Recent multimodal large language models (MLLMs) still struggle with long document understanding due to two fundamental challenges: information interference from abundant irrelevant content, and the quadratic computational cost of Transformer-based architectures. Existing approaches primarily fall into two categories: token compression, which sacrifices fine-grained details; and introducing external retrievers, which increase system complexity and prevent end-to-end optimization. To address these issues, we conduct an in-depth analysis and observe that MLLMs exhibit a human-like coarse-to-fine reasoning pattern: early Transformer layers attend broadly across the document, while deeper layers focus on relevant evidence pages. Motivated by this insight, we posit that the inherent evidence localization capabilities of MLLMs can be explicitly leveraged to perform retrieval during the reasoning process, facilitating efficient long document understanding. To this end, we propose URaG, a simple-yet-effective framework that Unifies Retrieval and Generation within a single MLLM. URaG introduces a lightweight cross-modal retrieval module that converts the early Transformer layers into an efficient evidence selector, identifying and preserving the most relevant pages while discarding irrelevant content. This design enables the deeper layers to concentrate computational resources on pertinent information, improving both accuracy and efficiency. Extensive experiments demonstrate that URaG achieves state-of-the-art performance while reducing computational overhead by 44-56%.

Visual Information Extraction (VIE) has experienced substantial growth and heightened interest due to its pivotal role in intelligent document processing. However, most existing related pre-trained models typically can only process the data from a certain (set of) language(s)—often just English, representing a distinct limitation. To solve it, we present a L anguage-subst i tutable L ayout-image T ransformer (LiLTv2). It can be pre-trained just once on mono-lingual documents and then collaborate with off-the-shelf textual models in other languages during fine-tuning. Firstly, LiLTv2 utilizes a new dual-stream model architecture, one stream for substitutable text information and the other for layout and image information. Then, LiLTv2 has improved upon the optimization strategy and the diverse tasks adopted in the pre-training stage. Finally, we innovatively propose a teacher-student knowledge distillation learning with segment-level multi-modal features named SegKD. Extensive experimental results on widely used benchmarks can demonstrate the superior effectiveness of our method.

Document pair extraction aims to identify key and value entities as well as their relationships from visually-rich documents. Most existing methods divide it into two separate tasks: semantic entity recognition (SER) and relation extraction (RE). However, simply concatenating SER and RE serially can lead to severe error propagation, and it fails to handle cases like multi-line entities in real scenarios. To address these issues, this paper introduces a novel framework, PEneo (Pair Extraction new decoder option), which performs document pair extraction in a unified pipeline, incorporating three concurrent sub-tasks: line extraction, line grouping, and entity linking. This approach alleviates the error accumulation problem and can handle the case of multi-line entities. Furthermore, to better evaluate the model's performance and to facilitate future research on pair extraction, we introduce RFUND, a re-annotated version of the commonly used FUNSD and XFUND datasets, to make them more accurate and cover realistic situations. Experiments on various benchmarks demonstrate PEneo's superiority over previous pipelines, boosting the performance by a large margin (e.g., 19.89%-22.91% F1 score on RFUND-EN) when combined with various backbones like LiLT and LayoutLMv3, showing its effectiveness and generality. Codes and the new annotations are available at https://github.com/ZeningLin/PEneo.