworks such as DocPruner (Yan et al., 2025) aim to discard redundant embeddings. These methods struggle to maintain performance under aggressive compression. 3 A third paradigm, exemplified by MetaEmbed (Xiao et al., 2025), introduces a set of abstract, learnable tokens to form a compact multi-vector representation. While innovative, these tokens lack an explicit grounding in the document's inherent layout structure, limiting their ability to capture crucial layout-specific semantics.", "source": "marker_v2", "marker_block_id": "/page/0/Text/24"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0005", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "To address the limitations of existing approaches, we introduce ColParse, a novel paradigm for constructing multivector representations that is fundamentally aligned with the structural nature of visual documents. Instead", "source": "marker_v2", "marker_block_id": "/page/0/Text/25"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0006", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Caption", "text": "Figure 2. The illustration of a multi-vector VDR model and three primary optimization strategies for its efficiency bottleneck.", "source": "marker_v2", "marker_block_id": "/page/1/Caption/5"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0007", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "of operating on a uniform grid of patches or abstract tokens, {\\tt ColParse} first employs a specialized document parsing model to intelligently segment each document page into a small set of k semantically meaningful, layout-informed sub-images (e.g., tables, figures, paragraphs), where k is typically less than 10. These k sub-images are then individually encoded by a standard single-vector retrieval model to yield k local vectors. In parallel, the entire document page is encoded to generate one global vector that captures the overall context. Finally, we fuse these representations by weighted element-wise adding the global vector to each of the k local vectors. This process results in k fused vectors for each document, which integrate both fine-grained, layout-specific details and holistic page-level context.", "source": "marker_v2", "marker_block_id": "/page/1/Text/6"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0008", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "We conducted comprehensive experiments on 24 diverse VDR datasets (Meng et al., 2025) to validate the effectiveness and robustness of our proposed framework. ColParse consistently delivers substantial performance improvements, achieving an average gain of over 10 points in nDCG@5 when applied to 10 different mainstream single-vector models. This demonstrates its remarkable flexibility as a training-free, plug-and-play module. By deeply integrating the unique structural properties of visual documents with the powerful mechanism of multi-vector retrieval, ColParse establishes a new trade-off between retrieval performance and storage efficiency. Our main contributions are as follows:", "source": "marker_v2", "marker_block_id": "/page/1/Text/7"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0009", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "ListGroup", "text": "A Novel Paradigm for Multi-Vector Construction: We introduce the first layout-informed paradigm for constructing multi-vector representations in VDR, which overcomes the storage efficiency bottleneck of conventional multi-vector models by leveraging document parsing. ② A Flexible and Robust Framework: Our method is designed as a training-free, plug-and-play framework that demonstrates robust and significant performance gains across a wide array of existing single-vector models, highlighting its versatility and ease of adoption. Superior Performance with Enhanced Interpretability: ColParse provides inherent interpretability by enabling retrieval results to be traced back to specific, parsed layout components, which significantly enhances its practicality and potential for real-world industrial applications.", "source": "marker_v2", "marker_block_id": "/page/1/ListGroup/390"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0010", "section": "2.1. Visual Document Retrieval", "page_start": 2, "page_end": 2, "type": "Text", "text": "VDR has become a crucial task for understanding visuallyrich documents, moving beyond traditional OCR-based pipelines that often lose critical layout information (Zhang et al., 2025b; Most et al., 2025). The advent of Vision-Language Models (VLMs) introduced end-to-end singlevector approaches (e.g., DSE (Ma et al., 2024a), GME (Zhang et al., 2024b), and UniSE (Liu et al., 2025b)), but these frequently struggle to capture the fine-grained semantics required for dense documents. A significant leap forward was made with the multi-vector paradigm, pioneered by ColPali (Faysse et al., 2024), which represents pages as numerous patch-level embeddings and employs late interaction for superior matching. Recent efforts have sought to optimize this paradigm at various levels: modellevel, by exploring bidirectional architectures like Modern-VBERT (Teiletche et al., 2025); data-level, through advanced data synthesis and hard-negative mining as seen in works like Llama Nemoretriever Colembed (Xu et al., 2025); and training-level, via new objectives and multi-task frameworks such as jina-embeddings-v4 (Günther et al., 2025). Despite their performance, these multi-vector models introduce a severe storage bottleneck.", "source": "marker_v2", "marker_block_id": "/page/1/Text/13"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0011", "section": "2.2. Mutli-Vector Retrieval", "page_start": 2, "page_end": 2, "type": "Text", "text": "The multi-vector paradigm, first popularized in text retrieval by ColBERT (Khattab & Zaharia, 2020), represents documents as sets of token-level embeddings to enable fine-grained matching through a late-interaction mechanism (Qian et al., 2022; Lee et al., 2023). This approach was further refined in the text domain by models like BGE-M3-Embedding (Chen et al., 2024) and Jina-ColBERT-v2 (Jha et al., 2024). The paradigm was successfully adapted for multimodal retrieval by ColPali (Faysse et al., 2024), shifting the focus to visual documents, which are inherently more complex than natural images. Despite their superior performance, these models face a critical efficiency bottleneck from the prohibitive storage cost of patch-level embeddings (Liu & Mao, 2023; Shrestha et al., 2024; Park et al., 2025). Current optimization efforts fall into three main categories, each with inherent drawbacks. (i) Pruning redundant embeddings, as seen in DocPruner (Yan et al., 2025) and Prune-then-Merge (Yan et al., 2026b), often struggles to maintain performance under aggressive compres-", "source": "marker_v2", "marker_block_id": "/page/1/Text/15"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0012", "section": "2.2. Mutli-Vector Retrieval", "page_start": 3, "page_end": 3, "type": "Text", "text": "sion. (ii) Merging similar embeddings via clustering, exemplified by Light-ColPali (Ma et al., 2025), can dilute fine-grained information, leading to unstable performance. (iii) Introducing abstract, learnable tokens, pioneered by MetaEmbed (Xiao et al., 2025) and CausalEmbed (Huo et al., 2026), creates compact representations that, however, lack an explicit grounding in the document's inherent layout structure. In contrast, ColParse addresses these limitations by leveraging document parsing to generate a compact set of layout-informed embeddings.", "source": "marker_v2", "marker_block_id": "/page/2/Text/1"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0013", "section": "2.3. Document Parsing VLM", "page_start": 3, "page_end": 3, "type": "Text", "text": "Document parsing VLMs have emerged as critical tools for converting visually-rich document images into structured formats like LaTeX or Markdown (Zhang et al., 2024a; Ouyang et al., 2025; Zhang et al., 2025c). Early models, such as Nougat (Blecher et al., 2023) and Donut (Kim et al., 2022), adopted an end-to-end, sequence-to-sequence approach but often struggled with the computational cost of high-resolution inputs. To balance accuracy and efficiency, a more recent multi-stage paradigm has gained traction. This is exemplified by models like MinerU2.5 (Niu et al., 2025), which first performs efficient layout analysis on a downsampled image before conducting targeted, high-resolution recognition on cropped regions. This coarse-to-fine strategy, also seen in models like Dolphin (Feng et al., 2025) and MonkeyOCR (Zhang et al., 2025a), effectively mitigates the O(N<sup>2</sup>) complexity of processing high-resolution images end-to-end. For ColParse, we select MinerU2.5 as our document parser, given its state-of-the-art accuracy and efficiency. A quantitative comparison with other document parsing models will be presented in Section 4.2.3.", "source": "marker_v2", "marker_block_id": "/page/2/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0014", "section": "3. Methodology", "page_start": 3, "page_end": 3, "type": "Text", "text": "In this section, we first formalize the task of VDR within the multi-vector paradigm. We then introduce the ColParse framework, detailing its multi-stage process for generating compact, layout-informed document representations. See our pseudo-code in Appendix A.", "source": "marker_v2", "marker_block_id": "/page/2/Text/5"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0015", "section": "3.1. Task Formulation", "page_start": 3, "page_end": 3, "type": "Text", "text": "The primary goal of VDR is, given a textual query q, to retrieve a ranked list of relevant document pages from a large-scale corpus \\mathcal{C}=\\{d_1,d_2,\\ldots,d_{|\\mathcal{C}|}\\} . In the conventional multi-vector retrieval paradigm, a document page d is first rendered as an image and then uniformly partitioned into a grid of N_p patches, \\{p_j\\}_{j=1}^{N_p} . A VLM, serving as an encoder \\Phi(\\cdot) , maps each patch p_j into a D-dimensional embedding, resulting in a large set of patch-level document embeddings \\mathbf{D}_{\\mathrm{grid}}=\\{\\mathbf{d}_j\\}_{j=1}^{N_p} , where each \\mathbf{d}_j\\in\\mathbb{R}^D . Concurrently, the same encoder maps the textual query q into a set", "source": "marker_v2", "marker_block_id": "/page/2/Text/7"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0016", "section": "3.1. Task Formulation", "page_start": 3, "page_end": 3, "type": "PictureGroup", "text": "Figure 3. The simplified illustration of ColParse framework.", "source": "marker_v2", "marker_block_id": "/page/2/PictureGroup/428"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0017", "section": "3.1. Task Formulation", "page_start": 3, "page_end": 3, "type": "Text", "text": "of N_q token-level embeddings \\mathbf{Q} = \\{\\mathbf{q}_i\\}_{i=1}^{N_q} , where each \\mathbf{q}_i \\in \\mathbb{R}^D . The relevance score s(q,d) between the query and the document is then computed using a late-interaction mechanism, typically MaxSim, as defined below:", "source": "marker_v2", "marker_block_id": "/page/2/Text/10"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0018", "section": "3.1. Task Formulation", "page_start": 3, "page_end": 3, "type": "Equation", "text": "s(q, d) = \\sum_{i=1}^{N_q} \\max_{j=1}^{N_p} (\\mathbf{q}_i^{\\mathsf{T}} \\mathbf{d}_j). \\tag{1}", "source": "marker_v2", "marker_block_id": "/page/2/Equation/11"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0019", "section": "3.1. Task Formulation", "page_start": 3, "page_end": 3, "type": "Text", "text": "where vectors are assumed to be L2-normalized. While this grid-based approach excels at fine-grained matching, it incurs a prohibitive storage cost of O(N_p \\times D) per document page, as N_p can be in the hundreds or thousands.", "source": "marker_v2", "marker_block_id": "/page/2/Text/12"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0020", "section": "3.1. Task Formulation", "page_start": 3, "page_end": 3, "type": "Text", "text": "The objective of our work is to address this critical bottleneck. We aim to replace the large, layout-agnostic set \\mathbf{D}_{\\text{grid}} with a highly compact, structurally-aware multi-vector representation \\mathbf{D}_{\\text{ColParse}} , which contains only k vectors, where k \\ll N_p . This new representation should significantly reduce storage requirements to O(k \\times D) while simultaneously enhancing retrieval performance by being explicitly grounded in the document's semantic layout.", "source": "marker_v2", "marker_block_id": "/page/2/Text/13"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0021", "section": "3.2. The ColParse Framework", "page_start": 3, "page_end": 3, "type": "Text", "text": "ColParse is a plug-and-play framework that revolutionizes the construction of multi-vector representations by moving \"beyond the grid.\" Instead of relying on uniform patches, it leverages structural understanding to generate a compact and semantically rich set of embeddings. As shown in Figure 3, our framework operates offline in a three-stage pipeline for each document image d \\in \\mathbb{R}^{H \\times W \\times 3} : (1) Layout-Informed Document Parsing, (2) Dual-Stream Encoding, and (3) Global-Local Fusion.", "source": "marker_v2", "marker_block_id": "/page/2/Text/15"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0022", "section": "3.2.1. LAYOUT-INFORMED DOCUMENT PARSING", "page_start": 3, "page_end": 3, "type": "Text", "text": "The foundational step of ColParse is to deconstruct the document image into its constituent semantic components. We employ a specialized, off-the-shelf document parsing model, \\Psi_{\\text{parse}}(\\cdot) , which functions as a layout detector. For a given document image d, the parser identifies a set of k distinct layout regions, outputting their corresponding bounding boxes and content type labels: [\\{b_j,c_j\\}_{j=1}^k]=\\Psi_{\\text{parse}}(d) . Here, b_j=(x_{j1},y_{j1},x_{j2},y_{j2}) is the bounding box for the j-th region, defining its coordinates within the original image. c_j", "source": "marker_v2", "marker_block_id": "/page/2/Text/17"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0023", "section": "3.2.1. LAYOUT-INFORMED DOCUMENT PARSING", "page_start": 4, "page_end": 4, "type": "Text", "text": "is a categorical label from a predefined set of content types \\mathcal{C}_{\\text{types}} = \\{\\text{'title'}, \\text{'table'}, \\text{'figure'}, \\dots \\} , indicating the semantic nature of the region. Using these bounding boxes, we crop the original image d to extract a set of k sub-images \\mathcal{S}_d . The number of sub-images, k, is dynamically determined by the parser based on the document's complexity and is typically very small (e.g., k < 10). The process can be formulated as \\mathcal{S}_d = \\{s_1, s_2, \\dots, s_k\\} where s_j = \\text{Crop}(d, b_j) . Each sub-image s_j \\in \\mathbb{R}^{H_j \\times W_j \\times 3} is of variable size, flexibly conforming to the dimensions of the detected layout component. This intelligent, content-aware segmentation forms the basis for our compact representation.", "source": "marker_v2", "marker_block_id": "/page/3/Text/1"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0024", "section": "3.2.2. DUAL-STREAM ENCODING", "page_start": 4, "page_end": 4, "type": "Text", "text": "Once the document is parsed, we generate embeddings using a standard single-vector retrieval model, \\Phi_{\\rm enc}(\\cdot) : \\mathbb{R}^{H' \\times W' \\times 3} \\to \\mathbb{R}^D , which serves as the base encoder. This stage operates in two parallel streams to capture both local, layout-specific details and global, page-level context.", "source": "marker_v2", "marker_block_id": "/page/3/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0025", "section": "3.2.2. DUAL-STREAM ENCODING", "page_start": 4, "page_end": 4, "type": "Text", "text": "Local Encoding. Each of the k variable-sized sub-images s_j from \\mathcal{S}_d is resized and independently passed through the encoder \\Phi_{\\mathrm{enc}}(\\cdot) to produce a corresponding D-dimensional local vector, \\mathbf{v}_{\\mathrm{local}}^{(j)} . This process yields a set of k local embeddings, each representing a distinct semantic unit: \\mathbf{D}_{\\mathrm{local}} = \\{\\mathbf{v}_{\\mathrm{local}}^{(j)} \\in \\mathbb{R}^D \\mid \\mathbf{v}_{\\mathrm{local}}^{(j)} = \\Phi_{\\mathrm{enc}}(s_j), \\forall s_j \\in \\mathcal{S}_d\\}_{j=1}^k .", "source": "marker_v2", "marker_block_id": "/page/3/Text/4"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0026", "section": "3.2.2. DUAL-STREAM ENCODING", "page_start": 4, "page_end": 4, "type": "Text", "text": "Global Encoding. In parallel, the entire , un-cropped document page image d is passed through the same encoder \\Phi_{\\rm enc}(\\cdot) to generate a single D-dimensional global vector, \\mathbf{v}_{\\rm global} . This vector serves as a holistic summary of the page, capturing the overall context and relationships between layout components: \\mathbf{v}_{\\rm global} = \\Phi_{\\rm enc}(d) \\in \\mathbb{R}^D . This dual-stream design ensures our final representation benefits from both the specificity of individual layout elements and the broader context of the entire page.", "source": "marker_v2", "marker_block_id": "/page/3/Text/5"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0027", "section": "3.2.3. GLOBAL-LOCAL FUSION FOR FINAL REPRESENTATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "Instead of a simple summation, we employ a weighted fusion strategy that allows for a tunable balance between these two critical streams of information. For each of the k local vectors, \\mathbf{v}_{\\text{local}}^{(j)} , we inject the holistic context captured by the single global vector, \\mathbf{v}_{\\text{global}} , using a balancing factor \\alpha \\in [0,1] . The resulting fused vector, \\mathbf{d}_{\\text{fused}}^{(j)} \\in \\mathbb{R}^D , synergistically integrates both fine-grained detail and overarching context. The fusion is performed as a weighted elementwise addition: \\mathbf{d}_{\\text{fused}}^{(j)} = \\alpha \\cdot \\mathbf{v}_{\\text{global}} + (1-\\alpha) \\cdot \\mathbf{v}_{\\text{local}}^{(j)}, \\quad \\forall j \\in \\{1,\\dots,k\\} . This weighted mechanism provides the flexibility to tailor the representation's focus.", "source": "marker_v2", "marker_block_id": "/page/3/Text/7"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0028", "section": "3.2.3. GLOBAL-LOCAL FUSION FOR FINAL REPRESENTATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "This fusion operation is performed for all k local vectors, producing the final multi-vector representation for document", "source": "marker_v2", "marker_block_id": "/page/3/Text/8"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0029", "section": "3.2.3. GLOBAL-LOCAL FUSION FOR FINAL REPRESENTATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "d, denoted as \\mathbf{D}_{\\mathtt{ColParse}} = \\{\\mathbf{d}_{\\mathtt{fused}}^{(j)}\\}_{j=1}^k \\subset \\mathbb{R}^{k \\times D} . This set, containing only k structurally-aware and context-enriched vectors, is then stored for online retrieval, achieving our goal of massive storage reduction.", "source": "marker_v2", "marker_block_id": "/page/3/Text/9"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0030", "section": "3.2.4. LATE-INTERACTION SCORING WITH ColParse", "page_start": 4, "page_end": 4, "type": "Text", "text": "During the online retrieval phase, the relevance score between a query q and a document d is computed efficiently using the compact representation \\mathbf{D}_{\\mathtt{ColParse}} . The query is first encoded into its token-level embeddings \\mathbf{Q} = \\{\\mathbf{q}_i \\in \\mathbb{R}^D\\}_{i=1}^{N_q} as standard. The MaxSim score is then calculated over the k fused document vectors:", "source": "marker_v2", "marker_block_id": "/page/3/Text/11"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0031", "section": "3.2.4. LATE-INTERACTION SCORING WITH ColParse", "page_start": 4, "page_end": 4, "type": "Equation", "text": "s_{\\texttt{ColParse}}(q, d) = \\sum_{i=1}^{N_q} \\max_{j=1}^k (\\mathbf{q}_i^\\top \\mathbf{d}_{\\text{fused}}^{(j)}). \\tag{2}", "source": "marker_v2", "marker_block_id": "/page/3/Equation/12"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0032", "section": "3.2.4. LATE-INTERACTION SCORING WITH ColParse", "page_start": 4, "page_end": 4, "type": "Text", "text": "By replacing the search over N_p grid-based vectors with a search over just k layout-informed vectors, ColParse not only dramatically reduces the storage footprint but also focuses the late-interaction mechanism on the most semantically salient parts of the document.", "source": "marker_v2", "marker_block_id": "/page/3/Text/13"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0033", "section": "3.3. Theoretical Foundation", "page_start": 4, "page_end": 4, "type": "Text", "text": "We provide a theoretical justification for ColParse from an Information Bottleneck (IB) perspective. We demonstrate that the framework's architecture serves as a principled surrogate for solving the intractable IB objective in VDR by (1) disentangling source information via parsing and (2) refining it with contextual side-information.", "source": "marker_v2", "marker_block_id": "/page/3/Text/15"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0034", "section": "3.3.1. THE VDR COMPRESSION PROBLEM AS AN INFORMATION BOTTLENECK", "page_start": 4, "page_end": 4, "type": "Text", "text": "Let D be a random variable representing a document image and Q be a random variable for a query, drawn from an unknown distribution P(Q). Let R be a relevance variable, a function of D and Q. The goal is to learn a compression function g that maps a document D to a compact representation Z = g(D) by