搜索结果：HuggingFace

Computational prediction of blood-brain barrier (BBB) permeability has emerged as a vital alternative to traditional experimental assays, which are often resource-intensive and lowthroughput to meet the demands of early-stage drug discovery. While early machine learning approaches have shown promise, integration of traditional chemical descriptors with deep learning embeddings remains an underexplored frontier. In this paper, we introduce TITAN-BBB , a multi-modal deep-learning architecture that utilizes tabular, image, and text-based features and combines them using attention mechanisms. To evaluate, we aggregated multiple literature sources to create the largest BBB permeability dataset to date, enabling robust training for both classification and regression tasks. Our results demonstrate that TITAN-BBB achieves 86.5% of balanced accuracy on classification tasks and 0.436 of mean absolute error for regression, outperforming the state-of-the-art by 3.1 percentage points in balanced accuracy and reducing the regression error by 20%. Our approach also outperforms state-of-the-art models in both classification and regression performance, demonstrating the benefits of combining deep and domain-specific representations. The source code is publicly available at https://github.com/pcdslab/TITAN-BBB . The inference-ready model is hosted on Hugging Face at https://huggingface.co/SaeedLab/TITAN-BBB , and the aggregated BBB permeability datasets are available at https://huggingface.co/datasets/SaeedLab/BBBP .

Foundational models that learn the "language" of molecules are essential for accelerating the material and drug discovery. These self-learning models can be trained on a large number of unlabelled molecules, enabling applications like property prediction, molecule design (de novo generation, optimization), and screening for specific functions. However, existing molecular language models are built upon first-generation transformer architectures and are pretrained using masked language modeling, a generic token-level objective that is agnostic to physicochemical and structural molecular properties. Here we introduce MolDe-BERTa , a structure-informed self-supervised molecular encoder that leverages a byte-level Byte-Pair Encoding (BPE) tokenization strategy. MolDeBERTa is pretrained on up to 123 million SMILES molecules from PubChem, representing one of the largest publicly available SMILES-based corpora. To achieve this, we introduce three novel pretraining objectives designed to inject strong inductive biases for molecular properties and structural similarity directly into the latent space, resulting in reduced gap between linguistic chemical representations and physical molecular properties. The model was then systematically investigated across three architectural scales, two dataset sizes, and five distinct pretraining objectives. MolDeBERTa when evaluated on 9 downstream MoleculeNet benchmarks outperformed existing masked language models, achieving up to a 16% reduction in regression error and improvements of up to 3.0 ROC-AUC points on classification benchmarks. MolDeBERTa advances unsupervised encoder-based foundational models at scale both for pretraining data and downstream evaluation, enabling data-efficient chemistry-informed representation learning. The source code is publicly available at https://github.com/pcdslab/MolDeBERTa , and Hugging Face at https://huggingface.co/collections/SaeedLab/moldeberta . All the pretraining datasets are available at https://huggingface.co/datasets/SaeedLab/MolDeBERTa .

Multiomics data analysis is essential for scientific discovery in precision medicine. However, translating analysis results of omics data analysis into novel scientific hypotheses remains a significant challenge. Human experts must manually review analysis results and generate new hypotheses based on extensive and interconnected biomedical prior knowledge, which is subjective and not scalable. While large language models can accelerate the discovery, their reasoning improves when grounded in structured, auditable, and comprehensive biomedical prior knowledge. However, biomedical knowledge is scattered across heterogeneous databases that use diverse and inconsistent nomenclature systems, making it difficult to integrate resources into a unified format for scalable analysis. This fragmentation limits the ability of artificial intelligence systems to fully leverage biomedical data for scientific discovery. We developed BioMedGraphica, a novel all-in-one platform that harmonizes fragmented biomedical resources by integrating 11 entity types and 30 relation types from 43 databases into a unified textual prior knowledge graph containing 2 306 921 entities and 27 232 091 relations. In addition, we present a novel textual-numeric graph (TNG) data structure concept, where textual information captures prior biological knowledge (e.g. transcription start sites, functions, mechanisms), numeric values represent quantitative biomedical features, and the integrated relations can help uncover mechanisms. By bridging prior knowledge with user-specific data, TNG is a novel and ideal data structure for developing novel graph analysis models. The code is available at: https://github.com/FuhaiLiAiLab/BioMedGraphica and BioMedGraphica knowledge graph database can be downloaded from huggingface dataset: https://huggingface.co/datasets/FuhaiLiAiLab/BioMedGraphica.

