Model merging has emerged as a promising approach for unifying independently fine-tuned models into an integrated framework, significantly enhancing computational efficiency in multi-task learning. Recently, several SVD-based techniques have been introduced to exploit low-rank structures for enhanced merging, but their reliance on such manually designed rank selection often leads to cross-task interference and suboptimal performance. In this paper, we propose AdaRank, a novel model merging framework that adaptively selects the most beneficial singular directions of task vectors to merge multiple models. W e empirically show that the dominant singular components of task vectors can cause critical interference with other tasks, and that naive truncation across tasks and layers degrades performance. In contrast, AdaRank dynamically prunes the singular components that cause interference and offers an optimal amount of information to each task vector by learning to prune ranks during test-time via entropy minimization. Our analysis demonstrates that such method mitigates detrimental overlaps among tasks, while empirical results show that AdaRank consistently achieves state-of-the-art performance with various backbones and number of tasks, reducing the performance gap between fine-tuned models to nearly 1%.

Video Moment Retrieval (MR) aims to localize moments within a video based on a given natural language query. Given the prevalent use of platforms like YouTube for information retrieval, the demand for MR techniques is significantly growing. Recent DETR-based models have made notable advances in performance but still struggle with accurately localizing short moments. Through data analysis, we identified limited feature diversity in short moments, which motivated the development of MomentMix. MomentMix employs two augmentation strategies: ForegroundMix and BackgroundMix, each enhancing the feature representations of the foreground and background, respectively. Additionally, our analysis of prediction bias revealed that short moments particularly struggle with accurately predicting their center positions of moments. To address this, we propose a Length-Aware Decoder, which conditions length through a novel bipartite matching process. Our extensive studies demonstrate the efficacy of our length-aware approach, especially in localizing short moments, leading to improved overall performance. Our method surpasses state-of-the-art DETR-based methods on benchmark datasets, achieving the highest R1 and mAP on QVHighlights and the highest R1@0.7 on TACoS and Charades-STA (such as a 2.46% gain in R1@0.7 and a 2.57% gain in mAP average for QVHighlights). The code is available at https://github.com/sjpark5800/LA-DETR.

Model merging aims to build a multi-task learner by combining the parameters of individually fine-tuned models without additional training. While a straightforward approach is to average model parameters across tasks, this often results in suboptimal performance due to interference among parameters across tasks. In this paper, we present intriguing results that weight averaging implicitly induces task vectors centered around the weight averaging itself and that applying a low-rank approximation to these centered task vectors significantly improves merging performance. Our analysis shows that centering the task vectors effectively separates core task-specific knowledge and nuisance noise within the fine-tuned parameters into the top and lower singular vectors, respectively, allowing us to reduce inter-task interference through its low-rank approximation. We evaluate our method on eight image classification tasks, demonstrating that it outperforms prior methods by a significant margin, narrowing the performance gap with traditional multi-task learning to within 1-3%

Score distillation sampling (SDS) has emerged as an effective framework in text-driven 3D editing tasks due to its inherent 3D consistency. However, existing SDS-based 3D editing methods suffer from extensive training time and lead to low-quality results, primarily because these methods deviate from the sampling dynamics of diffusion models. In this paper, we propose DreamCatalyst, a novel framework that interprets SDS-based editing as a diffusion reverse process. Our objective function considers the sampling dynamics, thereby making the optimization process of DreamCatalyst an approximation of the diffusion reverse process in editing tasks. DreamCatalyst aims to reduce training time and improve editing quality. DreamCatalyst presents two modes: (1) a faster mode, which edits the NeRF scene in only about 25 minutes, and (2) a high-quality mode, which produces superior results in less than 70 minutes. Specifically, our high-quality mode outperforms current state-of-the-art NeRF editing methods both in terms of speed and quality. See more extensive results on our project page: https://dream-catalyst.github.io.

