Selected work

Representative Papers

Role-SynthCLIP: A Role Play Driven Diverse Synthetic Data Approach

Nov 07, 2025

Existing synthetic data generation methods primarily focus on scaling data volume, yet suffer from limited semantic diversity and coarse-grained image-text alignment, leading to redundancy and superficial descriptions. To address this, we propose Role-SynthCLIP, a novel framework that introduces a role-playing mechanism—e.g., “Composition Analyst” and “Context Interpreter”—to guide multimodal large language models (MLLMs) in generating semantically rich, fine-grained aligned image-text pairs from diverse perspectives, without increasing dataset size. This enhances both semantic diversity and cross-modal alignment precision. The high-quality synthetic data produced by Role-SynthCLIP is employed for contrastive pretraining of CLIP. Using only 1 million synthetic pairs, our method achieves 64.1% Recall@1 on MS COCO—outperforming the 5-million-pair baseline by 2.8 percentage points. These results empirically validate the efficacy of role-driven prompting in improving synthetic data quality and downstream representation learning.

0 citationsRead paper

Towards Automated Quality Assurance of Patent Specifications: A Multi-Dimensional LLM Framework

Oct 29, 2025

Patent specification quality assessment lacks systematic methodologies, particularly amid the surge in AI-generated content, where multidimensional issues—including regulatory compliance, technical coherence, and figure-text consistency—demand automated solutions. This study proposes the first large language model (LLM)-based, multi-dimensional evaluation framework tailored for patent specifications, integrating rule-based reasoning with LLM capabilities. It comprises four core modules: compliance checking, technical coherence analysis, figure-text consistency verification, and improvement suggestion generation. Evaluated on 160 real-world patents, the three detection modules achieve balanced accuracies of 99.74%, 82.12%, and 91.2%, respectively. The framework uncovers, for the first time, systemic structural misalignment in AI-generated patents and identifies domain- and author-type–specific quality variations.

0 citationsRead paper

Large Language Model's Multi-Capability Alignment in Biomedical Domain

Aug 06, 2025

Addressing the challenge of aligning multiple capabilities—domain knowledge, reasoning, instruction following, and system integration—in biomedical large language models (LLMs), which often interfere with one another and lack sufficient clinical safety, this paper proposes BalancedBio. Our contributions are threefold: (1) the first formal Biomedical Multi-Capability Convergence Theorem; (2) a medical-knowledge-enhanced synthetic data generation method integrating clinical workflow constraints and ontology-based validation; and (3) capability-aware grouped relative policy optimization, combining rule-model hybrid reward shaping with Pareto-optimal analysis to ensure gradient orthogonality and clinical safety. Evaluated on a 0.5B-parameter model, BalancedBio achieves state-of-the-art performance: 80.95% domain knowledge accuracy, 61.94% reasoning score, 67.95% instruction-following fidelity, and 86.7% capability integration. In real-world deployment, it reduces operational costs by 78%, improves diagnostic accuracy by 23%, and attains 89% clinical acceptance.

0 citationsRead paper

Semantic Bridge: Universal Multi-Hop Question Generation via AMR-Driven Graph Synthesis

Aug 06, 2025

Current LLM training is hindered by the scarcity of high-quality, multi-hop reasoning question-answer (QA) pairs—especially in sparse domains such as PubMed and legal texts—while existing methods fail to generate controllable, complex, cross-document reasoning questions, limiting models’ deep semantic understanding. Method: We propose the first general-purpose controllable framework for synthesizing multi-hop QA pairs, leveraging Abstract Meaning Representation (AMR)-driven semantic graph weaving. It constructs cross-document reasoning paths via three strategies: entity bridging, predicate chain expansion, and causal inference. Integrated multimodal AMR analysis and graph-structured synthesis enable fine-grained control over question type, complexity, multilingual support, and cross-domain semantic relation extraction. Results: Experiments show our method outperforms baselines by 18.3%–25.4% across four languages; QA pairs generated from only 200 source documents surpass the performance of those derived from 600 human-annotated samples. Human evaluation confirms a 23.4% increase in question complexity and an 18.7% improvement in answerability.

0 citationsRead paper

Hierarchical Multi-Label Generation with Probabilistic Level-Constraint

Apr 30, 2025

Hierarchical Extreme Multi-Label Classification (HEMLC) faces significant challenges due to the complexity and scale of label taxonomies. To address this, we propose Hierarchical Multi-label Generation (HMG), a novel paradigm that reformulates HEMLC as end-to-end generation of cross-level relevant labels within a given taxonomy. We introduce the first Probabilistic Level Constraint (PLC) mechanism, explicitly controlling the number of generated labels, path length, and hierarchical depth—enabling strong controllability without relying on clustering or other preprocessing steps. Our method jointly leverages taxonomy structural priors and a PLC-guided probabilistic loss, augmented by a taxonomy-aware decoding strategy. Evaluated on standard HEMLC benchmarks, HMG achieves new state-of-the-art performance, improving hierarchical compliance rate by 23.6% over prior methods while demonstrating superior controllability and generation quality.

