Selected work

Representative Papers

A Comprehensive Survey on Vector Database: Storage and Retrieval Technique, Challenge

Oct 18, 2023arXiv.org

Managing and retrieving high-dimensional vector data poses significant challenges, particularly as traditional databases fail to meet performance requirements and the need for tight integration with large language models (LLMs) intensifies. Method: This paper systematically surveys four major approximate nearest neighbor search (ANNS) paradigms—hashing, tree-based indexing, graph-based methods (e.g., HNSW), and quantization (PQ/SQ)—and integrates hybrid optimization strategies. Contribution/Results: It introduces, for the first time, a “Four-Dimensional Methodology” framework tailored for industrial deployment of vector databases, analyzing trade-offs among accuracy, latency, memory footprint, and scalability. The work constructs a structured knowledge graph covering 200+ ANNS algorithms and proposes a novel paradigm for deep synergy between vector databases and LLMs. Collectively, these contributions provide both theoretical foundations and practical guidelines for system selection, architectural design, and development of AI-native database systems.

62 citations3 influentialRead paper

A Comprehensive Study on Visual Token Redundancy for Discrete Diffusion-based Multimodal Large Language Models

Nov 19, 2025

Discrete diffusion-based multimodal large language models (dMLLMs) incur high inference overhead due to full-sequence attention computation at every denoising step; existing acceleration methods largely overlook modality-specific visual token redundancy. Method: We first systematically reveal that visual redundancy occurs exclusively in *from-scratch trained* dMLLMs and exhibits dynamic evolution during long-answer generation—where pruning induces transient information loss but subsequent denoising steps enable self-recovery. Building on this insight, we propose a unified acceleration framework supporting both AR-to-diffusion and from-scratch training paradigms, integrating visual token pruning, layer skipping, KV cache optimization, and efficient sampling. Contribution/Results: We establish a multi-task evaluation benchmark to validate generalizability. Experiments demonstrate substantial reductions in inference latency and GPU memory consumption while preserving generation quality—establishing a new paradigm for efficient dMLLM deployment.

0 citationsRead paper

Aesthetic Image Captioning with Saliency Enhanced MLLMs

Sep 04, 2025

Existing aesthetic image captioning (AIC) methods primarily rely on fine-tuning general-purpose multimodal large language models (MLLMs), but they lack explicit modeling of aesthetic saliency, leading to insufficient attention to aesthetic content. Method: We propose the first end-to-end framework that explicitly integrates aesthetic saliency into MLLMs. It introduces an Image Aesthetic Saliency Module (IASM) to extract fine-grained aesthetic features and constructs an IAS-ViT encoder based on cross-attention to deeply fuse these features with visual representations. Contribution/Results: Our method requires no task-agnostic pretraining or additional annotations. It achieves state-of-the-art performance on major AIC benchmarks, significantly outperforming prior approaches. Experimental results validate that explicit aesthetic modeling is critical for improving both the accuracy and expressiveness of generated captions.

0 citationsRead paper

Joint Enhancement of Relational Reasoning for Long-Context LLMs

Aug 27, 2025

Large language models (LLMs) face critical challenges in long-context understanding, including memory constraints, logical inconsistency, frequent hallucinations, and poor interpretability. To address these issues, we propose JERR—a novel graph-structured reasoning framework that integrates summary-driven text chunking, directed acyclic graph (DAG)-based structural modeling, and Monte Carlo tree search (MCTS). JERR explicitly captures cross-paragraph semantic dependencies and reasoning paths, thereby enhancing transparency and controllability. Extensive experiments demonstrate that JERR consistently outperforms state-of-the-art baselines across ROUGE and F1 metrics. Moreover, it achieves the highest score in human evaluation via LLM-Rater, significantly improving accuracy and robustness on long-text question answering, abstractive summarization, and complex multi-step reasoning tasks.

0 citationsRead paper

City-VLM: Towards Multidomain Perception Scene Understanding via Multimodal Incomplete Learning

Jul 17, 2025

Current large vision-language models (LVLMs) struggle with large-scale outdoor scene understanding due to architectural limitations—namely, reliance on single-view 2D inputs and insufficient support for multi-view (e.g., bird’s-eye and ground-level), multi-modal (e.g., images and point clouds), and cross-scale outdoor perception data. To address this, we introduce SVM-City, the first outdoor-scene benchmark enabling instruction-tuned learning across multiple scales, views, and modalities. We further propose an incomplete multimodal learning framework that achieves robust 2D/3D heterogeneous data fusion via a shared probabilistic latent space, eliminating dependence on modality completeness inherent in conventional concatenation-based methods. Our model incorporates a cross-modal alignment encoder and a probabilistic fusion mechanism. Evaluated on three representative outdoor tasks, it achieves an average 18.14% improvement in question-answering performance over state-of-the-art LVLMs, demonstrating significantly enhanced generalization and deep semantic understanding of urban environments.