Skin cancer is the most common type of cancer worldwide, placing a substantial burden on healthcare systems. While numerous computer-assisted diagnostic systems have been developed, most offer limited transparency, making it difficult for healthcare professionals to understand how individual pieces of clinical information influence the diagnostic outcome. In this work, we propose PRISM: Probabilistic Reasoning Interpretable Stepwise Model, an interpretable multimodal skin cancer classification framework that integrates image-based predictions with clinical data using a Bayesian framework. Our approach allows the model to be evaluated with incrementally available metadata, enabling clinicians to interpret how each clinical feature contributes to the diagnostic decision. To address the compounding overconfidence inherent in sequential Bayesian updating, we implement a Stepwise Calibration protocol that dynamically scales with the volume of available evidence, ensuring statistically reliable confidence estimates at every intermediate diagnostic step. We validated the framework with four competitive vision backbones on three datasets. Compared to state-of-the-art attention-based methods, our framework achieves competitive or superior performance, reaching a peak balanced accuracy of [Formula: see text] on the PAD-UFES-20 dataset. Finally, qualitative case studies indicate that PRISM creates a reasoning process that is consistent with clinical logic, suggesting it is a viable alternative for clinical decision support. An interactive demo is available at https://huggingface.co/spaces/pedrobouzon/prism.

Accurate in silico identification of B-cell epitope residues is crucial for antibody design and structure-guided vaccine development. Although recent protein language models and structure-aware methods can capture spatial information of tertiary structure when generating residue embeddings, most existing epitope predictors use these embeddings to perform classification for individual residues one by one, without enforcing spatial continuity for reported epitope residues. Such methods often result in biologically implausible predictions because B-cell epitope residues always cluster together on the antigen surface. We present RoBep, a region-oriented B-cell epitope predictor that explicitly models the spatial clustering of epitope residues. RoBep introduces a novel region constraint mechanism and combines the advanced protein language model ESM-Cambrian with an equivariant graph neural network. Our method outperforms existing structure-based methods on the benchmark dataset, demonstrating improvements of 26%, 45%, 13%, and 43% in F1, Matthews correlation coefficient, area under the precision-recall curve, and AUROC0.1, respectively. In addition to residue-level predictions, RoBep can also provide antibody-antigen binding regions. Importantly, the predicted epitope residues are ensured to be spatially compact, enhancing biological plausibility and practical relevance for immunotherapeutic design. A user-friendly website for using RoBep is provided at https://huggingface.co/spaces/NielTT/RoBep. All datasets, source code used in this work, and implementation instructions of the website are publicly available at https://github.com/YitaoXU/RoBep.

The Genomic Data Commons (GDC) provides access to high quality, harmonized cancer genomics data through a unified curation and analysis platform centered around patient cohorts. While GDC users can interactively create complex cohorts through the graphical Cohort Builder, users (especially new ones) may struggle to find specific cohort descriptors across hundreds of possible fields and properties. However, users may be better able to describe their desired cohort in free-text natural language. We introduce GDC Cohort Copilot, an open-source copilot tool for curating cohorts from the GDC. GDC Cohort Copilot automatically generates the GDC cohort filter corresponding to a user-input natural language description of their desired cohort, before exporting the cohort back to the GDC for further analysis. An interactive user interface allows users to further refine the generated cohort. We develop and evaluate multiple large language models (LLMs) for GDC Cohort Copilot and demonstrate that our locally-served, open-source GDC Cohort LLM achieves better results than GPT-4o prompting in generating GDC cohorts. We implement and share GDC Cohort Copilot as a containerized Gradio app on HuggingFace Spaces, available at https://huggingface.co/spaces/uc-ctds/GDC-Cohort-Copilot. GDC Cohort LLM weights are available at https://huggingface.co/uc-ctds. All source code is available at https://github.com/uc-cdis/gdc-cohort-copilot.

