Sichuan University

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

FedConv: A Learning-on-Model Paradigm for Heterogeneous Federated Clients

Jun 03, 2024ACM SIGMOBILE International Conference on Mobile Systems, Applications, and Services

To address the uneven computational burden imposed by client-side resource heterogeneity in federated learning (FL), this paper proposes FedConv. It trains lightweight submodels directly in compressed convolutional form, eliminating decompression overhead. FedConv introduces the novel “learning-on-model” paradigm—the first approach enabling end-to-end training of compressed submodels. It further designs a transposed-convolution-based expansion mechanism to unify aggregation of heterogeneous submodels while preserving personalized parameters. Complemented by joint optimization on the server using a small public dataset, FedConv achieves an average accuracy improvement of 35.2% across six benchmark datasets, while reducing computational cost by 33.1% and communication cost by 24.8%, significantly outperforming existing FL methods.

7 citationsRead paper

SHANG++: Robust Stochastic Acceleration under Multiplicative Noise

Mar 10, 2026

本文针对乘性噪声下Nesterov加速的不稳定性问题，通过离散化Hessian驱动的加速梯度流提出SHANG及改进版SHANG++方法，增强了鲁棒性和收敛速度。

2 citationsRead paper

CogFlow: Bridging Perception and Reasoning through Knowledge Internalization for Visual Mathematical Problem Solving

Jan 05, 2026arXiv.org

This work addresses the challenge that existing multimodal large language models struggle to effectively integrate perceptual understanding with symbolic reasoning in visual mathematical tasks, often failing to faithfully leverage visual cues. To bridge this gap, the authors propose CogFlow—a human cognition-inspired three-stage framework (Perception → Internalization → Reasoning)—which explicitly introduces a knowledge internalization stage to establish a hierarchical flow from perception to reasoning. CogFlow is optimized through a synergistic combination of visual reward mechanisms, a knowledge internalization reward model, and a visual gating strategy, trained on MathCog, a newly curated high-quality dataset comprising over 120,000 samples. Experimental results demonstrate that CogFlow significantly outperforms current state-of-the-art methods across multiple visual mathematical reasoning benchmarks, substantially enhancing the model’s visually grounded reasoning capabilities.

1 citationsRead paper

Multi-Pair Temporal Sentence Grounding via Multi-Thread Knowledge Transfer Network

Dec 20, 2024arXiv.org

Existing temporal sentence grounding (TSG) methods train on untrimmed video–sentence query pairs independently, neglecting inter-pair correlations—leading to knowledge redundancy, inefficient training, and limited generalization. This paper proposes a novel multi-pair joint TSG paradigm, enabling a single model to collaboratively optimize multiple video–query pairs simultaneously. To this end, we design a multi-threaded knowledge transfer network featuring: (i) cross-modal contrastive learning to strengthen fine-grained alignment; (ii) a dual-granularity prototype matching mechanism—operating at both object/phrase level (spatial) and action/sentence level (temporal); and (iii) adaptive threshold-based hard negative mining coupled with self-supervised representation learning. Extensive experiments on multiple benchmarks demonstrate substantial improvements in both grounding accuracy and inference efficiency, achieving new state-of-the-art performance. Ablation studies confirm the effectiveness of inter-pair knowledge transfer and the model’s strong generalization capability across diverse queries and videos.

1 citationsRead paper

Recent publications