Selected work

Representative Papers

Model Adaptation: Unsupervised Domain Adaptation Without Source Data

Jun 01, 2020Computer Vision and Pattern Recognition

This paper addresses unsupervised model adaptation from a source domain to a target domain without access to any source-domain data or labels—termed *source-free* adaptation—aiming solely to improve the generalization of a pre-trained source model on unlabeled target data. To this end, we propose a Collaborative Class-Conditional Generative Adversarial Network (CC-GAN) framework: it models the semantic structure of the target domain via class-conditional generation; enforces weight constraints derived from the source model to preserve its discriminative capability; and incorporates clustering-driven feature regularization to enhance the discriminability of target-domain representations. Evaluated across multiple cross-domain vision tasks, our method achieves significant performance gains over conventional source-dependent adaptation approaches—using only unlabeled target data. It is the first to empirically validate the effectiveness, robustness, and scalability of model adaptation under the source-free setting.

464 citations53 influentialRead paper

Deep Kronecker Network

Oct 24, 2022Biometrika

Medical imaging analysis faces challenges of limited annotated samples and stringent requirements for model interpretability. Method: This paper proposes a novel deep tensor regression framework based on the Kronecker product, which implicitly enforces piecewise smoothness constraints while integrating a fully convolutional architecture with low-rank tensor priors. Model weights are explicitly generated as the Kronecker product of two low-dimensional factor matrices, enabling native support for multi-modal tensor data (e.g., MRI, fMRI, CT) and unifying classification and regression tasks. Contribution/Results: Evaluated on real-world ADNI MRI data, the method achieves significantly improved generalization stability under extreme few-shot settings (<100 subjects). Moreover, its analytically tractable weight decomposition yields both pixel-level and anatomically grounded interpretations, offering a new paradigm that jointly ensures high predictive performance and clinical trustworthiness.

29 citations4 influentialRead paper

Continuous Input Embedding Size Search For Recommender Systems

Apr 07, 2023Annual International ACM SIGIR Conference on Research and Development in Information Retrieval

To address memory inefficiency caused by fixed high-dimensional embeddings in recommender systems, this paper proposes a memory-constrained continuous embedding dimension optimization framework. Unlike conventional approaches that employ uniform high-dimensional embeddings or existing reinforcement learning (RL)-based methods limited to discrete dimension selection, our work introduces the first continuous-space embedding dimension search paradigm. We design a stochastic walk-driven exploration strategy to efficiently navigate the continuous dimension space, enabling joint optimization of recommendation accuracy and memory efficiency. The method is model-agnostic and plug-and-play. Extensive experiments on two real-world datasets and three state-of-the-art recommendation models demonstrate that our approach achieves superior performance across multiple memory budgets, consistently outperforming discrete-search baselines and establishing new state-of-the-art results.

21 citations1 influentialRead paper

Retrieval-Augmented Generation for Natural Language Processing: A Survey

Jul 18, 2024arXiv.org

To address hallucination, knowledge staleness, and poor domain adaptability in large language models (LLMs), this paper conducts a systematic study of retrieval-augmented generation (RAG). We propose a full-stack RAG framework encompassing retriever design (dense, sparse, and hybrid), query rewriting, context fusion, LLM fine-tuning, and prompt engineering. We introduce the first taxonomy for dynamic knowledge updating and establish a multidimensional evaluation benchmark that balances academic rigor with industrial practicality. Additionally, we release a structured RAG knowledge graph and fully reproducible open-source code. Our contributions significantly enhance RAG’s robustness and maintainability in real-world deployments, providing both theoretical foundations and engineering best practices for knowledge-enhanced generative systems.

15 citationsRead paper

Rethinking Membership Inference Attacks Against Transfer Learning

Jan 20, 2025IEEE Transactions on Information Forensics and Security

In transfer learning, teacher models’ training data are vulnerable to membership inference attacks (MIAs), yet existing work predominantly assumes attackers have black-box or white-box access to the teacher model—overlooking privacy threats when only the student model is accessible in a white-box setting. Method: This paper identifies that representational discrepancies between teacher and student hidden layers can be exploited for MIAs, and proposes the first MIA framework requiring only white-box access to the student model. It employs shadow models to calibrate student-layer representations, models teacher–student representation divergence, and infers teacher training membership via reverse inference—without requiring the teacher model itself. Contribution/Results: The method achieves high inference accuracy across four benchmark datasets and diverse transfer learning tasks. It significantly expands the threat surface of MIAs in transfer learning and provides novel empirical evidence and insights for safeguarding teacher model privacy.

