Reconstruction and Reflection of Positive Experiences through Resurfacing Laughter-indexed Everyday Moments
研究通过检测笑声来索引日常生活中的积极时刻，利用LaughAnchor系统帮助用户记录、回顾和反思这些时刻，从而支持正面经历的重构与反思。
0 citationsRead paper
Selected work
Representative Papers
Multi-Perspective Subimage CLIP with Keyword Guidance for Remote Sensing Image-Text Retrieval
This work addresses the limitations of existing remote sensing image-text retrieval methods, which rely on coarse-grained global alignment and overlook dense multi-scale semantic content in images, while full fine-tuning incurs high computational costs and risks catastrophic forgetting. To overcome these issues, we propose MPS-CLIP, a novel framework featuring a keyword-guided multi-view subregion alignment mechanism. Specifically, keywords generated by a large language model guide SamGeo to segment semantically meaningful subregions; fine-grained alignment is then achieved through a Gated Global Attention (G²A) adapter and a dynamic maximum-response view selection strategy. The model employs lightweight adapters and is jointly optimized with hybrid contrastive and weighted triplet losses. Experiments show that MPS-CLIP achieves average recall rates of 35.18% and 48.40% on RSICD and RSITMD, respectively, significantly outperforming state-of-the-art methods.
0 citationsRead paper
Can We Predict the Next Question? A Collaborative Filtering Approach to Modeling User Behavior
Existing conversational recommendation systems predominantly rely on static user modeling, failing to capture the dynamic evolution of user interests over interaction sequences. To address this, we propose CFQP—a collaborative filtering–enhanced question prediction framework—that models personalized temporal behavioral patterns via a dedicated memory module and incorporates a graph neural network–based preference propagation mechanism to jointly leverage individual historical interactions and collaborative signals from behaviorally similar users. CFQP is the first approach to achieve organic integration of language modeling and sequential behavioral modeling, significantly improving both predictive accuracy and behavioral plausibility in next-question prediction. This advances the paradigm for proactive, personalized conversational systems. Extensive experiments on multiple real-world datasets demonstrate that CFQP consistently outperforms state-of-the-art baselines.
0 citationsRead paper
SRLF: An Agent-Driven Set-Wise Reflective Learning Framework for Sequential Recommendation
Existing LLM-based recommender systems predominantly adopt pointwise item scoring, resulting in coarse-grained user preference modeling and rigid item semantic representations. To address this, we propose the Intelligent Reflective Learning Framework (IRLF), which establishes a closed-loop “evaluate–verify–reflect” mechanism to enable holistic judgment and relational modeling over item sets—marking the first such approach in recommendation. IRLF synergistically integrates LLM agents, in-context learning, and structured relational reasoning, overcoming the limitation of isolated item modeling inherent in conventional sequential recommendation. Extensive experiments on multiple standard sequential recommendation benchmarks demonstrate that IRLF significantly outperforms state-of-the-art baselines, especially in long-tail preference identification and cross-category behavioral modeling. These results validate the effectiveness of set-level reflective learning in enhancing both recommendation accuracy and semantic flexibility.
0 citationsRead paper
Semantic Palette-Guided Color Propagation
Traditional color propagation methods rely on low-level visual features and lack content awareness; while existing semantic approaches incorporate high-level information, they often induce unnatural global color casts. This paper proposes Semantic Palette, the first framework that jointly models user-provided local edits and image semantic segmentation to achieve semantically consistent pixel-level color transfer. Our method integrates differentiable palette optimization with energy minimization, enabling end-to-end training. Evaluated across diverse image categories, it significantly improves both accuracy and naturalness of local color editing: quantitative metrics surpass state-of-the-art by 12.3%, and user studies demonstrate a 37% increase in perceived realism. The core contributions are (i) a novel paradigm for constructing semantic palettes, and (ii) a unified formulation that jointly enforces local edit fidelity and global color consistency through semantic guidance.
