Kyungwoo Song

Changdae Oh, Hyeji Hwang, Hee-young Lee et al.

With the surge of large-scale pre-trained models (PTMs), fine-tuning these models to numerous downstream tasks becomes a crucial problem. Consequently, parameter efficient transfer learning (PETL) of large models has grasped huge attention. While recent PETL methods showcase impressive performance, they rely on optimistic assumptions: 1) the entire parameter set of a PTM is available, and 2) a sufficiently large memory capacity for the fine-tuning is equipped. However, in most real-world applications, PTMs are served as a black-box API or proprietary software without explicit parameter accessibility. Besides, it is hard to meet a large memory requirement for modern PTMs. In this work, we propose black-box visual prompting (BlackVIP), which efficiently adapts the PTMs without knowledge about model architectures and parameters. BlackVIP has two components; 1) Coordinator and 2) simultaneous perturbation stochastic approximation with gradient correction (SPSA-GC). The Coordinator designs input-dependent image-shaped visual prompts, which improves few-shot adaptation and robustness on distribution/location shift. SPSA-GC efficiently estimates the gradient of a target model to update Coordinator. Extensive experiments on 16 datasets demonstrate that BlackVIP enables robust adaptation to diverse domains without accessing PTMs' parameters, with minimal memory requirements. Code: \url{https://github.com/changdaeoh/BlackVIP}

Changdae Oh, Hyesu Lim, Mijoo Kim et al. · cmu, uw

Improving out-of-distribution (OOD) generalization during in-distribution (ID) adaptation is a primary goal of robust fine-tuning of zero-shot models beyond naive fine-tuning. However, despite decent OOD generalization performance from recent robust fine-tuning methods, confidence calibration for reliable model output has not been fully addressed. This work proposes a robust fine-tuning method that improves both OOD accuracy and confidence calibration simultaneously in vision language models. Firstly, we show that both OOD classification and OOD calibration errors have a shared upper bound consisting of two terms of ID data: 1) ID calibration error and 2) the smallest singular value of the ID input covariance matrix. Based on this insight, we design a novel framework that conducts fine-tuning with a constrained multimodal contrastive loss enforcing a larger smallest singular value, which is further guided by the self-distillation of a moving-averaged model to achieve calibrated prediction as well. Starting from empirical evidence supporting our theoretical statements, we provide extensive experimental results on ImageNet distribution shift benchmarks that demonstrate the effectiveness of our theorem and its practical implementation.

Changdae Oh, Junhyuk So, Hoyoon Byun et al.

Pre-trained multi-modal models, such as CLIP, provide transferable embeddings and show promising results in diverse applications. However, the analysis of learned multi-modal embeddings is relatively unexplored, and the embedding transferability can be improved. In this work, we observe that CLIP holds separated embedding subspaces for two different modalities, and then we investigate it through the lens of uniformity-alignment to measure the quality of learned representation. Both theoretically and empirically, we show that CLIP retains poor uniformity and alignment even after fine-tuning. Such a lack of alignment and uniformity might restrict the transferability and robustness of embeddings. To this end, we devise a new fine-tuning method for robust representation equipping better alignment and uniformity. First, we propose a Geodesic Multi-Modal Mixup that mixes the embeddings of image and text to generate hard negative samples on the hypersphere. Then, we fine-tune the model on hard negatives as well as original negatives and positives with contrastive loss. Based on the theoretical analysis about hardness guarantee and limiting behavior, we justify the use of our method. Extensive experiments on retrieval, calibration, few- or zero-shot classification (under distribution shift), embedding arithmetic, and image captioning further show that our method provides transferable representations, enabling robust model adaptation on diverse tasks. Code: https://github.com/changdaeoh/multimodal-mixup

HeeSun Bae, Seungjae Shin, Byeonghu Na et al.

