Selected work

Representative Papers

MonkeyOCRv2: A Visual-Text Foundation Model for Document AI

Jul 13, 2026

This work addresses the limitations of mainstream vision encoders—pretrained on natural images—in capturing character-level visual details and layout structures essential for document understanding. To bridge this gap, the authors propose a document-oriented vision-language pretraining approach, leveraging a newly constructed dataset, MonkeyDoc v2, comprising 113 million multilingual document images. Their method jointly optimizes image-to-text generation and pixel-level document reconstruction to effectively align visual and textual representations. As the first large-scale foundational model for documents, it employs a frozen vision encoder paired with a lightweight 0.7B-parameter language model, outperforming even 3B-parameter models by 2.8% on MDPBench and achieving state-of-the-art results across eight comprehension benchmarks and five categories of document analysis tasks, surpassing existing methods based on CLIP, DINO, and SAM.

0 citationsRead paper

STEM: Structure-Tracing Evidence Mining for Knowledge Graphs-Driven Retrieval-Augmented Generation

Apr 24, 2026

This work addresses the semantic mismatch caused by structural heterogeneity in knowledge graph question answering and the lack of global structural awareness in existing methods. It reframes multi-hop reasoning as a schema-guided graph search task, constructing an adaptive query schema graph via a semantic-structure projection mechanism. By integrating a Triple-Dependent Graph Neural Network, the approach enables globally guided node anchoring and subgraph retrieval, thereby incorporating global structural information into the retrieval phase for the first time to generate high-quality evidence reasoning graphs. This strategy substantially improves both retrieval accuracy and evidence completeness of multi-hop reasoning paths, achieving state-of-the-art performance across multiple benchmark datasets.

0 citationsRead paper

MDPBench: A Benchmark for Multilingual Document Parsing in Real-World Scenarios

Mar 30, 2026

This work addresses the lack of systematic evaluation benchmarks for document parsing under multilingual, multiscript, and real-world capture conditions—particularly for low-resource languages and non-ideal images. The authors introduce the first high-quality benchmark tailored to these challenging scenarios, encompassing 17 languages, diverse writing systems, and complex imaging conditions. A hybrid annotation pipeline combining expert model assistance with rigorous human verification ensures data fidelity, while a public/private split mitigates data leakage risks. Evaluation results reveal that closed-source models (e.g., Gemini3-Pro) exhibit robust performance, whereas open-source alternatives suffer significant degradation—14.0% on non-Latin scripts and 17.8% on captured documents—highlighting critical gaps in linguistic inclusivity and real-world deployment readiness of current systems.

0 citationsRead paper

CLFEC: A New Task for Unified Linguistic and Factual Error Correction in paragraph-level Chinese Professional Writing

Feb 27, 2026

This work addresses the frequent co-occurrence and mutual influence of linguistic errors (lexical, grammatical, and punctuation) and factual inaccuracies in Chinese professional writing—a challenge inadequately handled by traditional approaches that treat these issues in isolation. To bridge this gap, we introduce the first unified task, CLFEC (Combined Linguistic and Factual Error Correction), and construct a high-quality, multi-domain Chinese dataset spanning politics, finance, law, and medicine. Through systematic exploration of large language model–based paradigms—including prompt engineering, retrieval-augmented generation, and agent-based workflows—we demonstrate that joint in-context correction significantly outperforms decoupled methods. Moreover, when paired with an appropriate base model, the agent-driven workflow further enhances performance, offering both theoretical grounding and practical guidance for industrial-scale, fully automated proofreading systems.

0 citationsRead paper

UTDesign: A Unified Framework for Stylized Text Editing and Generation in Graphic Design Images

Dec 23, 2025

Diffusion models suffer from text style distortion and low recognition accuracy—particularly for small-font and multilingual (especially Chinese) text rendering. To address this, we propose a high-fidelity text generation method tailored for AI-assisted graphic design. Methodologically: (i) we introduce the first end-to-end text style transfer model built upon the DiT architecture; (ii) we construct the first bilingual (Chinese–English) synthetic text-image dataset; and (iii) we design a multimodal conditional encoder integrating a pre-trained text-to-image (T2I) model, an MLLM-based layout planner, and an RGBA transparent foreground generator, enabling background-aware, fully automated Text-to-Design (T2D) synthesis. Experiments demonstrate that our approach achieves state-of-the-art performance among open-source methods in both text accuracy and style consistency, and significantly outperforms commercial closed-source tools in Chinese font fidelity and typographic controllability.

