Selected work

Representative Papers

Real-Time Video Prediction With Fast Video Interpolation Model and Prediction Training

Oct 27, 2024International Conference on Information Photonics

To address perceptual latency degradation in real-time video transmission, which impairs interactive user experience, this paper proposes IFRVP—a zero-latency video prediction framework. Methodologically, we design IFRNet, a lightweight convolutional architecture incorporating ELAN-based residual modules to balance accuracy and efficiency, and introduce three novel frame interpolation training paradigms specifically tailored for predictive tasks. Furthermore, we propose a mid-level feature refinement mechanism to enable end-to-end inter-frame interpolation modeling. Experimental results demonstrate that IFRVP achieves a state-of-the-art trade-off between prediction accuracy and inference speed, enabling real-time prediction at over 30 FPS and significantly reducing end-to-end perceptual latency. The source code and demonstration videos are publicly available.

1 citationsRead paper

Dynamic SpectraFormer for Ultra-High-Definition Underwater Image Enhancement

Aug 19, 2026

本文针对水下图像因光线折射和吸收导致的颜色失真、雾霾及可见度低的问题，提出了一种基于频域变换的Dynamic SpectraFormer方法，有效提升了图像质量。

0 citationsRead paper

The Observer Effect in World Models: Invasive Adaptation Corrupts Latent Physics

Feb 12, 2026

Existing methods struggle to determine whether neural world models genuinely internalize physical laws or merely exploit statistical shortcuts, particularly in out-of-distribution settings. This work proposes PhyIP, a non-invasive evaluation protocol that assesses the decodability of key physical quantities—such as internal energy and inverse-square laws—using low-capacity linear probes on frozen self-supervised representations, thereby avoiding fine-tuning or other invasive operations that could disrupt latent physical structure. Experiments in fluid dynamics and orbital mechanics demonstrate that PhyIP efficiently recovers ground-truth physical quantities under out-of-distribution conditions (correlation coefficient ρ > 0.90), whereas invasive approaches cause representational collapse (ρ ≈ 0.05). This study thus reveals, for the first time, the critical influence of evaluation methodology on judgments of physical internalization.

0 citationsRead paper

POP: Online Structural Pruning Enables Efficient Inference of Large Foundation Models

Feb 06, 2026

This work addresses the limitations of existing structured pruning methods, which rely on static strategies and fail to adapt to the dynamically varying sparsity patterns inherent in autoregressive generation, thereby compromising both inference efficiency and accuracy. To overcome this, the authors propose POP, a lightweight online structured pruning framework that operates without preprocessing, retraining, or auxiliary predictors. POP constructs coarse-grained pruning partitions during a prefill phase and applies fine-grained masks to candidate regions during decoding, enabling a context-aware dynamic sparsity mechanism. By partitioning channels into retained, candidate, and pruned regions, POP effectively balances model accuracy and computational efficiency. Experiments demonstrate that POP consistently outperforms state-of-the-art pruning approaches across diverse large models—including LLMs, MoEs, and VLMs—achieving superior performance with lower computational overhead and reduced latency.

0 citationsRead paper

Formally Explaining Decision Tree Models with Answer Set Programming

Jan 07, 2026Electronic Proceedings in Theoretical Computer Science

This work addresses the limited interpretability of complex decision tree models—such as random forests and gradient-boosted trees—in safety-critical applications, where formal justifications for predictions are essential. To overcome this challenge, the authors propose a novel approach based on Answer Set Programming (ASP) that automatically generates diverse logical explanations, including sufficient, contrastive, majority-based, and tree-specific justifications. Compared to SAT-based methods, ASP offers greater flexibility in encoding user preferences and enables the enumeration of all feasible explanations, thereby significantly enhancing both the expressiveness and completeness of the generated justifications. Empirical evaluation across multiple datasets demonstrates the effectiveness of the proposed method in producing varied, formally verifiable explanations, while a systematic analysis highlights its strengths and limitations.