solving the IB Lagrangian:", "source": "marker_v2", "marker_block_id": "/page/3/Text/17"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0035", "section": "3.3.1. THE VDR COMPRESSION PROBLEM AS AN INFORMATION BOTTLENECK", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\min_{q} \\mathcal{L}(Z) = I(Z; D) - \\beta \\mathbb{E}_{Q}[I(Z; R(D, Q))]. \\quad (3)", "source": "marker_v2", "marker_block_id": "/page/3/Equation/18"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0036", "section": "3.3.1. THE VDR COMPRESSION PROBLEM AS AN INFORMATION BOTTLENECK", "page_start": 4, "page_end": 4, "type": "Text", "text": "This objective seeks to minimize the information Z retains about the source D (compression) while maximizing the information it preserves about the relevance R (prediction). The expectation \\mathbb{E}_Q over the unknown query distribution makes this problem intractable at indexing time.", "source": "marker_v2", "marker_block_id": "/page/3/Text/19"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0037", "section": "3.3.2. Information Disentanglement via Parsing", "page_start": 4, "page_end": 4, "type": "Text", "text": "ColParse's first stage, parsing, transforms the input space. Let \\Psi_{\\text{parse}}(D) = \\{S_1, \\dots, S_k\\} be the set of sub-images (semantic regions) derived from document D. These regions form a partition of the document's core semantic content. By the chain rule of mutual information, the information in the original document can be expressed through its", "source": "marker_v2", "marker_block_id": "/page/3/Text/21"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0038", "section": "3.3.2. Information Disentanglement via Parsing", "page_start": 5, "page_end": 5, "type": "Text", "text": "components: I(D;R) = I(S_1,S_2,\\ldots,S_k;R) . We posit the Semantic Concentration Axiom : for any given query Q, the relevance signal R is predominantly determined by a single primary semantic region S_{j^*} \\in \\{S_j\\} , where j^* is the index of the most relevant region. This implies near conditional independence for the remaining regions: I(S_{\\neg j^*};R|S_{j^*})\\approx 0 , where S_{\\neg j^*}=\\{S_j\\}_{j\\neq j^*} . This axiom leads to the approximation: I(D;R)\\approx I(S_{j^*};R) . This decomposition transforms the problem from compressing monolithic variable D to creating a set of representations, one for each potential primary channel S_j . This provides a structural prior to IB problem, justifying creation of a multivector set \\{g_j(S_j)\\}_{j=1}^k instead of a single vector g(D).", "source": "marker_v2", "marker_block_id": "/page/4/Text/1"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0039", "section": "3.3.2. Information Disentanglement via Parsing", "page_start": 5, "page_end": 5, "type": "Text", "text": "271272", "source": "marker_v2", "marker_block_id": "/page/4/Text/66"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0040", "section": "3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION", "page_start": 5, "page_end": 5, "type": "Text", "text": "ColParse generates two sets of intermediate representations: local vectors \\{V_j = \\Phi_{\\mathrm{enc}}(S_j)\\}_{j=1}^k and a global vector V_{\\mathrm{global}} = \\Phi_{\\mathrm{enc}}(D) . Each of these encoding steps is itself an information bottleneck, subject to the Data Processing Inequality (DPI): I(V_j;R) \\leq I(S_j;R), \\quad \\forall j \\in \\{1,\\dots,k\\}; I(V_{\\mathrm{global}};R) \\leq I(D;R) . The core of ColParse is the fusion step, which creates the final representation set \\{Z_j = V_j + V_{\\mathrm{global}}\\}_{j=1}^k . To analyze its benefit, consider the joint information held by the local-global pair (V_j, V_{\\mathrm{global}}) about the relevance R. Using the chain rule:", "source": "marker_v2", "marker_block_id": "/page/4/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0041", "section": "3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION", "page_start": 5, "page_end": 5, "type": "Equation", "text": "I(V_j,V_{\\rm global};R) = I(V_j;R) + \\underbrace{I(V_{\\rm global};R|V_j)}_{\\mbox{Contextual Information Gain}} \\ . \\ \\ (4)", "source": "marker_v2", "marker_block_id": "/page/4/Equation/4"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0042", "section": "3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION", "page_start": 5, "page_end": 5, "type": "Text", "text": "This gain term quantifies the new information about relevance that the global context provides, given the local region. It is non-zero if global context helps disambiguate local content. The fusion function f(V_j,V_{\\mathrm{global}})=V_j+V_{\\mathrm{global}} creates the vector Z_j . By DPI, this fusion is also a bottleneck:", "source": "marker_v2", "marker_block_id": "/page/4/Text/5"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0043", "section": "3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION", "page_start": 5, "page_end": 5, "type": "Equation", "text": "I(Z_i; R) = I(V_i + V_{\\text{global}}; R) \\le I(V_i, V_{\\text{global}}; R). (5)", "source": "marker_v2", "marker_block_id": "/page/4/Equation/6"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0044", "section": "3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION", "page_start": 5, "page_end": 5, "type": "Text", "text": "The objective of this fusion is to craft a compact vector Z_j that is more informative than the local vector V_j alone, by capturing a significant portion of the contextual gain. The net improvement in information for region j is:", "source": "marker_v2", "marker_block_id": "/page/4/Text/7"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0045", "section": "3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION", "page_start": 5, "page_end": 5, "type": "Equation", "text": "\\Delta I_j = I(Z_j; R) - I(V_j; R). \\tag{6}", "source": "marker_v2", "marker_block_id": "/page/4/Equation/8"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0046", "section": "3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION", "page_start": 5, "page_end": 5, "type": "Text", "text": "The success of ColParse relies on this fusion being effective, i.e., ensuring \\Delta I_j > 0 . This holds when the fusion operation successfully encodes the contextual information gain. The simple vector addition serves as a parameter-free, computationally efficient mechanism to achieve this. The final representation \\mathbf{D}_{\\text{ColParse}} = \\{Z_j\\}_{j=1}^k is thus a set of contextually-refined, disentangled vectors that more effectively preserve query-relevant information, providing a superior solution to the VDR compression problem.", "source": "marker_v2", "marker_block_id": "/page/4/Text/9"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0047", "section": "3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION", "page_start": 5, "page_end": 5, "type": "Text", "text": "See more theoretical analysis in Appendix B.", "source": "marker_v2", "marker_block_id": "/page/4/Text/10"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0048", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Benchmarks and Evaluation. To ensure a comprehensive evaluation, we assess the performance of ColParse across five mainstream VDR benchmark suites, encompassing a total of 24 diverse datasets: ViDoRe-V1 (Faysse et al., 2024), ViDoRe-V2 (Macé et al., 2025), VisRAG (Yu et al., 2024), ViDoSeek (Wang et al., 2025), and MMLongBench (Ma et al., 2024b). We validate the versatility and plug-andplay nature of our framework by applying it to ten prominent single-vector retrieval models: VLM2Vec-V1-2B/7B (Jiang et al., 2024), VLM2Vec-V2-2B (Meng et al., 2025), LamRA-Ret (Liu et al., 2025a), GME-2B/7B (Zhang et al., 2024b), UniME-V2-2B/7B (Gu et al., 2025), B3-2B/7B (Thirukovalluru et al., 2025). See details of benchmarks and models in Appendix C.1 and C.2, respectively. In alignment with established VDR practices (Wasserman et al., 2025; ILLUIN, 2025), we use the Normalized Discounted Cumulative Gain at 5 (nDCG@5) as the evaluation metric. Baselines. To demonstrate the superiority of ColParse, we compare it against a diverse set of baselines organized into five distinct categories. • Base: This category represents the original performance of the ten single-vector models without any multi-vector adaptation. ② Multi-img: In this approach, all sub-images parsed from a document are fed simultaneously into the base models, leveraging their native support for multi-image inputs to generate a single representative vector. 3 Chunking-layout: This category explores strategies that operate directly on the final layer of token embeddings from the base model. Guided by the layout parsing results, these tokens are chunked and aggregated, while any tokens outside the parsed bounding boxes are discarded. We evaluate four variants: type-cluster (tokens of the same content type are merged via semantic clustering), type-mean (tokens of the same content type are merged via mean pooling), subimg-cluster (tokens from the same sub-image region are clustered), and subimg-mean (tokens from the same sub-image region are pooled). 4 Chunking-semantic: In contrast to layoutguided methods, this baseline performs hierarchical clustering on the entire set of final-layer tokens to generate a fixed number of vectors (defaulting to 10), operating agnostically to the document's layout structure. 6 Single2multi: This category mimics ColParse by encoding each parsed sub-image separately but employs alternative scoring mechanisms instead of our vector fusion. The two variants are scoring-add and scoring-multiply, where the final document score is computed by either adding or multiplying the individual query-sub_img similarity scores with the global_img-sub_img similarity scores, respectively.", "source": "marker_v2", "marker_block_id": "/page/4/Text/13"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0049", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Implementation Details. To ensure a fair comparison and reproducibility, our entire evaluation pipeline is built", "source": "marker_v2", "marker_block_id": "/page/4/Text/14"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0050", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "FigureGroup", "text": "Figure 4. The performance comparison (evaluated by nDCG@5) between ColParse and baselines on five VDR benchmarks across ten mainstream single-vector multimodal retrieval models. Refer to Table 2 and Table 3 for detailed result records due to the space limit.", "source": "marker_v2", "marker_block_id": "/page/5/FigureGroup/402"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0051", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "Text", "text": "upon the MMEB codebase 1 . We employ MinerU2.5 2 (Niu et al., 2025) as the unified document parsing model across all experiments due to its state-of-the-art performance and efficiency (See details of MinerU2.5 in Appendix C.3) . For ColParse , we choose α ranging from 0.1 to 0.9 with an interval of 0.1, and we select the optimal hyperparameter for each base model to report the final results (validated in Section 4.2.3) . Experiments were conducted on a cluster of NVIDIA A100 (80G) GPUs. The complete codebase will be made publicly available upon acceptance.", "source": "marker_v2", "marker_block_id": "/page/5/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0052", "section": "4.2.1. MAIN RESULT", "page_start": 6, "page_end": 6, "type": "Text", "text": "ColParse consistently outperforms both single-vector base models and existing multi-vector optimization baselines across diverse benchmarks. As shown in Table 2 and Table 3 (Appendix C.4) , ColParse achieves a remarkable average nDCG@5 gain of 31.64 points for VLM2Vec-V1- 2B and 42.69 points for its 7B counterpart on the ViDoRe-V1 benchmark. Figure 4 further visualizes this superiority, where the red envelope representing ColParse consistently forms the outermost boundary across all ten mainstream models. This suggests that layout-informed sub-image representations capture critical fine-grained details that are typically \"diluted\" in monolithic global embeddings.", "source": "marker_v2", "marker_block_id": "/page/5/Text/6"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0053", "section": "4.2.1. MAIN RESULT", "page_start": 6, "page_end": 6, "type": "Text", "text": "The framework demonstrates exceptional versatility and robustness as a training-free, plug-and-play module for various VLM-based embeddings. Across ten distinct models including VLM2Vec, GME, UniME, and B3, ColParse consistently yields performance improvements regardless of the model's architecture or parameter scale. Notably, even for the high-performing GME-7B model, ColParse maintains state-of-the-art results while other optimization strategies like semantic chunking (c-sem) cause a drastic performance drop from 89.36 to 23.21 on ViDoRe-V1. This universality implies that layout awareness is a fundamental, model-agnostic enhancement for visual document understanding that can be unlocked at representation level.", "source": "marker_v2", "marker_block_id": "/page/5/Text/8"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0054", "section": "4.2.1. MAIN RESULT", "page_start": 6, "page_end": 6, "type": "Text", "text": "Layout-informed decomposition is substantially more effective for multi-vector construction than traditional token-level chunking or clustering. Quantitative results reveal that baselines like cl-t-c (type-clustering) and cl-s-m (sub-image mean pooling) often lead to significant performance degradation; for instance, VLM2Vec-V2-2B's performance plummets from 74.16 to 24.85 when using token-level clustering. In contrast, ColParse boosts the same model to 78.41 by preserving the visual integrity of semantic regions rather than aggregating abstract token embeddings. This phenomenon indicates that maintaining the raw visual-semantic alignment within parsed regions is superior to post-hoc heuristic aggregation of late-stage features.", "source": "marker_v2", "marker_block_id": "/page/5/Text/9"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0055", "section": "4.2.1. MAIN RESULT", "page_start": 6, "page_end": 6, "type": "Text", "text": "ColParse exhibits superior efficacy in handling complex, long-form documents that require multi-hop reasoning.", "source": "marker_v2", "marker_block_id": "/page/5/Text/10"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0056", "section": "4.2.1. MAIN RESULT", "page_start": 6, "page_end": 6, "type": "Footnote", "text": "1 2", "source": "marker_v2", "marker_block_id": "/page/5/Footnote/7"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0057", "section": "4.2.1. MAIN RESULT", "page_start": 7, "page_end": 7, "type": "Caption", "text": "Figure 5. Variant study of ColParse and its variants.", "source": "marker_v2", "marker_block_id": "/page/6/Caption/2"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0058", "section": "4.2.1. MAIN RESULT", "page_start": 7, "page_end": 7, "type": "Text", "text": "On the MMLongBench dataset (Table 2) , ColParse elevates the average nDCG@5 of VLM2Vec-V1-2B from 25.93 to 32.07 and UniME-V2-2B from 29.31 to 44.21, outperforming all other compression baselines. This performance leap is particularly evident in the cases which require cross-page information locating. We speculate that by intelligently segmenting pages into key layout components ( e.g., tables, figures), ColParse reduces the \"signal-to-noise\" ratio during the late-interaction phase, allowing the model to focus on the most semantically salient regions.", "source": "marker_v2", "marker_block_id": "/page/6/Text/3"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0059", "section": "4.2.1. MAIN RESULT", "page_start": 7, "page_end": 7, "type": "Text", "text": "The global-local fusion strategy is critical for providing necessary contextual grounding to isolated semantic regions. Comparative analysis with the single2multi (s2m-add/mul) baselines in Table 3 shows ColParse consistently leads on challenging tasks like VisRAG and ViDoSeek, achieving 51.96 on VisRAG with VLM2Vec-V1- 2B compared to only 38.94 for s2m-add. This significant gap highlights local sub-images alone often lack the holistic context ( e.g., a table without its preceding paragraph's context) needed for accurate retrieval. The synergy achieved through weighted fusion allows ColParse to retain both regional specificity and page-level semantics.", "source": "marker_v2", "marker_block_id": "/page/6/Text/4"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0060", "section": "4.2.2. VARIANT STUDY", "page_start": 7, "page_end": 7, "type": "Text", "text": "We evaluate three specific variants: (i) single2multi (s2m) variant serves as the baseline layout-decomposed representation, which extracts and encodes sub-images as independent vectors without any global information. (ii) single2multitype-cluster (s2m-t-c) variant extends s2m by aggregating all sub-image embeddings of the same semantic category into a single representative type-level vector to further compress the representation. (iii) single2multi-global-inclusion (s2m-g-i) variant appends the original holistic page-level embedding to the s2m sub-image set as a separate global vector to provide page-wide context.", "source": "marker_v2", "marker_block_id": "/page/6/Text/6"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0061", "section": "4.2.2. VARIANT STUDY", "page_start": 7, "page_end": 7, "type": "FigureGroup", "text": "Figure 6. The comparison of the average performance of ColParse across different balancing factors. The dash lines refer to the base results; and the star points refer to the best-performing balancing factors. See Figure 10 for the model-level comparisons.", "source": "marker_v2", "marker_block_id": "/page/6/FigureGroup/570"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0062", "section": "4.2.2. VARIANT STUDY", "page_start": 7, "page_end": 7, "type": "Text", "text": "Synergistic global-local fusion is significantly more effective than simple global vector inclusion for contextualizing layout-aware representations. As shown in Figure 5, ColParse consistently outperforms the s2m-g-i variant across all ten base models and benchmarks, for instance, achieving a gain of 2.44 points in nDCG@5 over s2m-g-i for VLM2Vec-V1-2B on ViDoRe-V1. This superiority is further visualized in the radar plots of Figure 9 (Appendix C.5) , where ColParse (red envelope) consistently covers the largest area, particularly in dense tasks like VisRAG where it leads s2m-g-i by over 2 points on the 7B model. This suggests that element-wise fusion serves as a deeper conditioning mechanism than simple inclusion, allowing local features to be fundamentally \"re-weighted\" by the global semantic environment of the document.", "source": "marker_v2", "marker_block_id": "/page/6/Text/9"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0063", "section": "4.2.2. VARIANT STUDY", "page_start": 7, "page_end": 7, "type": "Text", "text": "See more analysis in Appendix C.5 due to the space limit.", "source": "marker_v2", "marker_block_id": "/page/6/Text/10"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0064", "section": "4.2.3. HYPERPARAMETER ANALYSIS", "page_start": 7, "page_end": 7, "type": "Text", "text": "Effect of balancing factor. We investigate the sensitivity of ColParse to the balancing factor α, as illustrated in Figure 6 (details in Appendix C.6.1) . First, the retrieval performance is robust across the entire range of α ∈ [0.1, 0.9], consistently surpassing the single-vector baseline for all tested models. For instance, in VLM2Vec-V1-2B, even the lowest α of 0.1 achieves an overall score of 35.97, which is still a substantial improvement over the 31.18 base result. Second, the performance typically follows a convex trajectory, with the optimal balance point generally leaning towards the global context to effectively ground the isolated layout components. Quantitatively, most models reach their peak performance within the range of α ∈ [0.6, 0.8]. Finally, a moderate synergistic fusion is essential, as either excessive local specificity or excessive global dominance leads to suboptimal results. In B3-7B, performance steadily climbs from 62.09 at α = 0.1 to its peak of 68.51 at α = 0.7, before exhibiting a slight decline as the representation becomes overly dominated by the global vector at α = 0.9.", "source": "marker_v2", "marker_block_id": "/page/6/Text/12"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0065", "section": "4.2.3. HYPERPARAMETER ANALYSIS", "page_start": 8, "page_end": 8, "type": "Caption", "text": "Figure 7. Comparison of MinerU2.5 against its counterparts. The y-axis represents the overall score of OmniDocBench (Ouyang et al., 2025) , the x-axis shows end-to-end throughput (Pages/Sec), and bubble size indicates the parameter size.", "source": "marker_v2", "marker_block_id": "/page/7/Caption/2"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0066", "section": "4.2.3. HYPERPARAMETER ANALYSIS", "page_start": 8, "page_end": 8, "type": "TableGroup", "text": "Table 1. Efficiency analysis of ColParse on the best performing model GME-7B and its multi-vector counterpart (w/ original setting). Performance denotes the overall score of MMEB-visdoc, Storage refers to the number of vectors stored per document, and Latency represents the average encoding time per document. * denotes the multi-vector model is trained w/ aligned configuration. Orange arrows denote better and blue ones denote worse. Model Performance Storage Latency GME-7b 79.50 1.00 0.30 +ColParse 80.61↑1.11 5.90↑4.9 0.81↑0.51 ColQwen* 80.02↑0.52 768.00↑767 0.41↑0.11", "source": "marker_v2", "marker_block_id": "/page/7/TableGroup/590"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0067", "section": "4.2.3. HYPERPARAMETER ANALYSIS", "page_start": 