0 citationsRead paper

Recent publications

Latest Papers

A Comprehensive Study on Visual Token Redundancy for Discrete Diffusion-based Multimodal Large Language Models

Nov 19, 2025

Discrete diffusion-based multimodal large language models (dMLLMs) incur high inference overhead due to full-sequence attention computation at every denoising step; existing acceleration methods largely overlook modality-specific visual token redundancy. Method: We first systematically reveal that visual redundancy occurs exclusively in *from-scratch trained* dMLLMs and exhibits dynamic evolution during long-answer generation—where pruning induces transient information loss but subsequent denoising steps enable self-recovery. Building on this insight, we propose a unified acceleration framework supporting both AR-to-diffusion and from-scratch training paradigms, integrating visual token pruning, layer skipping, KV cache optimization, and efficient sampling. Contribution/Results: We establish a multi-task evaluation benchmark to validate generalizability. Experiments demonstrate substantial reductions in inference latency and GPU memory consumption while preserving generation quality—establishing a new paradigm for efficient dMLLM deployment.

0 citationsRead paper

Aesthetic Image Captioning with Saliency Enhanced MLLMs

Sep 04, 2025

Existing aesthetic image captioning (AIC) methods primarily rely on fine-tuning general-purpose multimodal large language models (MLLMs), but they lack explicit modeling of aesthetic saliency, leading to insufficient attention to aesthetic content. Method: We propose the first end-to-end framework that explicitly integrates aesthetic saliency into MLLMs. It introduces an Image Aesthetic Saliency Module (IASM) to extract fine-grained aesthetic features and constructs an IAS-ViT encoder based on cross-attention to deeply fuse these features with visual representations. Contribution/Results: Our method requires no task-agnostic pretraining or additional annotations. It achieves state-of-the-art performance on major AIC benchmarks, significantly outperforming prior approaches. Experimental results validate that explicit aesthetic modeling is critical for improving both the accuracy and expressiveness of generated captions.

0 citationsRead paper

Joint Enhancement of Relational Reasoning for Long-Context LLMs

Aug 27, 2025

Large language models (LLMs) face critical challenges in long-context understanding, including memory constraints, logical inconsistency, frequent hallucinations, and poor interpretability. To address these issues, we propose JERR—a novel graph-structured reasoning framework that integrates summary-driven text chunking, directed acyclic graph (DAG)-based structural modeling, and Monte Carlo tree search (MCTS). JERR explicitly captures cross-paragraph semantic dependencies and reasoning paths, thereby enhancing transparency and controllability. Extensive experiments demonstrate that JERR consistently outperforms state-of-the-art baselines across ROUGE and F1 metrics. Moreover, it achieves the highest score in human evaluation via LLM-Rater, significantly improving accuracy and robustness on long-text question answering, abstractive summarization, and complex multi-step reasoning tasks.

0 citationsRead paper

City-VLM: Towards Multidomain Perception Scene Understanding via Multimodal Incomplete Learning

Jul 17, 2025

Current large vision-language models (LVLMs) struggle with large-scale outdoor scene understanding due to architectural limitations—namely, reliance on single-view 2D inputs and insufficient support for multi-view (e.g., bird’s-eye and ground-level), multi-modal (e.g., images and point clouds), and cross-scale outdoor perception data. To address this, we introduce SVM-City, the first outdoor-scene benchmark enabling instruction-tuned learning across multiple scales, views, and modalities. We further propose an incomplete multimodal learning framework that achieves robust 2D/3D heterogeneous data fusion via a shared probabilistic latent space, eliminating dependence on modality completeness inherent in conventional concatenation-based methods. Our model incorporates a cross-modal alignment encoder and a probabilistic fusion mechanism. Evaluated on three representative outdoor tasks, it achieves an average 18.14% improvement in question-answering performance over state-of-the-art LVLMs, demonstrating significantly enhanced generalization and deep semantic understanding of urban environments.

0 citationsRead paper

Self-Guided Process Reward Optimization with Masked Step Advantage for Process Reinforcement Learning

Jul 02, 2025

To address two key challenges in Process Reinforcement Learning (PRL)—high computational overhead from external reward models and the absence of a unified theoretical framework for process-level advantage estimation—this paper proposes a Self-Guided Process Reward Optimization (SG-PRO) framework. Methodologically, it theoretically proves that the policy model itself can intrinsically generate process rewards; introduces Cumulative Process Reward and Masked Stepwise Advantage (MSA) to enable fine-grained, step-level advantage estimation under shared prompts, eliminating the need for auxiliary reward models; and integrates policy entropy regularization with response-length constraints. Experiments demonstrate that SG-PRO achieves 3.4× higher training efficiency and 17.5% higher test accuracy than GRPO, reduces average response length by ~33%, and maintains stable training entropy—effectively mitigating reward hacking.

0 citationsRead paper