document chunking

To address the high computational cost and input redundancy of vision-language models (VLMs) in document understanding, this paper proposes an index-preserving lightweight token pruning framework. Methodologically, it introduces a novel integration of a binarized block-level text classifier with a max-pooling refinement mechanism: first performing coarse-grained filtering to remove non-text background regions, then recovering fragmented text blocks via max-pooling—thereby preserving spatial coherence and maintaining original positional indices. The approach requires no fine-tuning of the backbone model and performs efficient token reduction solely during preprocessing. Experiments on real-world document datasets demonstrate significant inference acceleration (≈40% FLOPs reduction) while retaining baseline accuracy on key tasks such as OCR-enhanced visual question answering. The core contribution is the first realization of high-fidelity, index-aligned token pruning for document images—establishing a practical new paradigm for deploying VLMs under resource constraints.

This study addresses the challenge that implicit semantic connections between non-adjacent paragraphs in academic papers often elude readers. To tackle this issue, the authors propose a novel intra-document long-range semantic linking method grounded in a question-answering framework. Specifically, the approach leverages large language models to generate plausible questions from a given paragraph and retrieves corresponding answer paragraphs within the same paper, thereby explicitly establishing cross-paragraph semantic links. As the first work to implement such intra-paper semantic connections, this method achieves high precision in linking passages separated on average by ten paragraphs on standard benchmarks, significantly enhancing readers’ information retrieval efficiency and overall comprehension experience.

To address semantic discontinuity across paragraphs caused by segmentation-based preprocessing in long-document semantic segmentation, this paper proposes CrossFormer: an end-to-end, paragraph-level semantic segmentation model built upon the Transformer architecture. Its core innovation is a cross-fragment semantic fusion module that dynamically models implicit inter-paragraph semantic dependencies. Notably, CrossFormer is the first to directly integrate a learned semantic segmentation model into retrieval-augmented generation (RAG) systems—replacing heuristic, rule-based chunking. Leveraging self-attention and explicit cross-paragraph semantic alignment, it achieves global semantic modeling and is trained end-to-end under supervised learning. Experiments demonstrate state-of-the-art performance on multiple public semantic segmentation benchmarks. In RAG evaluation, CrossFormer significantly improves recall quality, response coherence, and—most notably—chunk-level semantic consistency (+23.6%).

This study addresses the lack of systematic, cross-domain evaluation of document chunking strategies in dense retrieval systems, which has hindered the effectiveness of retrieval-augmented applications. For the first time, it presents a large-scale benchmark across six knowledge domains, evaluating 36 chunking methods—including fixed-length, semantic, structure-aware, hierarchical, adaptive, and LLM-assisted approaches—using five embedding models and fine-grained relevance annotations generated by large language models. The results demonstrate that content-aware chunking significantly outperforms fixed-length splitting, with optimal strategies varying by domain: paragraph-group chunking achieves the best overall performance (average nDCG@5 = 0.459), dynamic chunking excels in biology, physics, and health domains, while paragraph-level chunking is superior in legal and mathematical contexts. The work also quantifies the trade-off between retrieval effectiveness and computational efficiency, offering practical guidance for system design.

This study addresses the challenging problem of automatically segmenting real-world multi-page heterogeneous document bundles, which often exhibit complex configurations such as page concatenation, misordering, and interleaving. We formally define the document bundle segmentation task for the first time, encompassing boundary detection, document type classification, and page order recovery. To facilitate research in this area, we introduce DocSplit, the first comprehensive benchmark dataset covering five multimodal document categories, along with standardized evaluation metrics tailored to these complex scenarios. Leveraging a multimodal large language model, our approach integrates visual and textual information for end-to-end understanding. Experimental results reveal that existing models perform substantially below desired levels on this task, underscoring the need for further innovation. DocSplit thus establishes a critical foundation for advancing document intelligence in domains such as legal, financial, and healthcare applications.

This work addresses the limitations of existing Retrieval-Augmented Generation (RAG) systems in handling literary texts, which often suffer from fragmented chunking and ambiguous coreference due to their disregard for complex narrative structures, thereby degrading retrieval and generation quality. To overcome this, the authors propose LitSeg, a novel framework that integrates narratological theory into document segmentation for the first time. LitSeg employs multi-stage large language model prompting to extract narrative events, disambiguate storylines, and identify structural turning points, enabling structure-aware text chunking. Furthermore, they introduce LitSeg-Lite, a lightweight variant based on knowledge distillation that supports efficient single-pass inference. Experimental results demonstrate that the proposed approach significantly improves retrieval accuracy, contextual relevance, and downstream question-answering performance, validating the efficacy of narrative-guided segmentation and model distillation.

This work addresses the limitation of traditional retrieval systems, which treat document representation as a static preprocessing step and thus struggle to adapt to downstream tasks. The authors propose AutoIndex, a novel framework that formulates document representation construction as a learnable program synthesis problem. AutoIndex dynamically generates retrieval-oriented representations by searching over executable transformation programs—such as slicing, augmentation, and normalization—and iteratively refines them using validation feedback. By integrating proxy-guided program search with retrieval quality evaluation, the method enables explicit optimization of document representations. Evaluated on the CRUMB benchmark across all eight tasks, AutoIndex consistently outperforms the full-document BM25 baseline, achieving average improvements of 8.4% in Recall@100 and 8.3% in nDCG@10, with peak gains reaching 30.5% and 43.6%, respectively.

This study addresses the degradation of semantic coherence in retrieval indexing caused by fragmentation of structured documents. To mitigate this issue, the authors propose a parser-agnostic pipeline that constructs semantically coherent “evidence units” by aggregating visual elements—such as figures and tables—with their surrounding textual context into unified retrieval blocks. The approach integrates ontology-driven role normalization, a global semantic allocation algorithm, and a graph-based decision-layer verification mechanism. It leverages an extended DoCO ontology, an optimized full similarity matrix, and Neo4j-based rule modeling to ensure compatibility with mainstream parsers like MinerU and Docling. Evaluated on OmniDocBench v1.0, the method achieves a 0.31 improvement in LCS, increases Recall@1 from 0.15 to 0.51 (a 3.4× gain), attains a Recall@1 of 0.47 for text queries, and demonstrates consistent performance across different document parsers.

This work addresses the challenge that existing retrieval systems struggle to ensure all conditions in multi-faceted queries are satisfied within a single long document, particularly when relevant evidence is scattered across different pages. To tackle this, the authors introduce the n-Clue benchmark and propose a “completeness-first” evaluation paradigm that requires retrieved documents to contain evidence for every query condition. Their approach integrates query decomposition, dense and sparse retrieval, vision–language joint modeling, and re-ranking strategies. Experimental results show that even the strongest hybrid system retrieves relevant documents for 81.1% of queries but achieves full condition coverage in only 35.8% of cases; purely vision-based systems perform substantially worse, with coverage below 5.3%, underscoring the difficulty of integrating cross-page, multi-condition evidence.

This work addresses the limitations of traditional document chunking methods, which often ignore user intent and consequently produce fragmented or irrelevant retrieved content, degrading retrieval performance. To overcome this, the authors propose an intent-aware dynamic chunking approach that first leverages a large language model to infer the underlying information needs behind user queries and then applies a dynamic programming algorithm to achieve globally optimal document segmentation, avoiding the suboptimal solutions typical of greedy strategies. As the first study to integrate dynamic programming into intent-driven chunking, the method significantly outperforms existing baselines across six question-answering datasets, improving top-1 retrieval accuracy by 5%–67%, reducing the number of chunks by 40%–60%, and maintaining answer coverage at 93%–100%.