8, "page_end": 8, "type": "Text", "text": "Effect of document parsing model. We evaluate the capability of various document parsing models and conclude that MinerU2.5 provides the optimal balance between parsing fidelity and inference efficiency, as shown in Figure 7 (Details in Appendix C.6.2) . First, MinerU2.5 demonstrates superior parsing accuracy across all semantic and structural dimensions compared to existing specialized VLMs. Quantitatively, it achieves a state-of-the-art Overall score of 90.67 on OmniDocBench (Ouyang et al., 2025) , significantly outperforming the best baseline MonkeyOCR-pro-3B (88.85) and maintaining the lowest error rates in both text (0.047) and reading order (0.044) recognition. Second, MinerU2.5 achieves industrial-grade throughput, ensuring the practicality of the ColParse pipeline for large-scale document corpora. For example, it delivers an end-to-end processing speed of 2.25 Pages/sec, which is 4× faster than highparameter models like Nanonets-OCR-s (0.55 Pages/sec). Consequently, we select MinerU2.5 as our unified parser as it sits at the optimal Pareto front, offering the highest parsing quality while sustaining a remarkable throughput.", "source": "marker_v2", "marker_block_id": "/page/7/Text/5"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0068", "section": "4.2.4. EFFICIENCY ANALYSIS", "page_start": 8, "page_end": 8, "type": "Text", "text": "To ensure a fair comparison, we evaluate ColParse using GME-7B, the best-performing single-vector model, and ColQwen, a multi-vector baseline with an aligned architecture and training configuration. We conclude that our", "source": "marker_v2", "marker_block_id": "/page/7/Text/7"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0069", "section": "4.2.4. EFFICIENCY ANALYSIS", "page_start": 8, "page_end": 8, "type": "Text", "text": "paradigm achieves superior retrieval performance while drastically slashing the storage overhead inherent in traditional multi-vector models. Specifically, as shown in Table 1, ColParse yields an overall score of 80.61 on MMEB-visdoc, outperforming the multi-vector counterpart ColQwen (80.02) while requiring only 5.9 vectors per page (See per-dataset records in Appendix C.6.3) , a massive storage reduction of over 99% compared to ColQwen's 768 vectors. Furthermore, although ColParse introduces a marginal increase in encoding latency, it remains a highly practical solution for real-world deployment. While the per-document latency rises to 0.81s, it is still significantly lower than the average 7s delay of conventional OCR-based pipelines, and the offline parsing stage can be optimized through parallel processing to mitigate indexing overhead.", "source": "marker_v2", "marker_block_id": "/page/7/Text/8"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0070", "section": "4.2.5. CASE STUDY", "page_start": 8, "page_end": 8, "type": "FigureGroup", "text": "Figure 8. The illustration of a representative case.", "source": "marker_v2", "marker_block_id": "/page/7/FigureGroup/591"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0071", "section": "4.2.5. CASE STUDY", "page_start": 8, "page_end": 8, "type": "Text", "text": "We conduct a case study in Figure 8 to illustrate the interpretability of ColParse beyond its superior retrieval performance. When handling a query requiring specific details, such as carbon emission reduction percentages, ColParse not only retrieves the correct document page but also pinpoints the exact parsed sub-vector ( e.g., a specific text block) that yields the highest MaxSim score. This layout-informed granularity enables fine-grained information back-tracing, allowing the system to present the specific evidence region directly to the user. Such explainability is highly valuable in practical industrial scenarios, such as financial auditing or legal review, where the ability to verify the source of information is as critical as retrieval accuracy.", "source": "marker_v2", "marker_block_id": "/page/7/Text/12"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0072", "section": "5. Conclusion", "page_start": 8, "page_end": 8, "type": "Text", "text": "In this paper, we introduced ColParse , a novel layoutinformed paradigm designed to overcome the critical storage efficiency bottleneck in multi-vector VDR. Our framework uniquely generates a compact multi-vector representation by fusing layout-aware sub-image embeddings from a document parser with a holistic global vector. Extensive experiments demonstrated that ColParse achieves over 95% storage compression while consistently improving retrieval performance across base models and datasets. Ultimately, ColParse charts a new course for the field, establishing that a deep understanding of document structure is critical for practical multimodal information systems.", "source": "marker_v2", "marker_block_id": "/page/7/Text/14"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0073", "section": "Impact Statement", "page_start": 9, "page_end": 9, "type": "Text", "text": "Ethical Considerations. We believe that our proposed ColParse framework raises no new ethical concerns. Its motivation is to advance the efficiency and performance of VDR systems in a principled and resource-conscious manner. By leveraging existing document parsing technologies to create more compact and interpretable representations, our method promotes responsible AI development, adhering to established ethical guidelines in information retrieval research without relying on sensitive or proprietary data.", "source": "marker_v2", "marker_block_id": "/page/8/Text/2"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0074", "section": "Impact Statement", "page_start": 9, "page_end": 9, "type": "Text", "text": "Societal Implications. ColParse introduces a new paradigm for multimodal information retrieval by shifting from storage-intensive, grid-based representations to a lightweight, layout-informed approach. It fundamentally resolves the critical conflict between fine-grained retrieval accuracy and the practical storage costs of large-scale deployment. By reducing storage requirements by over 95% while simultaneously enhancing performance, ColParse significantly lowers the barrier for deploying state-of-the-art visual document understanding systems. This has the potential to democratize access to advanced information retrieval in diverse domains, including academic research, enterprise knowledge management, and digital archives. Furthermore, its inherent interpretability, which links retrieval results to specific document components, fosters greater transparency and user trust in AI-powered information systems.", "source": "marker_v2", "marker_block_id": "/page/8/Text/3"}

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/marker_meta.json

@@ -0,0 +1,2949 @@
+{
+  "table_of_contents": [
+    {
+      "title": "()()()\n001\n011\n012",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          23.25,
+          41.765625
+        ],
+        [
+          41.25,
+          41.765625
+        ],
+        [
+          41.25,
+          505.828125
+        ],
+        [
+          23.25,
+          505.828125
+        ]
+      ]
+    },
+    {
+      "title": "006\n007",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          23.25,
+          90.87890625
+        ],
+        [
+          41.8359375,
+          90.87890625
+        ],
+        [
+          41.8359375,
+          164.35546875
+        ],
+        [
+          23.25,
+          164.35546875
+        ]
+      ]
+    },
+    {
+      "title": "002\n003\n004\n005",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          23.25,
+          81.2109375
+        ],
+        [
+          41.25,
+          81.2109375
+        ],
+        [
+          41.25,
+          139.9921875
+        ],
+        [
+          23.25,
+          139.9921875
+        ]
+      ]
+    },
+    {
+      "title": "008\n009",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          23.25,
+          155.84765625
+        ],
+        [
+          41.25,
+          155.84765625
+        ],
+        [
+          41.25,
+          201.48046875
+        ],
+        [
+          23.25,
+          201.48046875
+        ]
+      ]
+    },
+    {
+      "title": "014\n015\n016",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          23.25,
+          217.72265625
+        ],
+        [
+          45.720703125,
+          217.72265625
+        ],
+        [
+          45.720703125,
+          286.55859375
+        ],
+        [
+          23.25,
+          286.55859375
+        ]
+      ]
+    },
+    {
+      "title": "018\n022",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          23.25,
+          273.796875
+        ],
+        [
+          44.82421875,
+          273.796875
+        ],
+        [
+          44.82421875,
+          339.75
+        ],
+        [
+          23.25,
+          339.75
+        ]
+      ]
+    },
+    {
+      "title": "024\n025\n026",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          21.515625,
+          334.8984375
+        ],
+        [
+          42.0,
+          334.8984375
+        ],
+        [
+          42.0,
+          392.1328125
+        ],
+        [
+          21.515625,
+          392.1328125
+        ]
+      ]
+    },
+    {
+      "title": "028",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          21.515625,
+          390.19921875
+        ],
+        [
+          41.25,
+          390.19921875
+        ],
+        [
+          41.25,
+          438.92578125
+        ],
+        [
+          21.515625,
+          438.92578125
+        ]
+      ]
+    },
+    {
+      "title": "031",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          22.7109375,
+          423.0703125
+        ],
+        [
+          41.25,
+          423.0703125
+        ],
+        [
+          41.25,
+          474.1171875
+        ],
+        [
+          22.7109375,
+          474.1171875
+        ]
+      ]
+    },
+    {
+      "title": "034\n035",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          23.25,
+          467.15625
+        ],
+        [
+          41.25,
+          467.15625
+        ],
+        [
+          41.25,
+          510.46875
+        ],
+        [
+          23.25,
+          510.46875
+        ]
+      ]
+    },
+    {
+      "title": "038",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          23.25,
+          514.3359375
+        ],
+        [
+          41.8359375,
+          514.3359375
+        ],
+        [
+          41.8359375,
+          556.1015625
+        ],
+        [
+          23.25,
+          556.1015625
+        ]
+      ]
+    },
+    {
+      "title": "041\n043\n044\n045\n046\n047",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          23.25,
+          556.48828125
+        ],
+        [
+          42.134765625,
+          556.48828125
+        ],
+        [
+          42.134765625,
+          638.25
+        ],
+        [
+          23.25,
+          638.25
+        ]
+      ]
+    },
+    {
+      "title": "051\n052\n053",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          23.009765625,
+          667.86328125
+        ],
+        [
+          42.0,
+          667.86328125
+        ],
+        [
+          42.0,
+          711.94921875
+        ],
+        [
+          23.009765625,
+          711.94921875
+        ]
+      ]
+    },
+    {
+      "title": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval\nwith Parsed Visual Document Representations",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          117.0,
+          89.25
+        ],
+        [
+          479.3203125,
+          89.25
+        ],
+        [
+          479.3203125,
+          121.04296875
+        ],
+        [
+          117.0,
+          121.04296875
+        ]
+      ]
+    },
+    {
+      "title": "Anonymous Authors<sup>1</sup>",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          241.5,
+          159.71484375
+        ],
+        [
+          334.6875,
+          159.71484375
+        ],
+        [
+          334.6875,
+          169.76953125
+        ],
+        [
+          242.25,
+          171.0
+        ]
+      ]
+    },
+    {
+      "title": "Abstract",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          149.25,
+          192.97265625
+        ],
+        [
+          195.75,
+          192.97265625
+        ],
+        [
+          195.75,
+          203.25
+        ],
+        [
+          149.25,
+          203.25
+        ]
+      ]
+    },
+    {
+      "title": "1. Introduction",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          54.0,
+          533.25
+        ],
+        [
+          132.75,
+          533.25
+        ],
+        [
+          132.75,
+          543.0
+        ],
+        [
+          54.0,
+          543.0
+        ]
+      ]
+    },
+    {
+      "title": "2. Related Work",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          306.59765625,
+          149.66015625
+        ],
+        [
+          391.5,
+          149.66015625
+        ],
+        [
+          391.5,
+          159.75
+        ],
+        [
+          306.59765625,
+          159.75
+        ]
+      ]
+    },
+    {
+      "title": "2.1. Visual Document Retrieval",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          303.609375,
+          170.15625
+        ],
+        [
+          440.25,
+          170.15625
+        ],
+        [
+          440.25,
+          180.0
+        ],
+        [
+          303.609375,
+          180.0
+        ]
+      ]
+    },
+    {
+      "title": "2.2. Mutli-Vector Retrieval",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          306.0,
+          465.75
+        ],
+        [
+          422.25,
+          465.75
+        ],
+        [
+          422.25,
+          474.50390625
+        ],
+        [
+          306.0,
+          474.50390625
+        ]
+      ]
+    },
+    {
+      "title": "2.3. Document Parsing VLM",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          52.892578125,
+          200.70703125
+        ],
+        [
+          178.5,
+          200.70703125
+        ],
+        [
+          178.5,
+          210.76171875
+        ],
+        [
+          52.892578125,
+          210.76171875
+        ]
+      ]
+    },
+    {
+      "title": "3. Methodology",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          53.490234375,
+          475.27734375
+        ],
+        [
+          135.0,
+          475.27734375
+        ],
+        [
+          135.0,
+          484.5
+        ],
+        [
+          53.490234375,
+          484.5
+        ]
+      ]
+    },
+    {
+      "title": "3.1. Task Formulation",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          51.99609375,
+          568.08984375
+        ],
+        [
+          149.25,
+          568.08984375
+        ],
+        [
+          149.25,
+          577.5
+        ],
+        [
+          51.99609375,
+          577.5
+        ]
+      ]
+    },
+    {
+      "title": "3.2. The ColParse Framework",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          306.0,
+          453.62109375
+        ],
+        [
+          439.5,
+          453.62109375
+        ],
+        [
+          439.5,
+          463.5
+        ],
+        [
+          306.0,
+          463.5
+        ]
+      ]
+    },
+    {
+      "title": "3.2.1. LAYOUT-INFORMED DOCUMENT PARSING",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          306.0,
+          592.5
+        ],
+        [
+          514.5,
+          592.5
+        ],
+        [
+          514.5,
+          601.5
+        ],
+        [
+          306.0,
+          601.5
+        ]
+      ]
+    },
+    {
+      "title": "3.2.2. DUAL-STREAM ENCODING",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          52.59375,
+          224.25
+        ],
+        [
+          198.0,
+          224.25
+        ],
+        [
+          198.0,
+          233.25
+        ],
+        [
+          52.59375,
+          233.25
+        ]
+      ]
+    },
+    {
+      "title": "3.2.3. GLOBAL-LOCAL FUSION FOR FINAL REPRESENTATION",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          54.0,
+          522.75
+        ],
+        [
+          238.5,
+          522.75
+        ],
+        [
+          238.5,
+          545.66015625
+        ],
+        [
+          54.0,
+          545.66015625
+        ]
+      ]
+    },
+    {
+      "title": "3.2.4. LATE-INTERACTION SCORING WITH ColParse",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          305.40234375,
+          129.55078125
+        ],
+        [
+          534.0,
+          129.55078125
+        ],
+        [
+          534.0,
+          138.0
+        ],
+        [
+          305.40234375,
+          138.0
+        ]
+      ]
+    },
+    {
+      "title": "3.3. Theoretical Foundation",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          304.505859375,
+          325.5
+        ],
+        [
+          426.0,
+          325.5
+        ],
+        [
+          426.0,
+          334.5
+        ],
+        [
+          304.505859375,
+          334.5
+        ]
+      ]
+    },
+    {
+      "title": "3.3.1. THE VDR COMPRESSION PROBLEM AS AN INFORMATION BOTTLENECK",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          306.75,
+          426.9375
+        ],
+        [
+          516.75,
+          426.9375
+        ],
+        [
+          516.75,
+          448.5
+        ],
+        [
+          306.75,
+          448.5
+        ]
+      ]
+    },
+    {
+      "title": "3.3.2. Information Disentanglement via Parsing",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          305.40234375,
+          626.87109375
+        ],
+        [
+          542.25,
+          626.87109375
+        ],
+        [
+          542.25,
+          637.5
+        ],
+        [
+          305.40234375,
+          637.5
+        ]
+      ]
+    },
+    {
+      "title": "3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          54.0,
+          235.8984375
+        ],
+        [
+          281.25,
+          235.8984375
+        ],
+        [
+          281.25,
+          257.25
+        ],
+        [
+          54.0,
+          257.25
+        ]
+      ]
+    },
+    {
+      "title": "4. Experiment",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          306.75,
+          68.44921875
+        ],
+        [
+          381.3046875,
+          68.44921875
+        ],
+        [
+          381.3046875,
+          78.50390625
+        ],
+        [
+          306.75,
+          78.50390625
+        ]
+      ]
+    },
+    {
+      "title": "4.1. Experimental Setup",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          303.908203125,
+          88.171875
+        ],
+        [
+          410.25,
+          88.171875
+        ],
+        [
+          410.25,
+          98.25
+        ],
+        [
+          303.908203125,
+          98.25
+        ]
+      ]
+    },
+    {
+      "title": "4.2. Experimental Analysis",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          53.7890625,
+          498.72149658203125
+        ],
+        [
+          169.46194458007812,
+          498.72149658203125
+        ],
+        [
+          169.46194458007812,
+          508.68408203125
+        ],
+        [
+          53.7890625,
+          508.68408203125
+        ]
+      ]
+    },
+    {
+      "title": "4.2.1. MAIN RESULT",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          53.19140625,
+          517.4654541015625
+        ],
+        [
+          144.2446746826172,
+          517.4654541015625
+        ],
+        [
+          144.2446746826172,
+          527.4280700683594
+        ],
+        [
+          53.19140625,
+          527.4280700683594
+        ]
+      ]
+    },
+    {
+      "title": "4.2.2. VARIANT STUDY",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          53.19140625,
+          568.08984375
+        ],
+        [
+          154.822998046875,
+          568.08984375
+        ],
+        [
+          154.822998046875,
+          579.2640686035156
+        ],
+        [
+          53.19140625,
+          579.2640686035156
+        ]
+      ]
+    },
+    {
+      "title": "4.2.3. HYPERPARAMETER ANALYSIS",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          307.1953125,
+          472.18359375
+        ],
+        [
+          463.28118896484375,
+          472.18359375
+        ],
+        [
+          463.28118896484375,
+          483.6220703125
+        ],
+        [
+          307.1953125,
+          483.6220703125
+        ]
+      ]
+    },
+    {
+      "title": "4.2.4. EFFICIENCY ANALYSIS",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          51.99609375,
+          651.62109375
+        ],
+        [
+          180.4867401123047,
+          651.62109375
+        ],
+        [
+          180.4867401123047,
+          662.9500732421875
+        ],
+        [
+          51.99609375,
+          662.9500732421875
+        ]
+      ]
+    },
+    {
+      "title": "4.2.5. CASE STUDY",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          307.6889953613281,
+          259.1015625
+        ],
+        [
+          390.9485778808594,
+          259.1015625
+        ],
+        [
+          390.9485778808594,
+          270.49603271484375
+        ],
+        [
+          307.6889953613281,
+          270.49603271484375
+        ]
+      ]
+    },
+    {
+      "title": "5. Conclusion",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          307.44000244140625,
+          560.6741638183594
+        ],
+        [
+          376.52911376953125,
+          560.6741638183594
+        ],
+        [