Accurate species-level classification of prokaryotic 16S rRNA sequences remains difficult: existing tools rely on exact alignment, k-mer heuristics, or phylogenetic placement and are limited by incomplete reference databases. Deep learning approaches in microbial genomics have focused largely on whole-genome metagenomics, leaving 16S taxonomy under-supported. We present DeepTaxa, a hybrid CNN-BERT framework that pairs a multiscale CNN with a transformer trained from scratch on the DNABERT-2 BPE vocabulary, producing parallel rank-specific predictions across the seven Linnean ranks. On the Greengenes2 2024.09 test set, DeepTaxa achieves species-level accuracy of 92.96% and F1 of 0.9212 (3-seed mean; cross-seed standard deviation ≤ 0.0008 F1 at every rank), with F1 above 0.99 from domain through class and a species-level expected calibration error of 0.0242. DeepTaxa exceeds DADA2 (90.05%) and QIIME 2 (85.01%) at the species rank on the same held-out test set, with larger gains over the k-mer-based classifiers SINTAX and Kraken 2. Performance degrades smoothly with decreasing training-set similarity (species F1 from 0.95 to 0.45), and a dedicated V3-V4 amplicon checkpoint reaches 87.55% species accuracy from an approximately 420 bp window. Source code, trained checkpoints for full-length 16S and V3-V4 amplicons, curated datasets, and reproducible workflows are publicly available at github.com/systems-genomics-lab/deeptaxa and huggingface.co/systems-genomics-lab/deeptaxa.

Glycans are highly diverse biological sequences, but their functional understanding has lagged behind proteins and nucleic acids. Many glycans remain ambiguously annotated, limiting computational analyses. Existing computational approaches are primarily graph-based, capturing local structural features but struggling to model global patterns and incomplete sequences. We present GlycanGT, a graph-transformer-based pretrained model for glycans. Glycans were represented as graphs of monosaccharides and glycosidic bonds, and the model was pretrained using a masked language modeling objective. GlycanGT demonstrated higher performance than existing methods across 8 benchmark tasks (e.g., 0.844 AUPRC for immunogenicity classification), and its embeddings formed biologically relevant clusters that recovered known N- and O-glycan categories. Moreover, GlycanGT accurately proposed candidates for ambiguous sequences, maintaining >80% top-5 accuracy for both monosaccharide and glycosidic bond predictions under high masking levels. The source code used in this study is available at https://github.com/matsui-lab/GlycanGT and archived on Zenodo (DOI: 10.5281/zenodo.18636040); pretrained model weights are provided via Hugging Face (https://huggingface.co/Akikitani295/GlycanGT). matsui.yusuke.d4@f.mail.nagoya-u.ac.jp. Supplementary data are available at Bioinformatics online.

Computational models are central to cognitive neuroscience, but their rigorous application to experimental datasets is often constrained to a narrow set of canonical models that afford tractable analytical computations. We introduce the HSSM (Hierarchical Sequential Sampling Model) ecosystem, a Python toolbox that democratizes access to a broad, extensible array of neurocognitive process models through hierarchical Bayesian inference. Naturally leveraging simulation-based inference via likelihood surrogates, HSSM enables fast parameter estimation for models lacking closed-form likelihoods. Built atop PyMC and Bambi, HSSM provides a user-friendly formula syntax for specifying hierarchical mixed-effects regressions on model parameters, incorporating trial-by-trial neural or physiological covariates. The ecosystem allows fast model simulation and training data generation, as well as the neural network training utilities to deploy surrogate likelihood networks via HuggingFace. Contributions are designed to benefit not only the single researcher working on a problem, but organically, the entire research community. Together, the tools in the HSSM ecosystem bridge the interests of computational theorists as well as experimentalists, accelerating the cycle from model development to rigorous empirical testing.

BackgroundCommercially available large language models (LLMs) have demonstrated impressive capabilities in processing vast datasets and generating coherent narratives. However, their lack of domain-specific knowledge limits their reliability in clinical applications. This study aimed to develop and evaluate BariatricSurgeryGPT, a fine-tuned LLM specifically tailored for bariatric surgery to provide more accurate and clinically relevant responses to bariatric surgery-related questions.MethodsWe obtained 8764 bariatric surgery research abstracts published between January 1, 2020, and January 1, 2024, from PubMed. These abstracts were preprocessed and tokenized to fine-tune a pre-trained GPT-2 model using PyTorch and HuggingFace frameworks. The model's performance was evaluated using BLEU, METEOR, and ROUGE-1 scores on 20 clinically relevant bariatric surgery questions, each tested across nine temperature settings (0.1-0.9) for both the fine-tuned and baseline GPT-2 models, yielding 360 total evaluation instances.ResultsBariatricSurgeryGPT demonstrated consistent improvements over the baseline GPT-2 model across all metrics. The fine-tuned model achieved a BLEU score of 0.165 (vs 0.147 for baseline, 12.8% improvement), a METEOR score of 0.633 (vs 0.585, 8.2% improvement), and a ROUGE-1 score of 0.267 (vs 0.243, 9.7% improvement). These improvements indicate enhanced precision, recall, and semantic relevance in generating bariatric surgery-specific content.ConclusionBariatricSurgeryGPT represents the first domain-specific LLM for bariatric surgery and demonstrates the feasibility of developing specialty-specific AI tools with improved accuracy for clinical applications. The specialty-specific models could enhance surgical education through interactive learning tools, improve patient communication via personalized educational materials, and support clinical decision-making by providing evidence-based information synthesis.