0 citationsRead paper

Recent publications

Latest Papers

Dynamic SpectraFormer for Ultra-High-Definition Underwater Image Enhancement

Aug 19, 2026

本文针对水下图像因光线折射和吸收导致的颜色失真、雾霾及可见度低的问题，提出了一种基于频域变换的Dynamic SpectraFormer方法，有效提升了图像质量。

0 citationsRead paper

The Observer Effect in World Models: Invasive Adaptation Corrupts Latent Physics

Feb 12, 2026

Existing methods struggle to determine whether neural world models genuinely internalize physical laws or merely exploit statistical shortcuts, particularly in out-of-distribution settings. This work proposes PhyIP, a non-invasive evaluation protocol that assesses the decodability of key physical quantities—such as internal energy and inverse-square laws—using low-capacity linear probes on frozen self-supervised representations, thereby avoiding fine-tuning or other invasive operations that could disrupt latent physical structure. Experiments in fluid dynamics and orbital mechanics demonstrate that PhyIP efficiently recovers ground-truth physical quantities under out-of-distribution conditions (correlation coefficient ρ > 0.90), whereas invasive approaches cause representational collapse (ρ ≈ 0.05). This study thus reveals, for the first time, the critical influence of evaluation methodology on judgments of physical internalization.

0 citationsRead paper

POP: Online Structural Pruning Enables Efficient Inference of Large Foundation Models

Feb 06, 2026

This work addresses the limitations of existing structured pruning methods, which rely on static strategies and fail to adapt to the dynamically varying sparsity patterns inherent in autoregressive generation, thereby compromising both inference efficiency and accuracy. To overcome this, the authors propose POP, a lightweight online structured pruning framework that operates without preprocessing, retraining, or auxiliary predictors. POP constructs coarse-grained pruning partitions during a prefill phase and applies fine-grained masks to candidate regions during decoding, enabling a context-aware dynamic sparsity mechanism. By partitioning channels into retained, candidate, and pruned regions, POP effectively balances model accuracy and computational efficiency. Experiments demonstrate that POP consistently outperforms state-of-the-art pruning approaches across diverse large models—including LLMs, MoEs, and VLMs—achieving superior performance with lower computational overhead and reduced latency.

0 citationsRead paper

Formally Explaining Decision Tree Models with Answer Set Programming

Jan 07, 2026Electronic Proceedings in Theoretical Computer Science

This work addresses the limited interpretability of complex decision tree models—such as random forests and gradient-boosted trees—in safety-critical applications, where formal justifications for predictions are essential. To overcome this challenge, the authors propose a novel approach based on Answer Set Programming (ASP) that automatically generates diverse logical explanations, including sufficient, contrastive, majority-based, and tree-specific justifications. Compared to SAT-based methods, ASP offers greater flexibility in encoding user preferences and enables the enumeration of all feasible explanations, thereby significantly enhancing both the expressiveness and completeness of the generated justifications. Empirical evaluation across multiple datasets demonstrates the effectiveness of the proposed method in producing varied, formally verifiable explanations, while a systematic analysis highlights its strengths and limitations.

0 citationsRead paper

Generative AI for Self-Adaptive Systems: State of the Art and Research Roadmap

Dec 04, 2025

Despite growing interest in integrating generative AI (GenAI) into self-adaptive systems (SASs), its advantages and challenges remain poorly understood. To address this gap, this study conducts the first cross-domain systematic literature review spanning software engineering, human-computer interaction, autonomous systems, and AI—augmented by large language model–assisted data analysis and logical reasoning—to assess GenAI’s technical fit within each component of the MAPE-K feedback loop. We propose a novel dual-dimensional framework—“autonomy enhancement” and “human-AI collaboration”—to systematically characterize GenAI’s core strengths (e.g., dynamic modeling, intent understanding, policy generation) and critical limitations (e.g., explainability, real-time responsiveness, trustworthiness assurance). Finally, we derive a forward-looking research roadmap covering technical challenges, validation methodologies, and practical implementation pathways—providing a cohesive foundation for both theoretical advancement and industrial deployment of GenAI-powered SASs.

0 citationsRead paper