+          376.52911376953125,
+          572.6293640136719
+        ],
+        [
+          307.44000244140625,
+          572.6293640136719
+        ]
+      ]
+    },
+    {
+      "title": "Impact Statement",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          51.99609375,
+          66.515625
+        ],
+        [
+          146.74185180664062,
+          66.515625
+        ],
+        [
+          146.74185180664062,
+          79.80230712890625
+        ],
+        [
+          51.99609375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "References",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          54.087890625,
+          415.6041564941406
+        ],
+        [
+          110.9838638305664,
+          415.6041564941406
+        ],
+        [
+          110.9838638305664,
+          427.5593566894531
+        ],
+        [
+          54.087890625,
+          427.5593566894531
+        ]
+      ]
+    },
+    {
+      "title": "A. Algorithm Workflow",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          53.7890625,
+          68.0625
+        ],
+        [
+          176.25,
+          68.0625
+        ],
+        [
+          176.25,
+          78.890625
+        ],
+        [
+          53.7890625,
+          78.890625
+        ]
+      ]
+    },
+    {
+      "title": "Algorithm 1 The Offline Indexing Process of ColParse",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          51.99609375,
+          206.25
+        ],
+        [
+          285.75,
+          206.25
+        ],
+        [
+          285.75,
+          216.94921875
+        ],
+        [
+          51.99609375,
+          216.94921875
+        ]
+      ]
+    },
+    {
+      "title": "B. More Theoretical Analysis",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          52.59375,
+          579.75
+        ],
+        [
+          205.5,
+          579.75
+        ],
+        [
+          205.5,
+          589.5
+        ],
+        [
+          52.59375,
+          589.5
+        ]
+      ]
+    },
+    {
+      "title": "B.1. Information-Theoretic Preliminaries",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          52.59375,
+          660.0
+        ],
+        [
+          231.0,
+          660.0
+        ],
+        [
+          231.0,
+          669.75
+        ],
+        [
+          52.59375,
+          669.75
+        ]
+      ]
+    },
+    {
+      "title": "return score",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          305.103515625,
+          317.25
+        ],
+        [
+          360.75,
+          317.25
+        ],
+        [
+          360.75,
+          325.23046875
+        ],
+        [
+          305.103515625,
+          325.23046875
+        ]
+      ]
+    },
+    {
+      "title": "B.2. The Information Bottleneck (IB) Principle in VDR",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          53.19140625,
+          69.0
+        ],
+        [
+          289.5,
+          69.0
+        ],
+        [
+          289.5,
+          79.27734375
+        ],
+        [
+          53.19140625,
+          79.27734375
+        ]
+      ]
+    },
+    {
+      "title": "B.3. Justification for Structural Disentanglement",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          54.0,
+          309.375
+        ],
+        [
+          262.5,
+          309.375
+        ],
+        [
+          262.5,
+          318.75
+        ],
+        [
+          54.0,
+          318.75
+        ]
+      ]
+    },
+    {
+      "title": "B.4. Justification for Synergistic Fusion",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          305.40234375,
+          153.75
+        ],
+        [
+          474.75,
+          153.75
+        ],
+        [
+          474.75,
+          163.5
+        ],
+        [
+          305.40234375,
+          163.5
+        ]
+      ]
+    },
+    {
+      "title": "C. More Experimental Analysis",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          53.7890625,
+          67.84710693359375
+        ],
+        [
+          216.60804748535156,
+          67.84710693359375
+        ],
+        [
+          216.60804748535156,
+          79.80230712890625
+        ],
+        [
+          53.7890625,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "C.1. Benchmark Details",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          52.892578125,
+          88.74945068359375
+        ],
+        [
+          156.71978759765625,
+          88.74945068359375
+        ],
+        [
+          156.71978759765625,
+          98.7120361328125
+        ],
+        [
+          52.892578125,
+          98.7120361328125
+        ]
+      ]
+    },
+    {
+      "title": "C.2. Model Details",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          304.20703125,
+          189.10546875
+        ],
+        [
+          386.02496337890625,
+          189.10546875
+        ],
+        [
+          386.02496337890625,
+          200.11407470703125
+        ],
+        [
+          304.20703125,
+          200.11407470703125
+        ]
+      ]
+    },
+    {
+      "title": "C.3. MinerU2.5 Details",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          53.19140625,
+          375.1171875
+        ],
+        [
+          153.11331176757812,
+          375.1171875
+        ],
+        [
+          153.11331176757812,
+          385.41912841796875
+        ],
+        [
+          53.19140625,
+          385.41912841796875
+        ]
+      ]
+    },
+    {
+      "title": "C.4. Main Results",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          307.44000244140625,
+          660.90234375
+        ],
+        [
+          383.994140625,
+          660.90234375
+        ],
+        [
+          383.994140625,
+          671.4181060791016
+        ],
+        [
+          307.44000244140625,
+          671.4181060791016
+        ]
+      ]
+    },
+    {
+      "title": "C.5. Variant Study",
+      "heading_level": null,
+      "page_id": 20,
+      "polygon": [
+        [
+          51.3984375,
+          68.8359375
+        ],
+        [
+          135.0,
+          68.8359375
+        ],
+        [
+          135.0,
+          78.1171875
+        ],
+        [
+          51.3984375,
+          78.1171875
+        ]
+      ]
+    },
+    {
+      "title": "C.5.1. ALGORITHM WORKFLOW",
+      "heading_level": null,
+      "page_id": 20,
+      "polygon": [
+        [
+          52.294921875,
+          87.78515625
+        ],
+        [
+          192.75,
+          87.0
+        ],
+        [
+          192.75,
+          96.75
+        ],
+        [
+          52.294921875,
+          97.5
+        ]
+      ]
+    },
+    {
+      "title": "Algorithm 4 Integrated Offline Indexing for ColParse Vari-",
+      "heading_level": null,
+      "page_id": 20,
+      "polygon": [
+        [
+          51.99609375,
+          203.25
+        ],
+        [
+          291.0,
+          203.25
+        ],
+        [
+          291.0,
+          227.390625
+        ],
+        [
+          51.99609375,
+          227.390625
+        ]
+      ]
+    },
+    {
+      "title": "1119\n1120\n/* Stage 1: Layout-Informed Document Parsing",
+      "heading_level": null,
+      "page_id": 20,
+      "polygon": [
+        [
+          24.0,
+          295.5
+        ],
+        [
+          259.98046875,
+          295.5
+        ],
+        [
+          259.98046875,
+          316.5
+        ],
+        [
+          24.0,
+          316.5
+        ]
+      ]
+    },
+    {
+      "title": "1128\n/* Stage 2: Regional Encoding\n1129",
+      "heading_level": null,
+      "page_id": 20,
+      "polygon": [
+        [
+          24.0,
+          403.5
+        ],
+        [
+          215.15625,
+          403.5
+        ],
+        [
+          215.15625,
+          424.23046875
+        ],
+        [
+          24.0,
+          424.23046875
+        ]
+      ]
+    },
+    {
+      "title": "C.5.2. MORE ANALYSIS",
+      "heading_level": null,
+      "page_id": 20,
+      "polygon": [
+        [
+          306.75,
+          69.75
+        ],
+        [
+          410.291015625,
+          69.75
+        ],
+        [
+          410.291015625,
+          77.25
+        ],
+        [
+          306.75,
+          77.25
+        ]
+      ]
+    },
+    {
+      "title": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations",
+      "heading_level": null,
+      "page_id": 22,
+      "polygon": [
+        [
+          96.39400482177734,
+          46.79296875
+        ],
+        [
+          500.8359375,
+          46.79296875
+        ],
+        [
+          500.8359375,
+          56.1944580078125
+        ],
+        [
+          96.39400482177734,
+          56.1944580078125
+        ]
+      ]
+    },
+    {
+      "title": "Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations",
+      "heading_level": null,
+      "page_id": 24,
+      "polygon": [
+        [
+          96.39400482177734,
+          47.1796875
+        ],
+        [
+          500.8359375,
+          47.1796875
+        ],
+        [
+          500.8359375,
+          56.1944580078125
+        ],
+        [
+          96.39400482177734,
+          56.1944580078125
+        ]
+      ]
+    },
+    {
+      "title": "C.6. Hyperparameter Study",
+      "heading_level": null,
+      "page_id": 25,
+      "polygon": [
+        [
+          53.19140625,
+          69.22265625
+        ],
+        [
+          174.32369995117188,
+          69.22265625
+        ],
+        [
+          174.32369995117188,
+          79.33905029296875
+        ],
+        [
+          53.19140625,
+          79.33905029296875
+        ]
+      ]
+    },
+    {
+      "title": "C.6.1. EFFECT OF BALANCING FACTOR",
+      "heading_level": null,
+      "page_id": 25,
+      "polygon": [
+        [
+          53.19140625,
+          87.3984375
+        ],
+        [
+          222.63058471679688,
+          87.3984375
+        ],
+        [
+          222.63058471679688,
+          98.0830078125
+        ],
+        [
+          53.19140625,
+          98.0830078125
+        ]
+      ]
+    },
+    {
+      "title": "C.6.2. EFFECT OF DOCUMENT PARSING MODEL",
+      "heading_level": null,
+      "page_id": 25,
+      "polygon": [
+        [
+          54.38671875,
+          580.46484375
+        ],
+        [
+          259.2910461425781,
+          580.46484375
+        ],
+        [
+          259.2910461425781,
+          591.2190856933594
+        ],
+        [
+          54.38671875,
+          591.2190856933594
+        ]
+      ]
+    },
+    {
+      "title": "C.6.3. EFFICIENCY ANALYSIS",
+      "heading_level": null,
+      "page_id": 25,
+      "polygon": [
+        [
+          307.6889953613281,
+          372.0234375
+        ],
+        [
+          434.1504821777344,
+          372.0234375
+        ],
+        [
+          434.1504821777344,
+          382.1050720214844
+        ],
+        [
+          307.6889953613281,
+          382.1050720214844
+        ]
+      ]
+    }
+  ],
+  "page_stats": [
+    {
+      "page_id": 0,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          125
+        ],
+        [
+          "Span",
+          119
+        ],
+        [
+          "SectionHeader",
+          17
+        ],
+        [
+          "Text",
+          9
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 1,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          202
+        ],
+        [
+          "Span",
+          125
+        ],
+        [
+          "Text",
+          50
+        ],
+        [
+          "PageHeader",
+          5
+        ],
+        [
+          "ListItem",
+          3
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 2,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          208
+        ],
+        [
+          "Span",
+          149
+        ],
+        [
+          "Text",
+          61
+        ],
+        [
+          "SectionHeader",
+          5
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Picture",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Equation",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "PictureGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 3,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          212
+        ],
+        [
+          "Span",
+          188
+        ],
+        [
+          "Text",
+          64
+        ],
+        [
+          "SectionHeader",
+          6
+        ],
+        [
+          "Equation",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 4,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          219
+        ],
+        [
+          "Span",
+          176
+        ],
+        [
+          "Text",
+          61
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Equation",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 5,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          277
+        ],
+        [
+          "Line",
+          102
+        ],
+        [
+          "Text",
+          16
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 6,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          427
+        ],
+        [
+          "Line",
+          119
+        ],
+        [
+          "Text",
+          16
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "Figure",
+          2
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 7,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          423
+        ],
+        [
+          "Line",
+          147
+        ],
+        [
+          "TableCell",
+          16
+        ],
+        [
+          "Text",
+          9
+        ],
+        [
+          "Caption",
+          3
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "Figure",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 8,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          354
+        ],
+        [
+          "Line",
+          153
+        ],
+        [
+          "ListItem",
+          16
+        ],
+        [
+          "Reference",
+          16
+        ],
+        [
+          "Text",
+          3
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 9,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          326
+        ],
+        [
+          "Line",
+          138
+        ],
+        [
+          "ListItem",
+          22
+        ],
+        [
+          "Reference",
+          21
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 10,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          209
+        ],
+        [
+          "Line",
+          93
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "ListItem",
+          10
+        ],
+        [
+          "Reference",
+          10
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 11,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          231
+        ],
+        [
+          "Span",
+          179
+        ],
+        [
+          "Text",
+          77
+        ],
+        [
+          "SectionHeader",
+          5
+        ],
+        [
+          "Reference",
+          5
+        ],
+        [
+          "Equation",
+          4
+        ],
+        [
+          "Code",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 12,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          208
+        ],
+        [
+          "Span",
+          174
+        ],
+        [
+          "Text",
+          61
+        ],
+        [
+          "Equation",
+          7
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 13,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          379
+        ],
+        [
+          "Line",
+          162
+        ],
+        [
+          "Text",
+          14
+        ],
+        [
+          "Reference",
+          11
+        ],
+        [
+          "ListItem",
+          8
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Footnote",
+          3
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Code",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 14,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          634
+        ],
+        [
+          "Line",
+          152
+        ],
+        [
+          "Text",
+          8
+        ],
+        [
+          "Reference",
+          5
+        ],
+        [
+          "ListItem",
+          4
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Footnote",
+          2
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Code",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 15,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          1407
+        ],
+        [
+          "TableCell",
+          672
+        ],
+        [
+          "Line",
+          103
+        ],
+        [
+          "Text",
+          8
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 16,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          1409
+        ],
+        [
+          "TableCell",
+          663
+        ],
+        [
+          "Line",
+          82
+        ],
+        [
+          "Text",
+          14
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 17,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          313
+        ],
+        [
+          "TableCell",
+          114
+        ],
+        [
+          "Line",
+          63
+        ],
+        [
+          "Text",
+          4
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 18,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          2306
+        ],
+        [
+          "TableCell",
+          1056
+        ],
+        [
+          "Line",
+          75
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "Text",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 19,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          1271
+        ],
+        [
+          "TableCell",
+          576
+        ],
+        [
+          "Line",
+          83
+        ],
+        [
+          "Text",
+          9
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 20,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Span",
+          179
+        ],
+        [
+          "Line",
+          137
+        ],
+        [
+          "Text",
+          36
+        ],
+        [
+          "Code",
+          8
+        ],
+        [
+          "SectionHeader",
+          6
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 21,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          129
+        ],
+        [
+          "Line",
+          58
+        ],
+        [
+          "Figure",
+          2
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 22,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          535
+        ],
+        [
+          "TableCell",
+          252
+        ],
+        [
+          "Line",
+          48
+        ],
+        [
+          "Text",
+          11
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 23,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          1127
+        ],
+        [
+          "TableCell",
+          505
+        ],
+        [
+          "Line",
+          90
+        ],
+        [
+          "Text",
+          3
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 24,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          1885
+        ],
+        [
+          "TableCell",
+          832
+        ],
+        [
+          "Line",
+          111
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 25,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          258
+        ],
+        [
+          "Line",
+          100
+        ],
+        [
+          "TableCell",
+          28
+        ],
+        [
+          "Text",
+          5
+        ],
+        [
+          "Reference",
+          5
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 26,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          452
+        ],
+        [
+          "TableCell",
+          141
+        ],
+        [
+          "Line",
+          96
+        ],
+        [
+          "Caption",
+          11
+        ],
+        [
+          "Text",
+          5
+        ],
+        [
+          "Table",
+          4
+        ],
+        [
+          "Picture",
+          4
+        ],
+        [
+          "TableGroup",
+          4
+        ],
+        [
+          "PictureGroup",
+          3
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    }
+  ],
+  "debug_data_path": "debug_data/3250cb92-2f69-4e16-9df9-f569224173f0"
+}

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/model_text_v3.txt

@@ -0,0 +1,224 @@
+[p. 1 | section: Abstract | type: Text]
+Harnessing the full potential of visually-rich documents requires retrieval systems that understand not just text, but intricate layouts, a core challenge in Visual Document Retrieval (VDR). The prevailing multi-vector architectures, while powerful, face a crucial storage bottleneck that current optimization strategies, such as embedding merging, pruning, or introducing abstract tokens, fail to resolve without compromising performance or ignoring vital layout cues. To address this, we introduce ColParse, a novel paradigm that leverages a document parsing model to generate a small set of layout-informed sub-image embeddings, which are then fused with a global pagelevel vector to create a compact and structurallyaware multi-vector representation. Extensive experiments demonstrate that ColParse reduces storage requirements by over 95% while simultaneously yielding significant performance gains across numerous benchmarks and base models. ColParse thus bridges the critical gap between the fine-grained accuracy of multi-vector retrieval and the practical demands of large-scale deployment, offering a new path towards efficient and interpretable multimodal information systems.