Noisy labels are inevitable yet problematic in machine learning society. It ruins the generalization of a classifier by making the classifier over-fitted to noisy labels. Existing methods on noisy label have focused on modifying the classifier during the training procedure. It has two potential problems. First, these methods are not applicable to a pre-trained classifier without further access to training. Second, it is not easy to train a classifier and regularize all negative effects from noisy labels, simultaneously. We suggest a new branch of method, Noisy Prediction Calibration (NPC) in learning with noisy labels. Through the introduction and estimation of a new type of transition matrix via generative model, NPC corrects the noisy prediction from the pre-trained classifier to the true label as a post-processing scheme. We prove that NPC theoretically aligns with the transition matrix based methods. Yet, NPC empirically provides more accurate pathway to estimate true label, even without involvement in classifier learning. Also, NPC is applicable to any classifier trained with noisy label methods, if training instances and its predictions are available. Our method, NPC, boosts the classification performances of all baseline models on both synthetic and real-world datasets. The implemented code is available at https://github.com/BaeHeeSun/NPC.

Changdae Oh, Gyeongdeok Seo, Geunyoung Jung et al. · cmu, uw

With a surge of large-scale pre-trained models, parameter-efficient transfer learning (PETL) of large models has garnered significant attention. While promising, they commonly rely on two optimistic assumptions: 1) full access to the parameters of a PTM, and 2) sufficient memory capacity to cache all intermediate activations for gradient computation. However, in most real-world applications, PTMs serve as black-box APIs or proprietary software without full parameter accessibility. Besides, it is hard to meet a large memory requirement for modern PTMs. This work proposes black-box visual prompting (BlackVIP), which efficiently adapts the PTMs without knowledge of their architectures or parameters. BlackVIP has two components: 1) Coordinator and 2) simultaneous perturbation stochastic approximation with gradient correction (SPSA-GC). The Coordinator designs input-dependent visual prompts, which allow the target PTM to adapt in the wild. SPSA-GC efficiently estimates the gradient of PTM to update Coordinator. Besides, we introduce a variant, BlackVIP-SE, which significantly reduces the runtime and computational cost of BlackVIP. Extensive experiments on 19 datasets demonstrate that BlackVIPs enable robust adaptation to diverse domains and tasks with minimal memory requirements. We further provide a theoretical analysis on the generalization of visual prompting methods by presenting their connection to the certified robustness of randomized smoothing, and presenting an empirical support for improved robustness.

JoonHo Jang, Byeonghu Na, DongHyeok Shin et al.

Open-Set Domain Adaptation (OSDA) assumes that a target domain contains unknown classes, which are not discovered in a source domain. Existing domain adversarial learning methods are not suitable for OSDA because distribution matching with $\textit{unknown}$ classes leads to negative transfer. Previous OSDA methods have focused on matching the source and the target distribution by only utilizing $\textit{known}$ classes. However, this $\textit{known}$-only matching may fail to learn the target-$\textit{unknown}$ feature space. Therefore, we propose Unknown-Aware Domain Adversarial Learning (UADAL), which $\textit{aligns}$ the source and the target-$\textit{known}$ distribution while simultaneously $\textit{segregating}$ the target-$\textit{unknown}$ distribution in the feature alignment procedure. We provide theoretical analyses on the optimized state of the proposed $\textit{unknown-aware}$ feature alignment, so we can guarantee both $\textit{alignment}$ and $\textit{segregation}$ theoretically. Empirically, we evaluate UADAL on the benchmark datasets, which shows that UADAL outperforms other methods with better feature alignments by reporting state-of-the-art performances.

Hoyoon Byun, Youngjun Choi, Taero Kim et al.

Pre-Layer Normalization (Pre-LN) is the de facto choice for large language models (LLMs) and is crucial for stable pretraining and effective transfer learning. However, Pre-LN is inefficient due to repeated statistical calculations and suffers from the curse of depth. As layers grow, the magnitude and variance of the hidden state escalate, destabilizing training. Efficiency-oriented normalization-free methods such as Dynamic Tanh (DyT) improve speed but remain fragile at depth. To jointly address stability and efficiency, we propose Bounded Hyperbolic Tanh (BHyT), a drop-in replacement for Pre-LN. BHyT couples a tanh nonlinearity with explicit, data-driven input bounding to keep activations within a non-saturating range. It prevents depth-wise growth in activation magnitude and variance and comes with a theoretical stability guarantee. For efficiency, BHyT computes exact statistics once per block and replaces a second normalization with a lightweight variance approximation, enhancing efficiency. Empirically, BHyT demonstrates improved stability and efficiency during pretraining, achieving an average of 15.8% faster training and an average of 4.2% higher token generation throughput compared to RMSNorm., while matching or surpassing its inference performance and robustness across language understanding and reasoning benchmarks. Our code is available at: https://anonymous.4open.science/r/BHyT

Changdae Oh, Heeji Won, Junhyuk So et al.