Phylogenetic trees are ubiquitous and central to biology, but most published trees are available only as visual diagrams and not in the machine-readable Newick format. There are, thus, thousands of published trees in the scientific literature that are unavailable for follow-up analyses, comparisons, and supertree construction. Experts can easily read such diagrams, but the manual construction of a Newick string from a diagram is laborious, error-prone, and time-consuming. Previous attempts to semi-automate the reading of tree images relied on image processing techniques. These often encounter difficulties as typical published tree diagrams contain various graphical elements and annotations that overlap the branches, such as error bars on internal nodes. Here we introduce Treemble, a user-friendly desktop application for generating Newick strings from tree images. The user simply clicks to mark node locations, assisted by a deep learning-based node detection tool, and Treemble algorithmically assembles the tree from the node coordinates alone. Treemble also facilitates the automatic reading of tip name labels and can be used for both rectangular and circular trees. Treemble is a native desktop application for macOS and Windows and is freely available, with documentation, at treemble.org. Source code is available at github.com/John-Allard/Treemble. The trained node detection model is available at huggingface.co/John-Allard/treemble-1.

Electroencephalography (EEG) is a non-invasive method that records brain electrical activity from scalp electrodes, offering millisecond temporal resolution but limited spatial detail due to sparse sensor layouts. We present DiBiMa-EEGSR, a bidirectional Mamba-2 diffusion framework for spatio-temporal EEG super-resolution that reconstructs high-resolution signals from standard low-density recordings without additional hardware. The method formulates super-resolution as conditional generative inference and integrates a diffusion process with a bidirectional state-space backbone to model long-range temporal dependencies with linear complexity. Conditioning on low-resolution inputs, electrode positions and task labels enables anatomically coherent and context-aware reconstruction. A one-step sampling strategy substantially reduces inference time while preserving fidelity. Across two public benchmarks, the approach improves reconstruction accuracy, spatial coherence and spectral preservation over convolutional, transformer-based and prior diffusion models in both spatial and temporal upsampling tasks, providing a scalable pathway toward high-resolution electrophysiological imaging. Code to reproduce ablation experiments, training and evaluation of the proposed BiMa and DiBiMa EEGSR models are available at https://github.com/UgoLomoio/DiBiMa-EEGSR.git. Model weights are available at https://huggingface.co/Ugo96/DiBiMa-EEGSR while an interactive demo for EEG spatial super-resolution using our models can be found at https://huggingface.co/spaces/Ugo96/DiBiMa-EEGSR-Demo.

Recent advances in large language models have enabled the emergence of AI scientists that aim to autonomously analyze biological data and assist scientific discovery. Despite rapid progress, it remains unclear to what extent these systems can extract meaningful biological insights from real experimental data. Existing benchmarks either evaluate reasoning in the absence of data or focus on predefined analytical outputs, failing to reflect realistic, data-driven biological research. Here, we introduce BAISBench (Biological AI Scientist Benchmark), a benchmark for evaluating AI scientists on real single-cell transcriptomic datasets. BAISBench comprises two tasks: cell type annotation across 15 expert-labeled datasets, and scientific discovery through 193 multiple-choice questions derived from biological conclusions reported in 41 published single-cell studies. We evaluated several representative AI scientists using BAISBench and, to provide a human performance baseline, invited five graduate-level bioinformaticians to collectively complete the same tasks. The results show that while current AI scientists fall short of fully autonomous biological discovery, they already demonstrate substantial potential in supporting data-driven biological research. These results position BAISBench as a practical benchmark for characterizing the current capabilities and limitations of AI scientists in biological research. We expect BAISBench to serve as a practical evaluation framework for guiding the development of more capable AI scientists and for helping biologists identify AI systems that can effectively support real-world research workflows. https://github.com/EperLuo/BAISBench, https://huggingface.co/datasets/EperLuo/BaisBench.