0 citationsRead paper

Recent publications

Latest Papers

MonkeyOCRv2: A Visual-Text Foundation Model for Document AI

Jul 13, 2026

This work addresses the limitations of mainstream vision encoders—pretrained on natural images—in capturing character-level visual details and layout structures essential for document understanding. To bridge this gap, the authors propose a document-oriented vision-language pretraining approach, leveraging a newly constructed dataset, MonkeyDoc v2, comprising 113 million multilingual document images. Their method jointly optimizes image-to-text generation and pixel-level document reconstruction to effectively align visual and textual representations. As the first large-scale foundational model for documents, it employs a frozen vision encoder paired with a lightweight 0.7B-parameter language model, outperforming even 3B-parameter models by 2.8% on MDPBench and achieving state-of-the-art results across eight comprehension benchmarks and five categories of document analysis tasks, surpassing existing methods based on CLIP, DINO, and SAM.

0 citationsRead paper

STEM: Structure-Tracing Evidence Mining for Knowledge Graphs-Driven Retrieval-Augmented Generation

Apr 24, 2026

This work addresses the semantic mismatch caused by structural heterogeneity in knowledge graph question answering and the lack of global structural awareness in existing methods. It reframes multi-hop reasoning as a schema-guided graph search task, constructing an adaptive query schema graph via a semantic-structure projection mechanism. By integrating a Triple-Dependent Graph Neural Network, the approach enables globally guided node anchoring and subgraph retrieval, thereby incorporating global structural information into the retrieval phase for the first time to generate high-quality evidence reasoning graphs. This strategy substantially improves both retrieval accuracy and evidence completeness of multi-hop reasoning paths, achieving state-of-the-art performance across multiple benchmark datasets.

0 citationsRead paper

MDPBench: A Benchmark for Multilingual Document Parsing in Real-World Scenarios

Mar 30, 2026

This work addresses the lack of systematic evaluation benchmarks for document parsing under multilingual, multiscript, and real-world capture conditions—particularly for low-resource languages and non-ideal images. The authors introduce the first high-quality benchmark tailored to these challenging scenarios, encompassing 17 languages, diverse writing systems, and complex imaging conditions. A hybrid annotation pipeline combining expert model assistance with rigorous human verification ensures data fidelity, while a public/private split mitigates data leakage risks. Evaluation results reveal that closed-source models (e.g., Gemini3-Pro) exhibit robust performance, whereas open-source alternatives suffer significant degradation—14.0% on non-Latin scripts and 17.8% on captured documents—highlighting critical gaps in linguistic inclusivity and real-world deployment readiness of current systems.

0 citationsRead paper

CLFEC: A New Task for Unified Linguistic and Factual Error Correction in paragraph-level Chinese Professional Writing

Feb 27, 2026

This work addresses the frequent co-occurrence and mutual influence of linguistic errors (lexical, grammatical, and punctuation) and factual inaccuracies in Chinese professional writing—a challenge inadequately handled by traditional approaches that treat these issues in isolation. To bridge this gap, we introduce the first unified task, CLFEC (Combined Linguistic and Factual Error Correction), and construct a high-quality, multi-domain Chinese dataset spanning politics, finance, law, and medicine. Through systematic exploration of large language model–based paradigms—including prompt engineering, retrieval-augmented generation, and agent-based workflows—we demonstrate that joint in-context correction significantly outperforms decoupled methods. Moreover, when paired with an appropriate base model, the agent-driven workflow further enhances performance, offering both theoretical grounding and practical guidance for industrial-scale, fully automated proofreading systems.

0 citationsRead paper

UTDesign: A Unified Framework for Stylized Text Editing and Generation in Graphic Design Images

Dec 23, 2025

Diffusion models suffer from text style distortion and low recognition accuracy—particularly for small-font and multilingual (especially Chinese) text rendering. To address this, we propose a high-fidelity text generation method tailored for AI-assisted graphic design. Methodologically: (i) we introduce the first end-to-end text style transfer model built upon the DiT architecture; (ii) we construct the first bilingual (Chinese–English) synthetic text-image dataset; and (iii) we design a multimodal conditional encoder integrating a pre-trained text-to-image (T2I) model, an MLLM-based layout planner, and an RGBA transparent foreground generator, enabling background-aware, fully automated Text-to-Design (T2D) synthesis. Experiments demonstrate that our approach achieves state-of-the-art performance among open-source methods in both text accuracy and style consistency, and significantly outperforms commercial closed-source tools in Chinese font fidelity and typographic controllability.

0 citationsRead paper