The recent advent of play-to-earn (P2E) systems in massively multiplayer online role-playing games (MMORPGs) has made in-game goods interchangeable with real-world values more than ever before. The goods in the P2E MMORPGs can be directly exchanged with cryptocurrencies such as Bitcoin, Ethereum, or Klaytn via blockchain networks. Unlike traditional in-game goods, once they had been written to the blockchains, P2E goods cannot be restored by the game operation teams even with chargeback fraud such as payment fraud, cancellation, or refund. To tackle the problem, we propose a novel chargeback fraud prediction method, PU GNN, which leverages graph attention networks with PU loss to capture both the players' in-game behavior with P2E token transaction patterns. With the adoption of modified GraphSMOTE, the proposed model handles the imbalanced distribution of labels in chargeback fraud datasets. The conducted experiments on three real-world P2E MMORPG datasets demonstrate that PU GNN achieves superior performances over previously suggested methods.

Message-passing Graph Neural Networks (GNNs), which collect information from adjacent nodes achieve dismal performance on heterophilic graphs. Various schemes have been proposed to solve this problem, and propagating signed information on heterophilic edges has gained great attention. Recently, some works provided theoretical analysis that signed propagation always leads to performance improvement under a binary class scenario. However, we notice that prior analyses do not align well with multi-class benchmark datasets. This paper provides a new understanding of signed propagation for multi-class scenarios and points out two drawbacks in terms of message-passing and parameter update: (1) Message-passing: if two nodes belong to different classes but have a high similarity, signed propagation can decrease the separability. (2) Parameter update: the prediction uncertainty (e.g., conflict evidence) of signed neighbors increases during training, which can impede the stability of the algorithm. Based on the observation, we introduce two novel strategies for improving signed propagation under multi-class graphs. The proposed scheme combines calibration to secure robustness while reducing uncertainty. We show the efficacy of our theorem through extensive experiments on six benchmark graph datasets.

Graph neural networks (GNNs) are commonly used in semi-supervised settings. Previous research has primarily focused on finding appropriate graph filters (e.g. aggregation methods) to perform well on both homophilic and heterophilic graphs. While these methods are effective, they can still suffer from the sparsity of node features, where the initial data contain few non-zero elements. This can lead to overfitting in certain dimensions in the first projection matrix, as training samples may not cover the entire range of graph filters (hyperplanes). To address this, we propose a novel data augmentation strategy. Specifically, by flipping both the initial features and hyperplane, we create additional space for training, which leads to more precise updates of the learnable parameters and improved robustness for unseen features during inference. To the best of our knowledge, this is the first attempt to mitigate the overfitting caused by the initial features. Extensive experiments on real-world datasets show that our proposed technique increases node classification accuracy by up to 46.5% relatively.

Graph Neural Networks (GNNs) have proven to be powerful in many graph-based applications. However, they fail to generalize well under heterophilic setups, where neighbor nodes have different labels. To address this challenge, we employ a confidence ratio as a hyper-parameter, assuming that some of the edges are disassortative (heterophilic). Here, we propose a two-phased algorithm. Firstly, we determine edge coefficients through subgraph matching using a supplementary module. Then, we apply GNNs with a modified label propagation mechanism to utilize the edge coefficients effectively. Specifically, our supplementary module identifies a certain proportion of task-irrelevant edges based on a given confidence ratio. Using the remaining edges, we employ the widely used optimal transport to measure the similarity between two nodes with their subgraphs. Finally, using the coefficients as supplementary information on GNNs, we improve the label propagation mechanism which can prevent two nodes with smaller weights from being closer. The experiments on benchmark datasets show that our model alleviates over-smoothing and improves performance.

Recent advent in recommender systems, especially text-aided methods and CDR (Cross-Domain Recommendation) leads to promising results in solving data-sparsity and cold-start problems. Despite such progress, prior algorithms either require user overlapping or ignore domain-aware feature extraction. In addition, text-aided methods exceedingly emphasize aggregated documents and fail to capture the specifics embedded in individual reviews. To overcome such limitations, we propose a novel method, named DaRE (Domainaware Feature Extraction and Review Encoder), a comprehensive solution that consists of three key components; text-based representation learning, domain-aware feature extraction, and a review encoder. DaRE debilitate noises by separating domain-invariant features from domain-specific features through selective adversarial training. DaRE extracts features from aggregated documents, and the review encoder fine-tunes the representations by aligning them with the features extracted from individual reviews. Experiments on four real-world datasets show the superiority of DaRE over state-ofthe-art single-domain and cross-domain methodologies, achieving 9.2 % and 3.6 % improvements, respectively. We upload our implementations (https://anonymous.4open.science/r/DaRE-9CC9/) for a reproducibility