Malabar spinach is a nutrient-dense leafy vegetable widely cultivated and consumed in Bangladesh. Its productivity is often compromised by Alternaria leaf spot and straw mite infestations. This work proposes an efficient and interpretable deep learning framework for automatic Malabar spinach leaf disease classification. A curated dataset of Malabar spinach images collected from Habiganj Agricultural University and supplemented with public samples was categorized into three classes: Alternaria, straw mite, and healthy leaves. A lightweight SpinachCNN established a strong baseline, while Spinach-ResSENet, enhanced with squeeze-and-excitation modules, improved channel-wise attention and feature discrimination. A customized Vision Transformer (SpinachViT) and SwinV2-Base were further investigated to assess the benefits of transformer-based architectures under limited data. To mitigate annotation scarcity, we employed SimSiam-based self-supervised pretraining on unlabeled images, followed by supervised fine-tuning with cross-entropy or a hybrid objective combining cross-entropy and supervised contrastive loss. The best-performing domain-optimized model, SimSiam-CBAM-ResNet-50, incorporated Convolutional Block Attention Modules and achieved 97.31% test accuracy, 0.9983 macro ROC-AUC, and low calibration error, while maintaining robustness to Gaussian and salt-and-pepper noise. Although a SwinV2-Base benchmark pretrained on ImageNet-22k reached slightly higher accuracy (97.98%, 98.99% with test-time augmentation), its 86.9M parameters and reliance on large-scale pretraining reduce feasibility for edge deployment. In contrast, the SimSiam-CBAM model offers a more parameter-efficient and deployment-friendly solution for real-world agricultural applications. Model decisions are interpretable via Grad-CAM, Grad-CAM++, and LayerCAM, which consistently highlight biologically relevant lesion regions. The spinach dataset used in this study is publicly available on: https://huggingface.co/datasets/saifullah03/malabar_spinach_leaf_disease_dataset.

This work introduces, LLaMAntino-3-ANITA-8B-Inst-DPO-ITA, a Large Language Model (LLM) adapted for the Italian language based on the Meta-AI LLaMA-3 model family. The original 8B parameter instruction-tuned model is first fine-tuned using the Supervised Fine-tuning (SFT) technique on English datasets to enhance its baseline performance on instruction tasks. Subsequently, a Direct Preference Optimization (DPO) process is applied to align preferences, mitigate unsafe responses, and limiting biases. In the final stage, the model is adapted to the Italian language using a limited amount of high quality Italian language data. This methodology combines the efficiency of QLoRA, for fine-tuning on a smaller portion of the original model weights, with DPO to refine the model's output, adapting the model to the Italian linguistic structure while maintaining computational efficiency. Evaluation on Open LLM benchmarks for both Italian and English languages confirms the model's effectiveness, achieving state- of-the-art performance among Italian LLMs with an average accuracy score of 0.6160 on various Italian text comprehension and question-answering tasks. The model is released via the HuggingFace hub, with usage examples available in the GitHub repository.

Standardizing medication concepts across heterogeneous vocabularies is essential for interoperable analytics and observational research. In the Observational Medical Outcomes Partnership (OMOP) Common Data Model, local drug codes must be mapped to standardized RxNorm concepts, but automated mapping is challenging because drug strings encode clinically critical attributes, including strength, dosage form/route, release characteristics, and brand. We propose THIRAWAT (Terminology Harmonization using Late-Interaction Reranker With Alignment-tuned Transformers), a fine-tuned ColBERTv1 late-interaction reranker, and embed it within THIRAWAT Mapper, a retrieval-reranking pipeline with deterministic tie-breaking and stable ordering. Candidate generation used approximate nearest-neighbor retrieval with a bi-encoder (SapBERT-XLMR or BioLORD-2023). Candidates were reranked by THIRAWAT models that were fine-tuned using one-sided MaxSim and scored at inference using our adapted Bidirectional MaxSim (BiMaxSim) pooling. Finally, a deterministic tie-breaker extracted clinically salient cues, including strength, dosage form/route, release characteristics, and bracketed brand annotations, to resolve near-ties reproducibly. We evaluated three mapping settings: Branded Drugs, Clinical Drugs, and Thai Medicines Terminology (TMT). Using SapBERT-XLMR retrieval with THIRAWAT-Sap-BERT reranking and deterministic tie-breaking, THIRAWAT Mapper achieved MRR@100 values of 0.954 (95% confidence interval [CI], 0.921-0.983), 0.898 (95% CI, 0.866-0.925), and 0.912 (95% CI, 0.891-0.931), outperforming a lexical term frequency-inverse document frequency baseline (0.491, 0.216, and 0.143, respectively). Hits@1 improved to 0.942 (95% CI, 0.899-0.978), 0.859 (95% CI, 0.817-0.898), and 0.868 (95% CI, 0.838-0.896), respectively. BiMaxSim and deterministic tie-breaking improved drug mapping to RxNorm while preserving an efficient runtime profile and stable ordering. Overall, THIRAWAT Mapper offers a pragmatic combination of learned semantic matching and deterministic lexical constraints. Models and code are available on Hugging Face (https://huggingface.co/collections/sidataplus/thirawat) and GitHub (https://github.com/sidataplus/THIRAWAT-mapper).

