Attention mechanisms are central to the success of large language models (LLMs), enabling them to capture intricate token dependencies and implicitly assign importance to each token. Recent studies have revealed the sink token, which receives disproportionately high attention despite their limited semantic role. In this paper, we first expand the relationship between the sink token and other tokens, moving beyond attention to explore their similarity in hidden states, considering the layer depth. We observe that as the layers get deeper, the cosine similarity between the normalized hidden states of the sink token and those of other tokens increases, and that the normalized hidden states of the sink token exhibit negligible changes. These imply that other tokens consistently are directed toward the sink token throughout the layers. Next, we propose a dynamic token selection method, called OrthoRank, using these findings to select important tokens. Specifically, in a certain layer, we define token importance by the speed at which the token moves toward the sink token. This is converted into orthogonality with the sink token, meaning that tokens that are more orthogonal to the sink token are assigned greater importance. Finally, through extensive experiments, we demonstrated that our method results in lower perplexity and higher zero-shot accuracy compared to layer pruning methods at the same sparsity ratio with comparable throughput, while also achieving superior performance on LongBench.

Federated Learning (FL) aggregates locally trained models from individual clients to construct a global model. While FL enables learning a model with data privacy, it often suffers from significant performance degradation when clients have heterogeneous data distributions. This data heterogeneity causes the model to forget the global knowledge acquired from previously sampled clients after being trained on local datasets. Although the introduction of proximal objectives in local updates helps to preserve global knowledge, it can also hinder local learning by interfering with local objectives. To address this problem, we propose a novel method, Federated Stabilized Orthogonal Learning (FedSOL), which adopts an orthogonal learning strategy to balance the two conflicting objectives. FedSOL is designed to identify gradients of local objectives that are inherently orthogonal to directions affecting the proximal objective. Specifically, FedSOL targets parameter regions where learning on the local objective is minimally influenced by proximal weight perturbations. Our experiments demonstrate that FedSOL consistently achieves state-of-the-art performance across various scenarios.

Cross-domain few-shot learning (CD-FSL), where there are few target samples under extreme differences between source and target domains, has recently attracted huge attention. Recent studies on CD-FSL generally focus on transfer learning based approaches, where a neural network is pre-trained on popular labeled source domain datasets and then transferred to target domain data. Although the labeled datasets may provide suitable initial parameters for the target data, the domain difference between the source and target might hinder fine-tuning on the target domain. This paper proposes a simple yet powerful method that re-randomizes the parameters fitted on the source domain before adapting to the target data. The re-randomization resets source-specific parameters of the source pre-trained model and thus facilitates fine-tuning on the target domain, improving few-shot performance.

Translation has played a crucial role in improving the performance on multilingual tasks: (1) to generate the target language data from the source language data for training and (2) to generate the source language data from the target language data for inference.However, prior works have not considered the use of both translations simultaneously.This paper shows that combining them can synergize the results on various multilingual sentence classification tasks.We empirically find that translation artifacts stylized by translators are the main factor of the performance gain.Based on this analysis, we adopt two training methods, SupCon and MixUp, considering translation artifacts.Furthermore, we propose a crosslingual fine-tuning algorithm called MUSC, which uses SupCon and MixUp jointly and improves the performance.

Cross-domain few-shot learning (CD-FSL) has drawn increasing attention for handling large differences between the source and target domains--an important concern in real-world scenarios. To overcome these large differences, recent works have considered exploiting small-scale unlabeled data from the target domain during the pre-training stage. This data enables self-supervised pre-training on the target domain, in addition to supervised pre-training on the source domain. In this paper, we empirically investigate which pre-training is preferred based on domain similarity and few-shot difficulty of the target domain. We discover that the performance gain of self-supervised pre-training over supervised pre-training becomes large when the target domain is dissimilar to the source domain, or the target domain itself has low few-shot difficulty. We further design two pre-training schemes, mixed-supervised and two-stage learning, that improve performance. In this light, we present six findings for CD-FSL, which are supported by extensive experiments and analyses on three source and eight target benchmark datasets with varying levels of domain similarity and few-shot difficulty. Our code is available at https://github.com/sungnyun/understanding-cdfsl.

Knowledge distillation (KD), transferring knowledge from a cumbersome teacher model to a lightweight student model, has been investigated to design efficient neural architectures. Generally, the objective function of KD is the Kullback-Leibler (KL) divergence loss between the softened probability distributions of the teacher model and the student model with the temperature scaling hyperparameter τ. Despite its widespread use, few studies have discussed how such softening influences generalization. Here, we theoretically show that the KL divergence loss focuses on the logit matching when τ increases and the label matching when τ goes to 0 and empirically show that the logit matching is positively correlated to performance improvement in general. From this observation, we consider an intuitive KD loss function, the mean squared error (MSE) between the logit vectors, so that the student model can directly learn the logit of the teacher model. The MSE loss outperforms the KL divergence loss, explained by the penultimate layer representations difference between the two losses. Furthermore, we show that sequential distillation can improve performance and that KD, using the KL divergence loss with small τ particularly, mitigates the label noise. The code to reproduce the experiments is publicly available online at https://github.com/jhoon-oh/kd_data/.

Nowadays, online video platforms mostly recommend related videos by analyzing user-driven data such as viewing patterns, rather than the content of the videos. However, content is more important than any other element when videos aim to deliver knowledge. Therefore, we have developed a web application which recommends related TED lecture videos to the users, considering the content of the videos from the transcripts. TED Talk Recommender constructs a network for recommending videos that are similar content-wise and providing a user interface. Our demo system is available at http://dmserver6.kaist.ac.kr:24673/.