+
+[p. 1 | section: 1. Introduction | type: Text]
+Visual Document Retrieval (VDR), the task of retrieving relevant document pages from a large-scale corpus, has become a cornerstone of modern information retrieval (Mei et al., 2025; Yan et al., 2026a). Unlike natural image retrieval, visual documents, such as academic papers, financial reports, and invoices, are defined by a dense interplay of textual content, intricate layouts, and graphical elements, as illustrated in Figure 1. To effectively capture this fine-
+
+[p. 1 | section: 1. Introduction | type: FigureGroup]
+Figure 1. Comparison of natural image retrieval versus VDR.
+
+[p. 1 | section: 1. Introduction | type: Text]
+grained detail, the field has predominantly converged on multi-vector retrieval architectures (Faysse et al., 2024; Günther et al., 2025; Team, 2025). These models represent each document page as a set of patch-level embeddings and employ a late-interaction mechanism, such as MaxSim, to compute relevance (Khattab & Zaharia, 2020; Santhanam et al., 2022). This paradigm excels at aligning specific query phrases with corresponding visual or textual regions within a document, a capability essential for the high-precision information-seeking tasks inherent to VDR.
+
+[p. 1 | section: 1. Introduction | type: Text]
+Despite their superior performance, the widespread adoption of multi-vector VDR models is hindered by a critical bottleneck: prohibitive storage overhead (Jayaram et al., 2024; Shrestha et al., 2024; Liu & Mao, 2023). Storing hundreds or even thousands of embedding vectors for every page makes large-scale deployment practically challenging. To address this, the research community has explored several optimization strategies, as illustrated in Figure 2. • One line of work involves merging patch embeddings, where methods like Light-ColPali (Ma et al., 2025) use clustering techniques to aggregate similar vectors. However, this approach often leads to a dilution of fine-grained information, resulting in unstable performance. 2 Another direction is pruning , where frameworks such as DocPruner (Yan et al., 2025) aim to discard redundant embeddings. These methods struggle to maintain performance under aggressive compression. 3 A third paradigm, exemplified by MetaEmbed (Xiao et al., 2025), introduces a set of abstract, learnable tokens to form a compact multi-vector representation. While innovative, these tokens lack an explicit grounding in the document's inherent layout structure, limiting their ability to capture crucial layout-specific semantics.
+
+[p. 1 | section: 1. Introduction | type: Text]
+To address the limitations of existing approaches, we introduce ColParse, a novel paradigm for constructing multivector representations that is fundamentally aligned with the structural nature of visual documents. Instead
+
+[p. 2 | section: 1. Introduction | type: Caption]
+Figure 2. The illustration of a multi-vector VDR model and three primary optimization strategies for its efficiency bottleneck.
+
+[p. 2 | section: 1. Introduction | type: Text]
+of operating on a uniform grid of patches or abstract tokens, {\tt ColParse} first employs a specialized document parsing model to intelligently segment each document page into a small set of k semantically meaningful, layout-informed sub-images (e.g., tables, figures, paragraphs), where k is typically less than 10. These k sub-images are then individually encoded by a standard single-vector retrieval model to yield k local vectors. In parallel, the entire document page is encoded to generate one global vector that captures the overall context. Finally, we fuse these representations by weighted element-wise adding the global vector to each of the k local vectors. This process results in k fused vectors for each document, which integrate both fine-grained, layout-specific details and holistic page-level context.
+
+[p. 2 | section: 1. Introduction | type: Text]
+We conducted comprehensive experiments on 24 diverse VDR datasets (Meng et al., 2025) to validate the effectiveness and robustness of our proposed framework. ColParse consistently delivers substantial performance improvements, achieving an average gain of over 10 points in nDCG@5 when applied to 10 different mainstream single-vector models. This demonstrates its remarkable flexibility as a training-free, plug-and-play module. By deeply integrating the unique structural properties of visual documents with the powerful mechanism of multi-vector retrieval, ColParse establishes a new trade-off between retrieval performance and storage efficiency. Our main contributions are as follows:
+
+[p. 2 | section: 1. Introduction | type: ListGroup]
+A Novel Paradigm for Multi-Vector Construction: We introduce the first layout-informed paradigm for constructing multi-vector representations in VDR, which overcomes the storage efficiency bottleneck of conventional multi-vector models by leveraging document parsing. ② A Flexible and Robust Framework: Our method is designed as a training-free, plug-and-play framework that demonstrates robust and significant performance gains across a wide array of existing single-vector models, highlighting its versatility and ease of adoption. Superior Performance with Enhanced Interpretability: ColParse provides inherent interpretability by enabling retrieval results to be traced back to specific, parsed layout components, which significantly enhances its practicality and potential for real-world industrial applications.
+
+[p. 2 | section: 2.1. Visual Document Retrieval | type: Text]
+VDR has become a crucial task for understanding visuallyrich documents, moving beyond traditional OCR-based pipelines that often lose critical layout information (Zhang et al., 2025b; Most et al., 2025). The advent of Vision-Language Models (VLMs) introduced end-to-end singlevector approaches (e.g., DSE (Ma et al., 2024a), GME (Zhang et al., 2024b), and UniSE (Liu et al., 2025b)), but these frequently struggle to capture the fine-grained semantics required for dense documents. A significant leap forward was made with the multi-vector paradigm, pioneered by ColPali (Faysse et al., 2024), which represents pages as numerous patch-level embeddings and employs late interaction for superior matching. Recent efforts have sought to optimize this paradigm at various levels: modellevel, by exploring bidirectional architectures like Modern-VBERT (Teiletche et al., 2025); data-level, through advanced data synthesis and hard-negative mining as seen in works like Llama Nemoretriever Colembed (Xu et al., 2025); and training-level, via new objectives and multi-task frameworks such as jina-embeddings-v4 (Günther et al., 2025). Despite their performance, these multi-vector models introduce a severe storage bottleneck.
+
+[p. 2 | section: 2.2. Mutli-Vector Retrieval | type: Text]
+The multi-vector paradigm, first popularized in text retrieval by ColBERT (Khattab & Zaharia, 2020), represents documents as sets of token-level embeddings to enable fine-grained matching through a late-interaction mechanism (Qian et al., 2022; Lee et al., 2023). This approach was further refined in the text domain by models like BGE-M3-Embedding (Chen et al., 2024) and Jina-ColBERT-v2 (Jha et al., 2024). The paradigm was successfully adapted for multimodal retrieval by ColPali (Faysse et al., 2024), shifting the focus to visual documents, which are inherently more complex than natural images. Despite their superior performance, these models face a critical efficiency bottleneck from the prohibitive storage cost of patch-level embeddings (Liu & Mao, 2023; Shrestha et al., 2024; Park et al., 2025). Current optimization efforts fall into three main categories, each with inherent drawbacks. (i) Pruning redundant embeddings, as seen in DocPruner (Yan et al., 2025) and Prune-then-Merge (Yan et al., 2026b), often struggles to maintain performance under aggressive compres-
+
+[p. 3 | section: 2.2. Mutli-Vector Retrieval | type: Text]
+sion. (ii) Merging similar embeddings via clustering, exemplified by Light-ColPali (Ma et al., 2025), can dilute fine-grained information, leading to unstable performance. (iii) Introducing abstract, learnable tokens, pioneered by MetaEmbed (Xiao et al., 2025) and CausalEmbed (Huo et al., 2026), creates compact representations that, however, lack an explicit grounding in the document's inherent layout structure. In contrast, ColParse addresses these limitations by leveraging document parsing to generate a compact set of layout-informed embeddings.
+
+[p. 3 | section: 2.3. Document Parsing VLM | type: Text]
+Document parsing VLMs have emerged as critical tools for converting visually-rich document images into structured formats like LaTeX or Markdown (Zhang et al., 2024a; Ouyang et al., 2025; Zhang et al., 2025c). Early models, such as Nougat (Blecher et al., 2023) and Donut (Kim et al., 2022), adopted an end-to-end, sequence-to-sequence approach but often struggled with the computational cost of high-resolution inputs. To balance accuracy and efficiency, a more recent multi-stage paradigm has gained traction. This is exemplified by models like MinerU2.5 (Niu et al., 2025), which first performs efficient layout analysis on a downsampled image before conducting targeted, high-resolution recognition on cropped regions. This coarse-to-fine strategy, also seen in models like Dolphin (Feng et al., 2025) and MonkeyOCR (Zhang et al., 2025a), effectively mitigates the O(N<sup>2</sup>) complexity of processing high-resolution images end-to-end. For ColParse, we select MinerU2.5 as our document parser, given its state-of-the-art accuracy and efficiency. A quantitative comparison with other document parsing models will be presented in Section 4.2.3.
+
+[p. 3 | section: 3. Methodology | type: Text]
+In this section, we first formalize the task of VDR within the multi-vector paradigm. We then introduce the ColParse framework, detailing its multi-stage process for generating compact, layout-informed document representations. See our pseudo-code in Appendix A.
+
+[p. 3 | section: 3.1. Task Formulation | type: Text]
+The primary goal of VDR is, given a textual query q, to retrieve a ranked list of relevant document pages from a large-scale corpus \mathcal{C}=\{d_1,d_2,\ldots,d_{|\mathcal{C}|}\} . In the conventional multi-vector retrieval paradigm, a document page d is first rendered as an image and then uniformly partitioned into a grid of N_p patches, \{p_j\}_{j=1}^{N_p} . A VLM, serving as an encoder \Phi(\cdot) , maps each patch p_j into a D-dimensional embedding, resulting in a large set of patch-level document embeddings \mathbf{D}_{\mathrm{grid}}=\{\mathbf{d}_j\}_{j=1}^{N_p} , where each \mathbf{d}_j\in\mathbb{R}^D . Concurrently, the same encoder maps the textual query q into a set
+
+[p. 3 | section: 3.1. Task Formulation | type: PictureGroup]
+Figure 3. The simplified illustration of ColParse framework.
+
+[p. 3 | section: 3.1. Task Formulation | type: Text]
+of N_q token-level embeddings \mathbf{Q} = \{\mathbf{q}_i\}_{i=1}^{N_q} , where each \mathbf{q}_i \in \mathbb{R}^D . The relevance score s(q,d) between the query and the document is then computed using a late-interaction mechanism, typically MaxSim, as defined below:
+
+[p. 3 | section: 3.1. Task Formulation | type: Equation]
+s(q, d) = \sum_{i=1}^{N_q} \max_{j=1}^{N_p} (\mathbf{q}_i^{\mathsf{T}} \mathbf{d}_j). \tag{1}
+
+[p. 3 | section: 3.1. Task Formulation | type: Text]
+where vectors are assumed to be L2-normalized. While this grid-based approach excels at fine-grained matching, it incurs a prohibitive storage cost of O(N_p \times D) per document page, as N_p can be in the hundreds or thousands.
+
+[p. 3 | section: 3.1. Task Formulation | type: Text]
+The objective of our work is to address this critical bottleneck. We aim to replace the large, layout-agnostic set \mathbf{D}_{\text{grid}} with a highly compact, structurally-aware multi-vector representation \mathbf{D}_{\text{ColParse}} , which contains only k vectors, where k \ll N_p . This new representation should significantly reduce storage requirements to O(k \times D) while simultaneously enhancing retrieval performance by being explicitly grounded in the document's semantic layout.
+
+[p. 3 | section: 3.2. The ColParse Framework | type: Text]
+ColParse is a plug-and-play framework that revolutionizes the construction of multi-vector representations by moving "beyond the grid." Instead of relying on uniform patches, it leverages structural understanding to generate a compact and semantically rich set of embeddings. As shown in Figure 3, our framework operates offline in a three-stage pipeline for each document image d \in \mathbb{R}^{H \times W \times 3} : (1) Layout-Informed Document Parsing, (2) Dual-Stream Encoding, and (3) Global-Local Fusion.
+
+[p. 3 | section: 3.2.1. LAYOUT-INFORMED DOCUMENT PARSING | type: Text]
+The foundational step of ColParse is to deconstruct the document image into its constituent semantic components. We employ a specialized, off-the-shelf document parsing model, \Psi_{\text{parse}}(\cdot) , which functions as a layout detector. For a given document image d, the parser identifies a set of k distinct layout regions, outputting their corresponding bounding boxes and content type labels: [\{b_j,c_j\}_{j=1}^k]=\Psi_{\text{parse}}(d) . Here, b_j=(x_{j1},y_{j1},x_{j2},y_{j2}) is the bounding box for the j-th region, defining its coordinates within the original image. c_j
+
+[p. 4 | section: 3.2.1. LAYOUT-INFORMED DOCUMENT PARSING | type: Text]
+is a categorical label from a predefined set of content types \mathcal{C}_{\text{types}} = \{\text{'title'}, \text{'table'}, \text{'figure'}, \dots \} , indicating the semantic nature of the region. Using these bounding boxes, we crop the original image d to extract a set of k sub-images \mathcal{S}_d . The number of sub-images, k, is dynamically determined by the parser based on the document's complexity and is typically very small (e.g., k < 10). The process can be formulated as \mathcal{S}_d = \{s_1, s_2, \dots, s_k\} where s_j = \text{Crop}(d, b_j) . Each sub-image s_j \in \mathbb{R}^{H_j \times W_j \times 3} is of variable size, flexibly conforming to the dimensions of the detected layout component. This intelligent, content-aware segmentation forms the basis for our compact representation.
+
+[p. 4 | section: 3.2.2. DUAL-STREAM ENCODING | type: Text]
+Once the document is parsed, we generate embeddings using a standard single-vector retrieval model, \Phi_{\rm enc}(\cdot) : \mathbb{R}^{H' \times W' \times 3} \to \mathbb{R}^D , which serves as the base encoder. This stage operates in two parallel streams to capture both local, layout-specific details and global, page-level context.
+
+[p. 4 | section: 3.2.2. DUAL-STREAM ENCODING | type: Text]
+Local Encoding. Each of the k variable-sized sub-images s_j from \mathcal{S}_d is resized and independently passed through the encoder \Phi_{\mathrm{enc}}(\cdot) to produce a corresponding D-dimensional local vector, \mathbf{v}_{\mathrm{local}}^{(j)} . This process yields a set of k local embeddings, each representing a distinct semantic unit: \mathbf{D}_{\mathrm{local}} = \{\mathbf{v}_{\mathrm{local}}^{(j)} \in \mathbb{R}^D \mid \mathbf{v}_{\mathrm{local}}^{(j)} = \Phi_{\mathrm{enc}}(s_j), \forall s_j \in \mathcal{S}_d\}_{j=1}^k .
+
+[p. 4 | section: 3.2.2. DUAL-STREAM ENCODING | type: Text]
+Global Encoding. In parallel, the entire , un-cropped document page image d is passed through the same encoder \Phi_{\rm enc}(\cdot) to generate a single D-dimensional global vector, \mathbf{v}_{\rm global} . This vector serves as a holistic summary of the page, capturing the overall context and relationships between layout components: \mathbf{v}_{\rm global} = \Phi_{\rm enc}(d) \in \mathbb{R}^D . This dual-stream design ensures our final representation benefits from both the specificity of individual layout elements and the broader context of the entire page.