Pathology report generation has received increasing attention in recent years. However, existing pathology report generation methods still face two main limitations: (1) these methods utilize image-report datasets where some report contents are irrelevant to the given whole slide image (WSI), making it unreliable to train an AI model to generate such content; (2) existing methods trained on data from internal hospitals struggle to generalize to external hospitals due to significant and inevitable textual discrepancies across multiple centers. To address these challenges, we first introduce a multicenter microscopic findings (MMF) dataset, which consists of WSIs and corresponding reports of lung adenocarcinoma from multiple hospitals. The MMF dataset is designed to enable the model to generate relevant report contents and evaluate its generalization to external hospitals. Second, we propose a novel Multicenter In-Context Learning (MICL) method that effectively generalizes the model to external hospitals without the need for fine-tuning. Third, we propose a new WSI-MLLM that incorporates gigabyte-sized WSIs and their image pyramid structure into MLLMs for the first time. Experiments demonstrate that WSI-MLLM significantly outperforms existing pathology report generation methods, and MICL effectively accommodates multicenter discrepancies, achieving consistent performances across hospitals and improving BLEU-4 scores by up to 26.04%. These findings underscore the effectiveness of our WSI-MLLM and MICL in generating high-quality, generalizable pathology reports. This project is available at https://huggingface.co/datasets/yili7eli/MMF.

Machine-generated music (MGM) has become a powerful tool with applications in music therapy, personalised editing, and creative inspiration. However, its unregulated use threatens the entertainment, education, and arts sectors by diminishing the value of high-quality human compositions. Effective detection of machine-generated music (MGMD) is essential, yet progress is hindered by the lack of comprehensive datasets. To address this gap, we introduce M6, a large-scope benchmark dataset designed for MGMD research. M6 is distinguished by its diversity, encompassing multiple generators, domains, languages, cultural contexts, genres, and instruments, all provided in WAV format. We detail the data collection methodology and analysis, alongside baseline performance scores from foundational binary classification models, highlighting the complexity of MGMD and the need for further advancements. M6 serves as a robust resource to support future research in developing effective detection methods. The dataset is available at https://huggingface.co/datasets/yl7622/M6 to promote collaboration and innovation.

Visual information extraction (VIE) from visually rich documents remains challenging due to high layout variability and real-world impairments. Existing methods typically rely on sequential OCR pipelines or end-to-end models requiring extensive labeled data and layout-specific training, limiting their scalability. We propose a classification-guided large vision-language model (LVLM) framework for multi-type VIE that achieves high accuracy with minimal supervision. The approach decouples document-type classification from content extraction and employs in-context learning (ICL)-based dynamic prompt engineering to inject task-specific knowledge, enabling robust zero-shot inference across diverse layouts. From a theoretical perspective, the proposed method can be viewed as a form of conditional computation that reduces task uncertainty and improves information efficiency during prompt-based inference. Evaluated on a real-world bidding dataset with 16 certificate types, our zero-shot method (based on Qwen2.5-VL-7B) outperforms a strong supervised baseline by 18.35 percentage points in F1-score (86.43% vs. 68.08%) and 0.23 in normalized edit distance (0.90 vs. 0.67). Optional domain-specific fine-tuning further improves performance to 93.65% F1 and 0.93 NED, demonstrating superior robustness against seals, watermarks, and low contrast. The framework offers an efficient, scalable solution for complex document understanding in office automation. Code is available at https://github.com/FairmeHIT/Multi-VIE , and fine-tuned models at https://huggingface.co/fairme/Qwen2.5-VL-7B-SFT .