0 citationsRead paper
Recent publications
Latest Papers
Reconstruction and Reflection of Positive Experiences through Resurfacing Laughter-indexed Everyday Moments
研究通过检测笑声来索引日常生活中的积极时刻，利用LaughAnchor系统帮助用户记录、回顾和反思这些时刻，从而支持正面经历的重构与反思。
0 citationsRead paper
Multi-Perspective Subimage CLIP with Keyword Guidance for Remote Sensing Image-Text Retrieval
This work addresses the limitations of existing remote sensing image-text retrieval methods, which rely on coarse-grained global alignment and overlook dense multi-scale semantic content in images, while full fine-tuning incurs high computational costs and risks catastrophic forgetting. To overcome these issues, we propose MPS-CLIP, a novel framework featuring a keyword-guided multi-view subregion alignment mechanism. Specifically, keywords generated by a large language model guide SamGeo to segment semantically meaningful subregions; fine-grained alignment is then achieved through a Gated Global Attention (G²A) adapter and a dynamic maximum-response view selection strategy. The model employs lightweight adapters and is jointly optimized with hybrid contrastive and weighted triplet losses. Experiments show that MPS-CLIP achieves average recall rates of 35.18% and 48.40% on RSICD and RSITMD, respectively, significantly outperforming state-of-the-art methods.
0 citationsRead paper
Can We Predict the Next Question? A Collaborative Filtering Approach to Modeling User Behavior
Existing conversational recommendation systems predominantly rely on static user modeling, failing to capture the dynamic evolution of user interests over interaction sequences. To address this, we propose CFQP—a collaborative filtering–enhanced question prediction framework—that models personalized temporal behavioral patterns via a dedicated memory module and incorporates a graph neural network–based preference propagation mechanism to jointly leverage individual historical interactions and collaborative signals from behaviorally similar users. CFQP is the first approach to achieve organic integration of language modeling and sequential behavioral modeling, significantly improving both predictive accuracy and behavioral plausibility in next-question prediction. This advances the paradigm for proactive, personalized conversational systems. Extensive experiments on multiple real-world datasets demonstrate that CFQP consistently outperforms state-of-the-art baselines.
0 citationsRead paper
SRLF: An Agent-Driven Set-Wise Reflective Learning Framework for Sequential Recommendation
Existing LLM-based recommender systems predominantly adopt pointwise item scoring, resulting in coarse-grained user preference modeling and rigid item semantic representations. To address this, we propose the Intelligent Reflective Learning Framework (IRLF), which establishes a closed-loop “evaluate–verify–reflect” mechanism to enable holistic judgment and relational modeling over item sets—marking the first such approach in recommendation. IRLF synergistically integrates LLM agents, in-context learning, and structured relational reasoning, overcoming the limitation of isolated item modeling inherent in conventional sequential recommendation. Extensive experiments on multiple standard sequential recommendation benchmarks demonstrate that IRLF significantly outperforms state-of-the-art baselines, especially in long-tail preference identification and cross-category behavioral modeling. These results validate the effectiveness of set-level reflective learning in enhancing both recommendation accuracy and semantic flexibility.
0 citationsRead paper
Semantic Palette-Guided Color Propagation
Traditional color propagation methods rely on low-level visual features and lack content awareness; while existing semantic approaches incorporate high-level information, they often induce unnatural global color casts. This paper proposes Semantic Palette, the first framework that jointly models user-provided local edits and image semantic segmentation to achieve semantically consistent pixel-level color transfer. Our method integrates differentiable palette optimization with energy minimization, enabling end-to-end training. Evaluated across diverse image categories, it significantly improves both accuracy and naturalness of local color editing: quantitative metrics surpass state-of-the-art by 12.3%, and user studies demonstrate a 37% increase in perceived realism. The core contributions are (i) a novel paradigm for constructing semantic palettes, and (ii) a unified formulation that jointly enforces local edit fidelity and global color consistency through semantic guidance.
0 citationsRead paper