+
+[p. 4 | section: 3.2.3. GLOBAL-LOCAL FUSION FOR FINAL REPRESENTATION | type: Text]
+Instead of a simple summation, we employ a weighted fusion strategy that allows for a tunable balance between these two critical streams of information. For each of the k local vectors, \mathbf{v}_{\text{local}}^{(j)} , we inject the holistic context captured by the single global vector, \mathbf{v}_{\text{global}} , using a balancing factor \alpha \in [0,1] . The resulting fused vector, \mathbf{d}_{\text{fused}}^{(j)} \in \mathbb{R}^D , synergistically integrates both fine-grained detail and overarching context. The fusion is performed as a weighted elementwise addition: \mathbf{d}_{\text{fused}}^{(j)} = \alpha \cdot \mathbf{v}_{\text{global}} + (1-\alpha) \cdot \mathbf{v}_{\text{local}}^{(j)}, \quad \forall j \in \{1,\dots,k\} . This weighted mechanism provides the flexibility to tailor the representation's focus.
+
+[p. 4 | section: 3.2.3. GLOBAL-LOCAL FUSION FOR FINAL REPRESENTATION | type: Text]
+This fusion operation is performed for all k local vectors, producing the final multi-vector representation for document
+
+[p. 4 | section: 3.2.3. GLOBAL-LOCAL FUSION FOR FINAL REPRESENTATION | type: Text]
+d, denoted as \mathbf{D}_{\mathtt{ColParse}} = \{\mathbf{d}_{\mathtt{fused}}^{(j)}\}_{j=1}^k \subset \mathbb{R}^{k \times D} . This set, containing only k structurally-aware and context-enriched vectors, is then stored for online retrieval, achieving our goal of massive storage reduction.
+
+[p. 4 | section: 3.2.4. LATE-INTERACTION SCORING WITH ColParse | type: Text]
+During the online retrieval phase, the relevance score between a query q and a document d is computed efficiently using the compact representation \mathbf{D}_{\mathtt{ColParse}} . The query is first encoded into its token-level embeddings \mathbf{Q} = \{\mathbf{q}_i \in \mathbb{R}^D\}_{i=1}^{N_q} as standard. The MaxSim score is then calculated over the k fused document vectors:
+
+[p. 4 | section: 3.2.4. LATE-INTERACTION SCORING WITH ColParse | type: Equation]
+s_{\texttt{ColParse}}(q, d) = \sum_{i=1}^{N_q} \max_{j=1}^k (\mathbf{q}_i^\top \mathbf{d}_{\text{fused}}^{(j)}). \tag{2}
+
+[p. 4 | section: 3.2.4. LATE-INTERACTION SCORING WITH ColParse | type: Text]
+By replacing the search over N_p grid-based vectors with a search over just k layout-informed vectors, ColParse not only dramatically reduces the storage footprint but also focuses the late-interaction mechanism on the most semantically salient parts of the document.
+
+[p. 4 | section: 3.3. Theoretical Foundation | type: Text]
+We provide a theoretical justification for ColParse from an Information Bottleneck (IB) perspective. We demonstrate that the framework's architecture serves as a principled surrogate for solving the intractable IB objective in VDR by (1) disentangling source information via parsing and (2) refining it with contextual side-information.
+
+[p. 4 | section: 3.3.1. THE VDR COMPRESSION PROBLEM AS AN INFORMATION BOTTLENECK | type: Text]
+Let D be a random variable representing a document image and Q be a random variable for a query, drawn from an unknown distribution P(Q). Let R be a relevance variable, a function of D and Q. The goal is to learn a compression function g that maps a document D to a compact representation Z = g(D) by solving the IB Lagrangian:
+
+[p. 4 | section: 3.3.1. THE VDR COMPRESSION PROBLEM AS AN INFORMATION BOTTLENECK | type: Equation]
+\min_{q} \mathcal{L}(Z) = I(Z; D) - \beta \mathbb{E}_{Q}[I(Z; R(D, Q))]. \quad (3)
+
+[p. 4 | section: 3.3.1. THE VDR COMPRESSION PROBLEM AS AN INFORMATION BOTTLENECK | type: Text]
+This objective seeks to minimize the information Z retains about the source D (compression) while maximizing the information it preserves about the relevance R (prediction). The expectation \mathbb{E}_Q over the unknown query distribution makes this problem intractable at indexing time.
+
+[p. 4 | section: 3.3.2. Information Disentanglement via Parsing | type: Text]
+ColParse's first stage, parsing, transforms the input space. Let \Psi_{\text{parse}}(D) = \{S_1, \dots, S_k\} be the set of sub-images (semantic regions) derived from document D. These regions form a partition of the document's core semantic content. By the chain rule of mutual information, the information in the original document can be expressed through its
+
+[p. 5 | section: 3.3.2. Information Disentanglement via Parsing | type: Text]
+components: I(D;R) = I(S_1,S_2,\ldots,S_k;R) . We posit the Semantic Concentration Axiom : for any given query Q, the relevance signal R is predominantly determined by a single primary semantic region S_{j^*} \in \{S_j\} , where j^* is the index of the most relevant region. This implies near conditional independence for the remaining regions: I(S_{\neg j^*};R|S_{j^*})\approx 0 , where S_{\neg j^*}=\{S_j\}_{j\neq j^*} . This axiom leads to the approximation: I(D;R)\approx I(S_{j^*};R) . This decomposition transforms the problem from compressing monolithic variable D to creating a set of representations, one for each potential primary channel S_j . This provides a structural prior to IB problem, justifying creation of a multivector set \{g_j(S_j)\}_{j=1}^k instead of a single vector g(D).
+
+[p. 5 | section: 3.3.2. Information Disentanglement via Parsing | type: Text]
+271272
+
+[p. 5 | section: 3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION | type: Text]
+ColParse generates two sets of intermediate representations: local vectors \{V_j = \Phi_{\mathrm{enc}}(S_j)\}_{j=1}^k and a global vector V_{\mathrm{global}} = \Phi_{\mathrm{enc}}(D) . Each of these encoding steps is itself an information bottleneck, subject to the Data Processing Inequality (DPI): I(V_j;R) \leq I(S_j;R), \quad \forall j \in \{1,\dots,k\}; I(V_{\mathrm{global}};R) \leq I(D;R) . The core of ColParse is the fusion step, which creates the final representation set \{Z_j = V_j + V_{\mathrm{global}}\}_{j=1}^k . To analyze its benefit, consider the joint information held by the local-global pair (V_j, V_{\mathrm{global}}) about the relevance R. Using the chain rule:
+
+[p. 5 | section: 3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION | type: Equation]
+I(V_j,V_{\rm global};R) = I(V_j;R) + \underbrace{I(V_{\rm global};R|V_j)}_{\mbox{Contextual Information Gain}} \ . \ \ (4)
+
+[p. 5 | section: 3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION | type: Text]
+This gain term quantifies the new information about relevance that the global context provides, given the local region. It is non-zero if global context helps disambiguate local content. The fusion function f(V_j,V_{\mathrm{global}})=V_j+V_{\mathrm{global}} creates the vector Z_j . By DPI, this fusion is also a bottleneck:
+
+[p. 5 | section: 3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION | type: Equation]
+I(Z_i; R) = I(V_i + V_{\text{global}}; R) \le I(V_i, V_{\text{global}}; R). (5)
+
+[p. 5 | section: 3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION | type: Text]
+The objective of this fusion is to craft a compact vector Z_j that is more informative than the local vector V_j alone, by capturing a significant portion of the contextual gain. The net improvement in information for region j is:
+
+[p. 5 | section: 3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION | type: Equation]
+\Delta I_j = I(Z_j; R) - I(V_j; R). \tag{6}
+
+[p. 5 | section: 3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION | type: Text]
+The success of ColParse relies on this fusion being effective, i.e., ensuring \Delta I_j > 0 . This holds when the fusion operation successfully encodes the contextual information gain. The simple vector addition serves as a parameter-free, computationally efficient mechanism to achieve this. The final representation \mathbf{D}_{\text{ColParse}} = \{Z_j\}_{j=1}^k is thus a set of contextually-refined, disentangled vectors that more effectively preserve query-relevant information, providing a superior solution to the VDR compression problem.
+
+[p. 5 | section: 3.3.3. CONTEXTUAL REFINEMENT VIA SYNERGISTIC FUSION | type: Text]
+See more theoretical analysis in Appendix B.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+Benchmarks and Evaluation. To ensure a comprehensive evaluation, we assess the performance of ColParse across five mainstream VDR benchmark suites, encompassing a total of 24 diverse datasets: ViDoRe-V1 (Faysse et al., 2024), ViDoRe-V2 (Macé et al., 2025), VisRAG (Yu et al., 2024), ViDoSeek (Wang et al., 2025), and MMLongBench (Ma et al., 2024b). We validate the versatility and plug-andplay nature of our framework by applying it to ten prominent single-vector retrieval models: VLM2Vec-V1-2B/7B (Jiang et al., 2024), VLM2Vec-V2-2B (Meng et al., 2025), LamRA-Ret (Liu et al., 2025a), GME-2B/7B (Zhang et al., 2024b), UniME-V2-2B/7B (Gu et al., 2025), B3-2B/7B (Thirukovalluru et al., 2025). See details of benchmarks and models in Appendix C.1 and C.2, respectively. In alignment with established VDR practices (Wasserman et al., 2025; ILLUIN, 2025), we use the Normalized Discounted Cumulative Gain at 5 (nDCG@5) as the evaluation metric. Baselines. To demonstrate the superiority of ColParse, we compare it against a diverse set of baselines organized into five distinct categories. • Base: This category represents the original performance of the ten single-vector models without any multi-vector adaptation. ② Multi-img: In this approach, all sub-images parsed from a document are fed simultaneously into the base models, leveraging their native support for multi-image inputs to generate a single representative vector. 3 Chunking-layout: This category explores strategies that operate directly on the final layer of token embeddings from the base model. Guided by the layout parsing results, these tokens are chunked and aggregated, while any tokens outside the parsed bounding boxes are discarded. We evaluate four variants: type-cluster (tokens of the same content type are merged via semantic clustering), type-mean (tokens of the same content type are merged via mean pooling), subimg-cluster (tokens from the same sub-image region are clustered), and subimg-mean (tokens from the same sub-image region are pooled). 4 Chunking-semantic: In contrast to layoutguided methods, this baseline performs hierarchical clustering on the entire set of final-layer tokens to generate a fixed number of vectors (defaulting to 10), operating agnostically to the document's layout structure. 6 Single2multi: This category mimics ColParse by encoding each parsed sub-image separately but employs alternative scoring mechanisms instead of our vector fusion. The two variants are scoring-add and scoring-multiply, where the final document score is computed by either adding or multiplying the individual query-sub_img similarity scores with the global_img-sub_img similarity scores, respectively.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+Implementation Details. To ensure a fair comparison and reproducibility, our entire evaluation pipeline is built
+
+[p. 6 | section: 4.1. Experimental Setup | type: FigureGroup]
+Figure 4. The performance comparison (evaluated by nDCG@5) between ColParse and baselines on five VDR benchmarks across ten mainstream single-vector multimodal retrieval models. Refer to Table 2 and Table 3 for detailed result records due to the space limit.
+
+[p. 6 | section: 4.1. Experimental Setup | type: Text]
+upon the MMEB codebase 1 . We employ MinerU2.5 2 (Niu et al., 2025) as the unified document parsing model across all experiments due to its state-of-the-art performance and efficiency (See details of MinerU2.5 in Appendix C.3) . For ColParse , we choose α ranging from 0.1 to 0.9 with an interval of 0.1, and we select the optimal hyperparameter for each base model to report the final results (validated in Section 4.2.3) . Experiments were conducted on a cluster of NVIDIA A100 (80G) GPUs. The complete codebase will be made publicly available upon acceptance.
+
+[p. 6 | section: 4.2.1. MAIN RESULT | type: Text]
+ColParse consistently outperforms both single-vector base models and existing multi-vector optimization baselines across diverse benchmarks. As shown in Table 2 and Table 3 (Appendix C.4) , ColParse achieves a remarkable average nDCG@5 gain of 31.64 points for VLM2Vec-V1- 2B and 42.69 points for its 7B counterpart on the ViDoRe-V1 benchmark. Figure 4 further visualizes this superiority, where the red envelope representing ColParse consistently forms the outermost boundary across all ten mainstream models. This suggests that layout-informed sub-image representations capture critical fine-grained details that are typically "diluted" in monolithic global embeddings.
+
+[p. 6 | section: 4.2.1. MAIN RESULT | type: Text]
+The framework demonstrates exceptional versatility and robustness as a training-free, plug-and-play module for various VLM-based embeddings. Across ten distinct models including VLM2Vec, GME, UniME, and B3, ColParse consistently yields performance improvements regardless of the model's architecture or parameter scale. Notably, even for the high-performing GME-7B model, ColParse maintains state-of-the-art results while other optimization strategies like semantic chunking (c-sem) cause a drastic performance drop from 89.36 to 23.21 on ViDoRe-V1. This universality implies that layout awareness is a fundamental, model-agnostic enhancement for visual document understanding that can be unlocked at representation level.
+
+[p. 6 | section: 4.2.1. MAIN RESULT | type: Text]
+Layout-informed decomposition is substantially more effective for multi-vector construction than traditional token-level chunking or clustering. Quantitative results reveal that baselines like cl-t-c (type-clustering) and cl-s-m (sub-image mean pooling) often lead to significant performance degradation; for instance, VLM2Vec-V2-2B's performance plummets from 74.16 to 24.85 when using token-level clustering. In contrast, ColParse boosts the same model to 78.41 by preserving the visual integrity of semantic regions rather than aggregating abstract token embeddings. This phenomenon indicates that maintaining the raw visual-semantic alignment within parsed regions is superior to post-hoc heuristic aggregation of late-stage features.
+
+[p. 6 | section: 4.2.1. MAIN RESULT | type: Text]
+ColParse exhibits superior efficacy in handling complex, long-form documents that require multi-hop reasoning.
+
+[p. 6 | section: 4.2.1. MAIN RESULT | type: Footnote]
+1 2
+
+[p. 7 | section: 4.2.1. MAIN RESULT | type: Caption]
+Figure 5. Variant study of ColParse and its variants.
+
+[p. 7 | section: 4.2.1. MAIN RESULT | type: Text]
+On the MMLongBench dataset (Table 2) , ColParse elevates the average nDCG@5 of VLM2Vec-V1-2B from 25.93 to 32.07 and UniME-V2-2B from 29.31 to 44.21, outperforming all other compression baselines. This performance leap is particularly evident in the cases which require cross-page information locating. We speculate that by intelligently segmenting pages into key layout components ( e.g., tables, figures), ColParse reduces the "signal-to-noise" ratio during the late-interaction phase, allowing the model to focus on the most semantically salient regions.
+
+[p. 7 | section: 4.2.1. MAIN RESULT | type: Text]
+The global-local fusion strategy is critical for providing necessary contextual grounding to isolated semantic regions. Comparative analysis with the single2multi (s2m-add/mul) baselines in Table 3 shows ColParse consistently leads on challenging tasks like VisRAG and ViDoSeek, achieving 51.96 on VisRAG with VLM2Vec-V1- 2B compared to only 38.94 for s2m-add. This significant gap highlights local sub-images alone often lack the holistic context ( e.g., a table without its preceding paragraph's context) needed for accurate retrieval. The synergy achieved through weighted fusion allows ColParse to retain both regional specificity and page-level semantics.
+
+[p. 7 | section: 4.2.2. VARIANT STUDY | type: Text]
+We evaluate three specific variants: (i) single2multi (s2m) variant serves as the baseline layout-decomposed representation, which extracts and encodes sub-images as independent vectors without any global information. (ii) single2multitype-cluster (s2m-t-c) variant extends s2m by aggregating all sub-image embeddings of the same semantic category into a single representative type-level vector to further compress the representation. (iii) single2multi-global-inclusion (s2m-g-i) variant appends the original holistic page-level embedding to the s2m sub-image set as a separate global vector to provide page-wide context.
+
+[p. 7 | section: 4.2.2. VARIANT STUDY | type: FigureGroup]
+Figure 6. The comparison of the average performance of ColParse across different balancing factors. The dash lines refer to the base results; and the star points refer to the best-performing balancing factors. See Figure 10 for the model-level comparisons.
+
+[p. 7 | section: 4.2.2. VARIANT STUDY | type: Text]
+Synergistic global-local fusion is significantly more effective than simple global vector inclusion for contextualizing layout-aware representations. As shown in Figure 5, ColParse consistently outperforms the s2m-g-i variant across all ten base models and benchmarks, for instance, achieving a gain of 2.44 points in nDCG@5 over s2m-g-i for VLM2Vec-V1-2B on ViDoRe-V1. This superiority is further visualized in the radar plots of Figure 9 (Appendix C.5) , where ColParse (red envelope) consistently covers the largest area, particularly in dense tasks like VisRAG where it leads s2m-g-i by over 2 points on the 7B model. This suggests that element-wise fusion serves as a deeper conditioning mechanism than simple inclusion, allowing local features to be fundamentally "re-weighted" by the global semantic environment of the document.
+
+[p. 7 | section: 4.2.2. VARIANT STUDY | type: Text]
+See more analysis in Appendix C.5 due to the space limit.
+
+[p. 7 | section: 4.2.3. HYPERPARAMETER ANALYSIS | type: Text]
+Effect of balancing factor. We investigate the sensitivity of ColParse to the balancing factor α, as illustrated in Figure 6 (details in Appendix C.6.1) . First, the retrieval performance is robust across the entire range of α ∈ [0.1, 0.9], consistently surpassing the single-vector baseline for all tested models. For instance, in VLM2Vec-V1-2B, even the lowest α of 0.1 achieves an overall score of 35.97, which is still a substantial improvement over the 31.18 base result. Second, the performance typically follows a convex trajectory, with the optimal balance point generally leaning towards the global context to effectively ground the isolated layout components. Quantitatively, most models reach their peak performance within the range of α ∈ [0.6, 0.8]. Finally, a moderate synergistic fusion is essential, as either excessive local specificity or excessive global dominance leads to suboptimal results. In B3-7B, performance steadily climbs from 62.09 at α = 0.1 to its peak of 68.51 at α = 0.7, before exhibiting a slight decline as the representation becomes overly dominated by the global vector at α = 0.9.