0 citationsRead paper

Recent publications

Latest Papers

Role-SynthCLIP: A Role Play Driven Diverse Synthetic Data Approach

Nov 07, 2025

Existing synthetic data generation methods primarily focus on scaling data volume, yet suffer from limited semantic diversity and coarse-grained image-text alignment, leading to redundancy and superficial descriptions. To address this, we propose Role-SynthCLIP, a novel framework that introduces a role-playing mechanism—e.g., “Composition Analyst” and “Context Interpreter”—to guide multimodal large language models (MLLMs) in generating semantically rich, fine-grained aligned image-text pairs from diverse perspectives, without increasing dataset size. This enhances both semantic diversity and cross-modal alignment precision. The high-quality synthetic data produced by Role-SynthCLIP is employed for contrastive pretraining of CLIP. Using only 1 million synthetic pairs, our method achieves 64.1% Recall@1 on MS COCO—outperforming the 5-million-pair baseline by 2.8 percentage points. These results empirically validate the efficacy of role-driven prompting in improving synthetic data quality and downstream representation learning.

0 citationsRead paper

Towards Automated Quality Assurance of Patent Specifications: A Multi-Dimensional LLM Framework

Oct 29, 2025

Patent specification quality assessment lacks systematic methodologies, particularly amid the surge in AI-generated content, where multidimensional issues—including regulatory compliance, technical coherence, and figure-text consistency—demand automated solutions. This study proposes the first large language model (LLM)-based, multi-dimensional evaluation framework tailored for patent specifications, integrating rule-based reasoning with LLM capabilities. It comprises four core modules: compliance checking, technical coherence analysis, figure-text consistency verification, and improvement suggestion generation. Evaluated on 160 real-world patents, the three detection modules achieve balanced accuracies of 99.74%, 82.12%, and 91.2%, respectively. The framework uncovers, for the first time, systemic structural misalignment in AI-generated patents and identifies domain- and author-type–specific quality variations.

0 citationsRead paper

Large Language Model's Multi-Capability Alignment in Biomedical Domain

Aug 06, 2025

Addressing the challenge of aligning multiple capabilities—domain knowledge, reasoning, instruction following, and system integration—in biomedical large language models (LLMs), which often interfere with one another and lack sufficient clinical safety, this paper proposes BalancedBio. Our contributions are threefold: (1) the first formal Biomedical Multi-Capability Convergence Theorem; (2) a medical-knowledge-enhanced synthetic data generation method integrating clinical workflow constraints and ontology-based validation; and (3) capability-aware grouped relative policy optimization, combining rule-model hybrid reward shaping with Pareto-optimal analysis to ensure gradient orthogonality and clinical safety. Evaluated on a 0.5B-parameter model, BalancedBio achieves state-of-the-art performance: 80.95% domain knowledge accuracy, 61.94% reasoning score, 67.95% instruction-following fidelity, and 86.7% capability integration. In real-world deployment, it reduces operational costs by 78%, improves diagnostic accuracy by 23%, and attains 89% clinical acceptance.

0 citationsRead paper

Semantic Bridge: Universal Multi-Hop Question Generation via AMR-Driven Graph Synthesis

Aug 06, 2025

Current LLM training is hindered by the scarcity of high-quality, multi-hop reasoning question-answer (QA) pairs—especially in sparse domains such as PubMed and legal texts—while existing methods fail to generate controllable, complex, cross-document reasoning questions, limiting models’ deep semantic understanding. Method: We propose the first general-purpose controllable framework for synthesizing multi-hop QA pairs, leveraging Abstract Meaning Representation (AMR)-driven semantic graph weaving. It constructs cross-document reasoning paths via three strategies: entity bridging, predicate chain expansion, and causal inference. Integrated multimodal AMR analysis and graph-structured synthesis enable fine-grained control over question type, complexity, multilingual support, and cross-domain semantic relation extraction. Results: Experiments show our method outperforms baselines by 18.3%–25.4% across four languages; QA pairs generated from only 200 source documents surpass the performance of those derived from 600 human-annotated samples. Human evaluation confirms a 23.4% increase in question complexity and an 18.7% improvement in answerability.

0 citationsRead paper

Hierarchical Multi-Label Generation with Probabilistic Level-Constraint

Apr 30, 2025

Hierarchical Extreme Multi-Label Classification (HEMLC) faces significant challenges due to the complexity and scale of label taxonomies. To address this, we propose Hierarchical Multi-label Generation (HMG), a novel paradigm that reformulates HEMLC as end-to-end generation of cross-level relevant labels within a given taxonomy. We introduce the first Probabilistic Level Constraint (PLC) mechanism, explicitly controlling the number of generated labels, path length, and hierarchical depth—enabling strong controllability without relying on clustering or other preprocessing steps. Our method jointly leverages taxonomy structural priors and a PLC-guided probabilistic loss, augmented by a taxonomy-aware decoding strategy. Evaluated on standard HEMLC benchmarks, HMG achieves new state-of-the-art performance, improving hierarchical compliance rate by 23.6% over prior methods while demonstrating superior controllability and generation quality.

0 citationsRead paper