13 citationsRead paper

Recent publications

Latest Papers

Practice Makes Unsafe: Skill Misevolution in Self-Improving LLM Agents

Aug 13, 2026

This work addresses the risk that self-improving large language model agents may inadvertently固化 unsafe successful experiences into reusable strategies during skill evolution, leading to cross-task risk propagation. The paper introduces the concept of “skill mis-evolution” and establishes two evaluation frameworks—SkillMisevo-Gym and SkillMisevo-Bench—to systematically assess this phenomenon. To mitigate such risks, the authors propose SafeEvolve, a mechanism integrating trajectory distillation, skill versioning, adversarial exposure testing, and safety wrappers to enable end-to-end risk tracking, intervention, and remediation throughout the skill lifecycle. Experimental results demonstrate that unsafe skill generation occurs across all 25 evaluated configurations; SafeEvolve reduces unsafe retrieval rates by 26.7 percentage points and new-session harm by 17.3 percentage points, while preserving benign task utility within a narrow margin (±0.4).

0 citationsRead paper

Co-leading Teams Drive Scientific Novelty in Large-scale Research Infrastructures

Aug 13, 2026

This study investigates how collaboration between external users and in-house researchers in large-scale scientific facilities influences scientific novelty—a relationship that remains poorly understood. Leveraging a dataset of 270,000 publications, the authors develop a hybrid machine learning framework to identify three distinct collaboration patterns and quantify their impact on novelty. They find, for the first time, that scientific novelty peaks when in-house researchers participate as co-leads and the ratio of users to staff is balanced. Notably, experienced users benefit significantly only under this specific configuration, providing empirical support for the “knowledge saturation effect.” These findings uncover a dynamic optimization mechanism underlying collaborative structures and offer evidence-based guidance for the design of scientific organizations.

0 citationsRead paper

Refine After Generation: Toward Correct and Concise Patches in LLM-based Program Repair

Aug 13, 2026

This study addresses the pervasive redundancy in program repair patches generated by large language models (LLMs), which are significantly larger and more complex than developer-written patches, thereby undermining their reviewability and practicality. The work presents the first systematic characterization of this redundancy and introduces a novel post-processing paradigm that decouples minimization from patch generation, overcoming limitations of prior approaches reliant on prompting or format constraints. To this end, the authors construct a multi-source patch-pair dataset and train RECAP—a plug-and-play refiner—using supervised fine-tuning, direct preference optimization, and reasoning trace distillation. Experiments demonstrate that RECAP reduces total code changes from +242.14% to +4.24% and net changes from +348.24% to −39.75% relative to developer patches, while preserving correctness and yielding up to 42 additional successfully repaired instances.

0 citationsRead paper

ATOM: Geometry-Aware Microgesture towards Object-Agnostic Tangible Interaction

Aug 12, 2026

This work proposes a cross-object tangible interaction method for microgestures that operates without requiring predefined object models. By integrating fingertip sensing with generative 2D/3D geometric modeling, the system automatically extracts object geometry and leverages ergonomic principles to identify high-usability interaction regions, thereby transforming arbitrary handheld objects into tangible interfaces supporting 0D, 1D, and 2D microgestures. This study presents the first geometry-aware microgesture interaction framework independent of specific object identities and introduces an ergonomics-based mechanism for prioritizing interactive elements. Evaluations across ten everyday objects, including kitchen utensils, demonstrate that the system significantly outperforms ablation baselines in task completion rate, System Usability Scale (SUS) scores, and NASA-TLX cognitive workload metrics, confirming its strong generalization capability and practical utility.

0 citationsRead paper

Benchmarking Trustworthiness of SLMs: Pre-trained vs. Compressed

Aug 12, 2026

It remains unclear how to effectively construct small language models (SLMs) that exhibit high trustworthiness across multiple dimensions—fairness, robustness, privacy, and ethics. This work presents the first systematic comparison between training SLMs from scratch and compressing large language models (LLMs), introducing a comprehensive evaluation framework that assesses trustworthiness along these four axes. The study investigates the impact of pruning, quantization, and knowledge distillation on the trust-related properties of SLMs. Findings reveal that quantization preserves trustworthiness significantly better than pruning. Moreover, SLMs derived via quantization from trustworthy LLMs outperform natively trained small models in both trustworthiness and task adaptability. Further gains in reliability are achievable by incorporating knowledge distillation into the compression pipeline.

0 citationsRead paper