+
+[p. 8 | section: 4.2.3. HYPERPARAMETER ANALYSIS | type: Caption]
+Figure 7. Comparison of MinerU2.5 against its counterparts. The y-axis represents the overall score of OmniDocBench (Ouyang et al., 2025) , the x-axis shows end-to-end throughput (Pages/Sec), and bubble size indicates the parameter size.
+
+[p. 8 | section: 4.2.3. HYPERPARAMETER ANALYSIS | type: TableGroup]
+Table 1. Efficiency analysis of ColParse on the best performing model GME-7B and its multi-vector counterpart (w/ original setting). Performance denotes the overall score of MMEB-visdoc, Storage refers to the number of vectors stored per document, and Latency represents the average encoding time per document. * denotes the multi-vector model is trained w/ aligned configuration. Orange arrows denote better and blue ones denote worse. Model Performance Storage Latency GME-7b 79.50 1.00 0.30 +ColParse 80.61↑1.11 5.90↑4.9 0.81↑0.51 ColQwen* 80.02↑0.52 768.00↑767 0.41↑0.11
+
+[p. 8 | section: 4.2.3. HYPERPARAMETER ANALYSIS | type: Text]
+Effect of document parsing model. We evaluate the capability of various document parsing models and conclude that MinerU2.5 provides the optimal balance between parsing fidelity and inference efficiency, as shown in Figure 7 (Details in Appendix C.6.2) . First, MinerU2.5 demonstrates superior parsing accuracy across all semantic and structural dimensions compared to existing specialized VLMs. Quantitatively, it achieves a state-of-the-art Overall score of 90.67 on OmniDocBench (Ouyang et al., 2025) , significantly outperforming the best baseline MonkeyOCR-pro-3B (88.85) and maintaining the lowest error rates in both text (0.047) and reading order (0.044) recognition. Second, MinerU2.5 achieves industrial-grade throughput, ensuring the practicality of the ColParse pipeline for large-scale document corpora. For example, it delivers an end-to-end processing speed of 2.25 Pages/sec, which is 4× faster than highparameter models like Nanonets-OCR-s (0.55 Pages/sec). Consequently, we select MinerU2.5 as our unified parser as it sits at the optimal Pareto front, offering the highest parsing quality while sustaining a remarkable throughput.
+
+[p. 8 | section: 4.2.4. EFFICIENCY ANALYSIS | type: Text]
+To ensure a fair comparison, we evaluate ColParse using GME-7B, the best-performing single-vector model, and ColQwen, a multi-vector baseline with an aligned architecture and training configuration. We conclude that our
+
+[p. 8 | section: 4.2.4. EFFICIENCY ANALYSIS | type: Text]
+paradigm achieves superior retrieval performance while drastically slashing the storage overhead inherent in traditional multi-vector models. Specifically, as shown in Table 1, ColParse yields an overall score of 80.61 on MMEB-visdoc, outperforming the multi-vector counterpart ColQwen (80.02) while requiring only 5.9 vectors per page (See per-dataset records in Appendix C.6.3) , a massive storage reduction of over 99% compared to ColQwen's 768 vectors. Furthermore, although ColParse introduces a marginal increase in encoding latency, it remains a highly practical solution for real-world deployment. While the per-document latency rises to 0.81s, it is still significantly lower than the average 7s delay of conventional OCR-based pipelines, and the offline parsing stage can be optimized through parallel processing to mitigate indexing overhead.
+
+[p. 8 | section: 4.2.5. CASE STUDY | type: FigureGroup]
+Figure 8. The illustration of a representative case.
+
+[p. 8 | section: 4.2.5. CASE STUDY | type: Text]
+We conduct a case study in Figure 8 to illustrate the interpretability of ColParse beyond its superior retrieval performance. When handling a query requiring specific details, such as carbon emission reduction percentages, ColParse not only retrieves the correct document page but also pinpoints the exact parsed sub-vector ( e.g., a specific text block) that yields the highest MaxSim score. This layout-informed granularity enables fine-grained information back-tracing, allowing the system to present the specific evidence region directly to the user. Such explainability is highly valuable in practical industrial scenarios, such as financial auditing or legal review, where the ability to verify the source of information is as critical as retrieval accuracy.
+
+[p. 8 | section: 5. Conclusion | type: Text]
+In this paper, we introduced ColParse , a novel layoutinformed paradigm designed to overcome the critical storage efficiency bottleneck in multi-vector VDR. Our framework uniquely generates a compact multi-vector representation by fusing layout-aware sub-image embeddings from a document parser with a holistic global vector. Extensive experiments demonstrated that ColParse achieves over 95% storage compression while consistently improving retrieval performance across base models and datasets. Ultimately, ColParse charts a new course for the field, establishing that a deep understanding of document structure is critical for practical multimodal information systems.
+
+[p. 9 | section: Impact Statement | type: Text]
+Ethical Considerations. We believe that our proposed ColParse framework raises no new ethical concerns. Its motivation is to advance the efficiency and performance of VDR systems in a principled and resource-conscious manner. By leveraging existing document parsing technologies to create more compact and interpretable representations, our method promotes responsible AI development, adhering to established ethical guidelines in information retrieval research without relying on sensitive or proprietary data.
+
+[p. 9 | section: Impact Statement | type: Text]
+Societal Implications. ColParse introduces a new paradigm for multimodal information retrieval by shifting from storage-intensive, grid-based representations to a lightweight, layout-informed approach. It fundamentally resolves the critical conflict between fine-grained retrieval accuracy and the practical storage costs of large-scale deployment. By reducing storage requirements by over 95% while simultaneously enhancing performance, ColParse significantly lowers the barrier for deploying state-of-the-art visual document understanding systems. This has the potential to democratize access to advanced information retrieval in diverse domains, including academic research, enterprise knowledge management, and digital archives. Furthermore, its inherent interpretability, which links retrieval results to specific document components, fosters greater transparency and user trust in AI-powered information systems.

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/parse_report.json

@@ -0,0 +1,80 @@
+{
+  "paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0",
+  "pipeline": "marker_non_llm_v3",
+  "parser": "marker_single",
+  "formats": [
+    "markdown",
+    "chunks"
+  ],
+  "llm_enabled": false,
+  "pdf_path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/icml26_20260429_1952_duequeue/raw/3250cb92-2f69-4e16-9df9-f569224173f0.pdf",
+  "pdf_sha256": "6792e28617d10eeebaf4b9677383f9b6d4705973786ce23b000bcf5e2d52e2e5",
+  "bytes": 7873527,
+  "source": "https://koala.science/storage/pdfs/3250cb92-2f69-4e16-9df9-f569224173f0.pdf",
+  "page_count": 27,
+  "ok": true,
+  "elapsed_seconds": 194.01999999999998,
+  "paper2markdown_v3": {
+    "paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0",
+    "pipeline": "Paper2Markdown-V3",
+    "ok": true,
+    "page_count": 27,
+    "chunk_count": 200,
+    "main_body_chunk_count": 75,
+    "appendix_chunk_count": 117,
+    "reference_chunk_count": 8,
+    "model_text_chars": 42730,
+    "raw_markdown_chars": 191958,
+    "sanitized_chars": 130384,
+    "page_provenance": {
+      "min_page": 1,
+      "max_page": 27,
+      "invalid_count": 0
+    },
+    "marker_block_type_counts": {
+      "Caption": 12,
+      "Code": 13,
+      "Equation": 17,
+      "Figure": 3,
+      "FigureGroup": 6,
+      "Footnote": 7,
+      "ListGroup": 10,
+      "PageFooter": 27,
+      "PageHeader": 24,
+      "Picture": 1,
+      "PictureGroup": 4,
+      "SectionHeader": 68,
+      "Table": 7,
+      "TableGroup": 7,
+      "Text": 573
+    },
+    "asset_count_raw": 14,
+    "asset_count_model_kept": 9,
+    "asset_count_rejected": 5,
+    "asset_reject_reasons": {
+      "kept": 9,
+      "too_small": 5
+    },
+    "artifact_leak_audit": {
+      "ok": true,
+      "hits": {
+        "Anonymous Authors": [],
+        "ACKNOWLEDGMENT": [],
+        "OpenReview": [],
+        "\"accept_label\"": [],
+        "\"decision\"": [],
+        "\"decision_tier\"": [],
+        "\"source_status\"": [],
+        "Meta-review": [],
+        "Official Review": [],
+        "official_reviews": [],
+        "meta_reviews": [],
+        "suggested_verdict_score": []
+      },
+      "artifact_count": 2
+    },
+    "default_model_input": "model_text_v3.txt",
+    "appendix_input": "appendix_text_v3.txt",
+    "reference_input": "reference_text_v3.txt"
+  }
+}

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/reference_chunks.jsonl

@@ -0,0 +1,8 @@
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0075", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Blecher, L., Cucurull, G., Scialom, T., and Stojnic, R. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418 , 2023. Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D., and Liu, Z. M3-embedding: Multi-linguality, multifunctionality, multi-granularity text embeddings through self-knowledge distillation. In Findings of the Association for Computational Linguistics ACL 2024 , pp. 2318–2335, 2024. Faysse, M., Sibille, H., Wu, T., Omrani, B., Viaud, G., Hudelot, C., and Colombo, P. Colpali: Efficient document retrieval with vision language models. arXiv preprint arXiv:2407.01449 , 2024. Feng, H., Wei, S., Fei, X., Shi, W., Han, Y., Liao, L., Lu, J., Wu, B., Liu, Q., Lin, C., et al. Dolphin: Document image parsing via heterogeneous anchor prompting. arXiv preprint arXiv:2505.14059 , 2025. Gu, T., Yang, K., Zhang, K., An, X., Feng, Z., Zhang, Y., Cai, W., Deng, J., and Bing, L. Unime-v2: Mllm-as-ajudge for universal multimodal embedding learning, 2025. URL .", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/530"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0076", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Gunther, M., Sturua, S., Akram, M. K., Mohr, I., Ungureanu, ¨ A., Wang, B., Eslami, S., Martens, S., Werk, M., Wang, N., et al. jina-embeddings-v4: Universal embeddings for multimodal multilingual retrieval. arXiv preprint arXiv:2506.18902 , 2025. Huo, J., Huang, Y., Yan, Y., Pan, Y., Cao, Y., Ou, M., Yu, P. S., and Hu, X. Causalembed: Auto-regressive multivector generation in latent space for visual document embedding. arXiv preprint arXiv:2601.21262 , 2026. ILLUIN. ViDoRe V3: a comprehensive evaluation of retrieval for enterprise use-cases, nov 2025. URL QuentinJG/introducing-vidore-v3 . Jayaram, R., Dhulipala, L., Hadian, M., Lee, J. D., and Mirrokni, V. Muvera: Multi-vector retrieval via fixed dimensional encoding. Advances in Neural Information Processing Systems , 37:101042–101073, 2024. Jha, R., Wang, B., Gunther, M., Mastrapas, G., Sturua, ¨ S., Mohr, I., Koukounas, A., Akram, M. K., Wang, N., and Xiao, H. Jina-colbert-v2: A general-purpose multilingual late interaction retriever. arXiv preprint arXiv:2408.16672 , 2024. Jiang, Z., Meng, R., Yang, X., Yavuz, S., Zhou, Y., and Chen, W. Vlm2vec: Training vision-language models for massive multimodal embedding tasks. arXiv preprint arXiv:2410.05160 , 2024. Khattab, O. and Zaharia, M. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pp. 39–48, 2020. Kim, G., Hong, T., Yim, M., Nam, J., Park, J., Yim, J., Hwang, W., Yun, S., Han, D., and Park, S. Ocr-free document understanding transformer. In European Conference on Computer Vision , pp. 498–517. Springer, 2022. Lee, J., Dai, Z., Duddu, S. M. K., Lei, T., Naim, I., Chang, M.-W., and Zhao, V. Rethinking the role of token retrieval in multi-vector retrieval. Advances in Neural Information Processing Systems , 36:15384–15405, 2023. Liu, Q. and Mao, J. Understanding the multi-vector dense retrieval models. In Proceedings of the 32nd ACM In ternational Conference on Information and Knowledge Management , pp. 4110–4114, 2023. Liu, Y., Zhang, Y., Cai, J., Jiang, X., Hu, Y., Yao, J., Wang, Y., and Xie, W. Lamra: Large multimodal model as your advanced retrieval assistant. In Proceedings of the Computer Vision and Pattern Recognition Conference , pp. 4015–4025, 2025a.", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/531"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0077", "section": "References", "page_start": 10, "page_end": 10, "type": "Text", "text": "495 496 497 498 499 500 Liu, Z., Liu, Z., Liang, Z., Zhou, J., Xiao, S., Gao, C., Zhang, C. J., and Lian, D. Any information is just worth one single screenshot: Unifying search with visualized information retrieval. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 19238–19261, 2025b.", "source": "marker_v2", "marker_block_id": "/page/9/Text/501"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0078", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "Ma, X., Lin, S.-C., Li, M., Chen, W., and Lin, J. Unifying multimodal retrieval via document screenshot embedding. In Al-Onaizan, Y., Bansal, M., and Chen, Y.-N. (eds.), Proceedings of the 2024 Conference on Empirical Meth ods in Natural Language Processing , pp. 6492–6505, Miami, Florida, USA, November 2024a. Association for Computational Linguistics. doi: 10.18653/v1/2024. emnlp-main.373. URL org/2024.emnlp-main.373/ . Ma, Y., Zang, Y., Chen, L., Chen, M., Jiao, Y., Li, X., Lu, X., Liu, Z., Ma, Y., Dong, X., et al. Mmlongbenchdoc: Benchmarking long-context document understanding with visualizations. Advances in Neural Information Processing Systems , 37:95963–96010, 2024b. Ma, Y., Li, J., Zang, Y., Wu, X., Dong, X., Zhang, P., Cao, Y., Duan, H., Wang, J., Cao, Y., et al. Towards storage-efficient visual document retrieval: An empirical study on reducing patch-level embeddings. arXiv preprint arXiv:2506.04997 , 2025. Mace, Q., Loison, A., and Faysse, M. Vidore benchmark ´ v2: Raising the bar for visual retrieval. arXiv preprint arXiv:2505.17166 , 2025. Mei, L., Mo, S., Yang, Z., and Chen, C. A survey of multimodal retrieval-augmented generation. arXiv preprint arXiv:2504.08748 , 2025. Meng, R., Jiang, Z., Liu, Y., Su, M., Yang, X., Fu, Y., Qin, C., Chen, Z., Xu, R., Xiong, C., et al. Vlm2vec-v2: Advancing multimodal embedding for videos, images, and visual documents. arXiv preprint arXiv:2507.04590 , 2025. Most, A., Winjum, J., Bhattarai, M., Jones, S., Ranasinghe, N. R., Biswas, A., and O'Malley, D. Lost in ocr translation? vision-based approaches to robust document retrieval. In Proceedings of the 2025 ACM Symposium on Document Engineering , pp. 1–10, 2025. Niu, J., Liu, Z., Gu, Z., Wang, B., Ouyang, L., Zhao, Z., Chu, T., He, T., Wu, F., Zhang, Q., et al. Mineru2. 5: A decoupled vision-language model for efficient high-resolution document parsing. arXiv preprint arXiv:2509.22186 , 2025. Ouyang, L., Qu, Y., Zhou, H., Zhu, J., Zhang, R., Lin, Q., Wang, B., Zhao, Z., Jiang, M., Zhao, X., et al. Omnidocbench: Benchmarking diverse pdf document parsing", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/499"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0079", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "with comprehensive annotations. In Proceedings of the Computer Vision and Pattern Recognition Conference , pp. 24838–24848, 2025. Park, C., Jeong, S., Kim, M., Lim, K., and Lee, Y.-H. Scv: Light and effective multi-vector retrieval with sequence compressive vectors. In Proceedings of the 31st Interna tional Conference on Computational Linguistics: Indus try Track , pp. 760–770, 2025. Qian, Y., Lee, J., Duddu, S. M. K., Dai, Z., Brahma, S., Naim, I., Lei, T., and Zhao, V. Y. Multi-vector retrieval as sparse alignment. arXiv preprint arXiv:2211.01267 , 2022. Santhanam, K., Khattab, O., Saad-Falcon, J., Potts, C., and Zaharia, M. Colbertv2: Effective and efficient retrieval via lightweight late interaction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan guage Technologies , pp. 3715–3734, 2022. Shrestha, S., Reddy, N., and Li, Z. Espn: Memory-efficient multi-vector information retrieval. In Proceedings of the 2024 ACM SIGPLAN International Symposium on Memory Management , pp. 95–107, 2024. Team, N. Nomic embed multimodal: Interleaved text, image, and screenshots for visual document retrieval, 2025. URL nomic-embed-multimodal . Teiletche, P., Mace, Q., Conti, M., Loison, A., Viaud, ´ G., Colombo, P., and Faysse, M. Modernvbert: Towards smaller visual document retrievers. arXiv preprint arXiv:2510.01149 , 2025. Thirukovalluru, R., Meng, R., Liu, Y., Su, M., Nie, P., Yavuz, S., Zhou, Y., Chen, W., Dhingra, B., et al. Breaking the batch barrier (b3) of contrastive learning via smart batch mining. arXiv preprint arXiv:2505.11293 , 2025. Tishby, N., Pereira, F. C., and Bialek, W. The information bottleneck method. arXiv preprint physics/0004057 , 2000. Wang, Q., Ding, R., Chen, Z., Wu, W., Wang, S., Xie, P., and Zhao, F. Vidorag: Visual document retrieval-augmented generation via dynamic iterative reasoning agents. arXiv preprint arXiv:2502.18017 , 2025. Wasserman, N., Pony, R., Naparstek, O., Goldfarb, A. R., Schwartz, E., Barzelay, U., and Karlinsky, L. Real-mmrag: A real-world multi-modal retrieval benchmark. arXiv preprint arXiv:2502.12342 , 2025. Xiao, Z., Ma, Q., Gu, M., Chen, C.-c. J., Chen, X., Ordonez, V., and Mohan, V. Metaembed: Scaling multimodal", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/500"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0080", "section": "References", "page_start": 11, "page_end": 11, "type": "Text", "text": "retrieval at test-time with flexible late interaction. arXiv preprint arXiv:2509.18095 , 2025.", "source": "marker_v2", "marker_block_id": "/page/10/Text/326"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0081", "section": "References", "page_start": 11, "page_end": 11, "type": "ListGroup", "text": "Xu, M., Moreira, G., Ak, R., Osmulski, R., Babakhin, Y., Yu, Z., Schifferer, B., and Oldridge, E. Llama nemoretriever colembed: Top-performing text-image retrieval model. arXiv preprint arXiv:2507.05513 , 2025. Yan, Y., Xu, G., Zou, X., Liu, S., Kwok, J., and Hu, X. Docpruner: A storage-efficient framework for multivector visual document retrieval via adaptive patch-level embedding pruning. arXiv preprint arXiv:2509.23883 , 2025. Yan, Y., Huo, J., Feng, G., Ou, M., Cao, Y., Zou, X., Liu, S., Lyu, Y., Huang, Y., Li, J., et al. Unlocking multimodal document intelligence: From current triumphs to future frontiers of visual document retrieval. arXiv preprint arXiv:2602.19961 , 2026a. Yan, Y., Ou, M., Cao, Y., Zou, X., Huo, J., Liu, S., Kwok, J., and Hu, X. Sculpting the vector space: Towards efficient multi-vector visual document retrieval via prunethen-merge framework. arXiv preprint arXiv:2602.19549 , 2026b. Yu, S., Tang, C., Xu, B., Cui, J., Ran, J., Yan, Y., Liu, Z., Wang, S., Han, X., Liu, Z., et al. Visrag: Visionbased retrieval-augmented generation on multi-modality documents. arXiv preprint arXiv:2410.10594 , 2024. Zhang, J., Liu, Y., Wu, Z., Pang, G., Ye, Z., Zhong, Y., Ma, J., Wei, T., Xu, H., Chen, W., et al. Monkeyocr v1. 5 technical report: Unlocking robust document parsing for complex patterns. arXiv preprint arXiv:2511.10390 , 2025a. Zhang, J., Zhang, Q., Wang, B., Ouyang, L., Wen, Z., Li, Y., Chow, K.-H., He, C., and Zhang, W. Ocr hinders rag: Evaluating the cascading impact of ocr on retrievalaugmented generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pp. 17443– 17453, 2025b. Zhang, Q., Wang, B., Huang, V. S.-J., Zhang, J., Wang, Z., Liang, H., He, C., and Zhang, W. Document parsing unveiled: Techniques, challenges, and prospects for structured information extraction. arXiv preprint arXiv:2410.21169 , 2024a. Zhang, Q., Zhang, J., Ren, Z., Ouyang, L., Wen, Z., Niu, J., Qu, Y., Wang, B., Chow, K.-H., He, C., et al. Docr-inspector: Fine-grained and automated evaluation of document parsing with vlm. arXiv preprint arXiv:2512.10619 , 2025c.", "source": "marker_v2", "marker_block_id": "/page/10/ListGroup/325"}
+{"paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0", "chunk_id": "3250cb92-2f69-4e16-9df9-f569224173f0:0082", "section": "References", "page_start": 11, "page_end": 11, "type": "Text", "text": "Zhang, X., Zhang, Y., Xie, W., Li, M., Dai, Z., Long, D., Xie, P., Zhang, M., Li, W., and Zhang, M. Gme: Improving universal multimodal retrieval by multimodal llms. arXiv preprint arXiv:2412.16855 , 2024b.", "source": "marker_v2", "marker_block_id": "/page/10/Text/11"}

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/reference_text_v3.txt

@@ -0,0 +1,23 @@
+[p. 9 | section: References | type: ListGroup]
+Blecher, L., Cucurull, G., Scialom, T., and Stojnic, R. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418 , 2023. Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D., and Liu, Z. M3-embedding: Multi-linguality, multifunctionality, multi-granularity text embeddings through self-knowledge distillation. In Findings of the Association for Computational Linguistics ACL 2024 , pp. 2318–2335, 2024. Faysse, M., Sibille, H., Wu, T., Omrani, B., Viaud, G., Hudelot, C., and Colombo, P. Colpali: Efficient document retrieval with vision language models. arXiv preprint arXiv:2407.01449 , 2024. Feng, H., Wei, S., Fei, X., Shi, W., Han, Y., Liao, L., Lu, J., Wu, B., Liu, Q., Lin, C., et al. Dolphin: Document image parsing via heterogeneous anchor prompting. arXiv preprint arXiv:2505.14059 , 2025. Gu, T., Yang, K., Zhang, K., An, X., Feng, Z., Zhang, Y., Cai, W., Deng, J., and Bing, L. Unime-v2: Mllm-as-ajudge for universal multimodal embedding learning, 2025. URL .
+
+[p. 9 | section: References | type: ListGroup]
+Gunther, M., Sturua, S., Akram, M. K., Mohr, I., Ungureanu, ¨ A., Wang, B., Eslami, S., Martens, S., Werk, M., Wang, N., et al. jina-embeddings-v4: Universal embeddings for multimodal multilingual retrieval. arXiv preprint arXiv:2506.18902 , 2025. Huo, J., Huang, Y., Yan, Y., Pan, Y., Cao, Y., Ou, M., Yu, P. S., and Hu, X. Causalembed: Auto-regressive multivector generation in latent space for visual document embedding. arXiv preprint arXiv:2601.21262 , 2026. ILLUIN. ViDoRe V3: a comprehensive evaluation of retrieval for enterprise use-cases, nov 2025. URL QuentinJG/introducing-vidore-v3 . Jayaram, R., Dhulipala, L., Hadian, M., Lee, J. D., and Mirrokni, V. Muvera: Multi-vector retrieval via fixed dimensional encoding. Advances in Neural Information Processing Systems , 37:101042–101073, 2024. Jha, R., Wang, B., Gunther, M., Mastrapas, G., Sturua, ¨ S., Mohr, I., Koukounas, A., Akram, M. K., Wang, N., and Xiao, H. Jina-colbert-v2: A general-purpose multilingual late interaction retriever. arXiv preprint arXiv:2408.16672 , 2024. Jiang, Z., Meng, R., Yang, X., Yavuz, S., Zhou, Y., and Chen, W. Vlm2vec: Training vision-language models for massive multimodal embedding tasks. arXiv preprint arXiv:2410.05160 , 2024. Khattab, O. and Zaharia, M. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pp. 39–48, 2020. Kim, G., Hong, T., Yim, M., Nam, J., Park, J., Yim, J., Hwang, W., Yun, S., Han, D., and Park, S. Ocr-free document understanding transformer. In European Conference on Computer Vision , pp. 498–517. Springer, 2022. Lee, J., Dai, Z., Duddu, S. M. K., Lei, T., Naim, I., Chang, M.-W., and Zhao, V. Rethinking the role of token retrieval in multi-vector retrieval. Advances in Neural Information Processing Systems , 36:15384–15405, 2023. Liu, Q. and Mao, J. Understanding the multi-vector dense retrieval models. In Proceedings of the 32nd ACM In ternational Conference on Information and Knowledge Management , pp. 4110–4114, 2023. Liu, Y., Zhang, Y., Cai, J., Jiang, X., Hu, Y., Yao, J., Wang, Y., and Xie, W. Lamra: Large multimodal model as your advanced retrieval assistant. In Proceedings of the Computer Vision and Pattern Recognition Conference , pp. 4015–4025, 2025a.
+
+[p. 10 | section: References | type: Text]
+495 496 497 498 499 500 Liu, Z., Liu, Z., Liang, Z., Zhou, J., Xiao, S., Gao, C., Zhang, C. J., and Lian, D. Any information is just worth one single screenshot: Unifying search with visualized information retrieval. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 19238–19261, 2025b.
+
+[p. 10 | section: References | type: ListGroup]
+Ma, X., Lin, S.-C., Li, M., Chen, W., and Lin, J. Unifying multimodal retrieval via document screenshot embedding. In Al-Onaizan, Y., Bansal, M., and Chen, Y.-N. (eds.), Proceedings of the 2024 Conference on Empirical Meth ods in Natural Language Processing , pp. 6492–6505, Miami, Florida, USA, November 2024a. Association for Computational Linguistics. doi: 10.18653/v1/2024. emnlp-main.373. URL org/2024.emnlp-main.373/ . Ma, Y., Zang, Y., Chen, L., Chen, M., Jiao, Y., Li, X., Lu, X., Liu, Z., Ma, Y., Dong, X., et al. Mmlongbenchdoc: Benchmarking long-context document understanding with visualizations. Advances in Neural Information Processing Systems , 37:95963–96010, 2024b. Ma, Y., Li, J., Zang, Y., Wu, X., Dong, X., Zhang, P., Cao, Y., Duan, H., Wang, J., Cao, Y., et al. Towards storage-efficient visual document retrieval: An empirical study on reducing patch-level embeddings. arXiv preprint arXiv:2506.04997 , 2025. Mace, Q., Loison, A., and Faysse, M. Vidore benchmark ´ v2: Raising the bar for visual retrieval. arXiv preprint arXiv:2505.17166 , 2025. Mei, L., Mo, S., Yang, Z., and Chen, C. A survey of multimodal retrieval-augmented generation. arXiv preprint arXiv:2504.08748 , 2025. Meng, R., Jiang, Z., Liu, Y., Su, M., Yang, X., Fu, Y., Qin, C., Chen, Z., Xu, R., Xiong, C., et al. Vlm2vec-v2: Advancing multimodal embedding for videos, images, and visual documents. arXiv preprint arXiv:2507.04590 , 2025. Most, A., Winjum, J., Bhattarai, M., Jones, S., Ranasinghe, N. R., Biswas, A., and O'Malley, D. Lost in ocr translation? vision-based approaches to robust document retrieval. In Proceedings of the 2025 ACM Symposium on Document Engineering , pp. 1–10, 2025. Niu, J., Liu, Z., Gu, Z., Wang, B., Ouyang, L., Zhao, Z., Chu, T., He, T., Wu, F., Zhang, Q., et al. Mineru2. 5: A decoupled vision-language model for efficient high-resolution document parsing. arXiv preprint arXiv:2509.22186 , 2025. Ouyang, L., Qu, Y., Zhou, H., Zhu, J., Zhang, R., Lin, Q., Wang, B., Zhao, Z., Jiang, M., Zhao, X., et al. Omnidocbench: Benchmarking diverse pdf document parsing
+
+[p. 10 | section: References | type: ListGroup]
+with comprehensive annotations. In Proceedings of the Computer Vision and Pattern Recognition Conference , pp. 24838–24848, 2025. Park, C., Jeong, S., Kim, M., Lim, K., and Lee, Y.-H. Scv: Light and effective multi-vector retrieval with sequence compressive vectors. In Proceedings of the 31st Interna tional Conference on Computational Linguistics: Indus try Track , pp. 760–770, 2025. Qian, Y., Lee, J., Duddu, S. M. K., Dai, Z., Brahma, S., Naim, I., Lei, T., and Zhao, V. Y. Multi-vector retrieval as sparse alignment. arXiv preprint arXiv:2211.01267 , 2022. Santhanam, K., Khattab, O., Saad-Falcon, J., Potts, C., and Zaharia, M. Colbertv2: Effective and efficient retrieval via lightweight late interaction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan guage Technologies , pp. 3715–3734, 2022. Shrestha, S., Reddy, N., and Li, Z. Espn: Memory-efficient multi-vector information retrieval. In Proceedings of the 2024 ACM SIGPLAN International Symposium on Memory Management , pp. 95–107, 2024. Team, N. Nomic embed multimodal: Interleaved text, image, and screenshots for visual document retrieval, 2025. URL nomic-embed-multimodal . Teiletche, P., Mace, Q., Conti, M., Loison, A., Viaud, ´ G., Colombo, P., and Faysse, M. Modernvbert: Towards smaller visual document retrievers. arXiv preprint arXiv:2510.01149 , 2025. Thirukovalluru, R., Meng, R., Liu, Y., Su, M., Nie, P., Yavuz, S., Zhou, Y., Chen, W., Dhingra, B., et al. Breaking the batch barrier (b3) of contrastive learning via smart batch mining. arXiv preprint arXiv:2505.11293 , 2025. Tishby, N., Pereira, F. C., and Bialek, W. The information bottleneck method. arXiv preprint physics/0004057 , 2000. Wang, Q., Ding, R., Chen, Z., Wu, W., Wang, S., Xie, P., and Zhao, F. Vidorag: Visual document retrieval-augmented generation via dynamic iterative reasoning agents. arXiv preprint arXiv:2502.18017 , 2025. Wasserman, N., Pony, R., Naparstek, O., Goldfarb, A. R., Schwartz, E., Barzelay, U., and Karlinsky, L. Real-mmrag: A real-world multi-modal retrieval benchmark. arXiv preprint arXiv:2502.12342 , 2025. Xiao, Z., Ma, Q., Gu, M., Chen, C.-c. J., Chen, X., Ordonez, V., and Mohan, V. Metaembed: Scaling multimodal
+
+[p. 11 | section: References | type: Text]
+retrieval at test-time with flexible late interaction. arXiv preprint arXiv:2509.18095 , 2025.
+
+[p. 11 | section: References | type: ListGroup]
+Xu, M., Moreira, G., Ak, R., Osmulski, R., Babakhin, Y., Yu, Z., Schifferer, B., and Oldridge, E. Llama nemoretriever colembed: Top-performing text-image retrieval model. arXiv preprint arXiv:2507.05513 , 2025. Yan, Y., Xu, G., Zou, X., Liu, S., Kwok, J., and Hu, X. Docpruner: A storage-efficient framework for multivector visual document retrieval via adaptive patch-level embedding pruning. arXiv preprint arXiv:2509.23883 , 2025. Yan, Y., Huo, J., Feng, G., Ou, M., Cao, Y., Zou, X., Liu, S., Lyu, Y., Huang, Y., Li, J., et al. Unlocking multimodal document intelligence: From current triumphs to future frontiers of visual document retrieval. arXiv preprint arXiv:2602.19961 , 2026a. Yan, Y., Ou, M., Cao, Y., Zou, X., Huo, J., Liu, S., Kwok, J., and Hu, X. Sculpting the vector space: Towards efficient multi-vector visual document retrieval via prunethen-merge framework. arXiv preprint arXiv:2602.19549 , 2026b. Yu, S., Tang, C., Xu, B., Cui, J., Ran, J., Yan, Y., Liu, Z., Wang, S., Han, X., Liu, Z., et al. Visrag: Visionbased retrieval-augmented generation on multi-modality documents. arXiv preprint arXiv:2410.10594 , 2024. Zhang, J., Liu, Y., Wu, Z., Pang, G., Ye, Z., Zhong, Y., Ma, J., Wei, T., Xu, H., Chen, W., et al. Monkeyocr v1. 5 technical report: Unlocking robust document parsing for complex patterns. arXiv preprint arXiv:2511.10390 , 2025a. Zhang, J., Zhang, Q., Wang, B., Ouyang, L., Wen, Z., Li, Y., Chow, K.-H., He, C., and Zhang, W. Ocr hinders rag: Evaluating the cascading impact of ocr on retrievalaugmented generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pp. 17443– 17453, 2025b. Zhang, Q., Wang, B., Huang, V. S.-J., Zhang, J., Wang, Z., Liang, H., He, C., and Zhang, W. Document parsing unveiled: Techniques, challenges, and prospects for structured information extraction. arXiv preprint arXiv:2410.21169 , 2024a. Zhang, Q., Zhang, J., Ren, Z., Ouyang, L., Wen, Z., Niu, J., Qu, Y., Wang, B., Chow, K.-H., He, C., et al. Docr-inspector: Fine-grained and automated evaluation of document parsing with vlm. arXiv preprint arXiv:2512.10619 , 2025c.
+
+[p. 11 | section: References | type: Text]
+Zhang, X., Zhang, Y., Xie, W., Li, M., Dai, Z., Long, D., Xie, P., Zhang, M., Li, W., and Zhang, M. Gme: Improving universal multimodal retrieval by multimodal llms. arXiv preprint arXiv:2412.16855 , 2024b.

icml26/3250cb92-2f69-4e16-9df9-f569224173f0/sanitization_report.json

@@ -0,0 +1,63 @@
+{
+  "paper_id": "3250cb92-2f69-4e16-9df9-f569224173f0",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 27,
+  "chunk_count": 200,
+  "main_body_chunk_count": 75,
+  "appendix_chunk_count": 117,
+  "reference_chunk_count": 8,
+  "model_text_chars": 42730,
+  "raw_markdown_chars": 191958,
+  "sanitized_chars": 130384,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 27,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 12,
+    "Code": 13,
+    "Equation": 17,
+    "Figure": 3,
+    "FigureGroup": 6,
+    "Footnote": 7,
+    "ListGroup": 10,
+    "PageFooter": 27,
+    "PageHeader": 24,
+    "Picture": 1,
+    "PictureGroup": 4,
+    "SectionHeader": 68,
+    "Table": 7,
+    "TableGroup": 7,
+    "Text": 573
+  },
+  "asset_count_raw": 14,
+  "asset_count_model_kept": 9,
+  "asset_count_rejected": 5,
+  "asset_reject_reasons": {
+    "kept": 9,
+    "too_small": 5
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt"
+}