Learning fair representation is crucial for achieving fairness or debiasing sensitive information. Most existing works rely on adversarial representation learning to inject some invariance into representation. However, adversarial learning methods are known to suffer from relatively unstable training, and this might harm the balance between fairness and predictiveness of representation. We propose a new approach, learning FAir Representation via distributional CONtrastive Variational AutoEncoder (FarconVAE), which induces the latent space to be disentangled into sensitive and nonsensitive parts. We first construct the pair of observations with different sensitive attributes but with the same labels. Then, FarconVAE enforces each non-sensitive latent to be closer, while sensitive latents to be far from each other and also far from the non-sensitive latent by contrasting their distributions. We provide a new type of contrastive loss motivated by Gaussian and Student-t kernels for distributional contrastive learning with theoretical analysis. Besides, we adopt a new swap-reconstruction loss to boost the disentanglement further. FarconVAE shows superior performance on fairness, pretrained model debiasing, and domain generalization tasks from various modalities, including tabular, image, and text.

Geunyoung Jung, Soohong Kim, Kyungwoo Song et al.

With the rise of pre-trained models in the 3D point cloud domain for a wide range of real-world applications, adapting them to downstream tasks has become increasingly important. However, conventional full fine-tuning methods are computationally expensive and storage-intensive. Although prompt tuning has emerged as an efficient alternative, it often suffers from overfitting, thereby compromising generalization capability. To address this issue, we propose Prototypical Point-level Prompt Tuning (P$^3$T), a parameter-efficient prompt tuning method designed for pre-trained 3D vision-language models (VLMs). P$^3$T consists of two components: 1) \textit{Point Prompter}, which generates instance-aware point-level prompts for the input point cloud, and 2) \textit{Text Prompter}, which employs learnable prompts into the input text instead of hand-crafted ones. Since both prompters operate directly on input data, P$^3$T enables task-specific adaptation of 3D VLMs without sacrificing generalizability. Furthermore, to enhance embedding space alignment, which is key to fine-tuning 3D VLMs, we introduce a prototypical loss that reduces intra-category variance. Extensive experiments demonstrate that our method matches or outperforms full fine-tuning in classification and few-shot learning, and further exhibits robust generalization under data shift in the cross-dataset setting. The code is available at \textcolor{violet}{https://github.com/gyjung975/P3T}.

Seunghwan An, Kyungwoo Song, Jong-June Jeon

We present a new supervised learning technique for the Variational AutoEncoder (VAE) that allows it to learn a causally disentangled representation and generate causally disentangled outcomes simultaneously. We call this approach Causally Disentangled Generation (CDG). CDG is a generative model that accurately decodes an output based on a causally disentangled representation. Our research demonstrates that adding supervised regularization to the encoder alone is insufficient for achieving a generative model with CDG, even for a simple task. Therefore, we explore the necessary and sufficient conditions for achieving CDG within a specific model. Additionally, we introduce a universal metric for evaluating the causal disentanglement of a generative model. Empirical results from both image and tabular datasets support our findings.

Taero Kim, Subeen Park, Sungjun Lim et al.

Learning robust models under distribution shifts between training and test datasets is a fundamental challenge in machine learning. While learning invariant features across environments is a popular approach, it often assumes that these features are fully observed in both training and test sets, a condition frequently violated in practice. When models rely on invariant features absent in the test set, their robustness in new environments can deteriorate. To tackle this problem, we introduce a novel learning principle called the Sufficient Invariant Learning (SIL) framework, which focuses on learning a sufficient subset of invariant features rather than relying on a single feature. After demonstrating the limitation of existing invariant learning methods, we propose a new algorithm, Adaptive Sharpness-aware Group Distributionally Robust Optimization (ASGDRO), to learn diverse invariant features by seeking common flat minima across the environments. We theoretically demonstrate that finding a common flat minima enables robust predictions based on diverse invariant features. Empirical evaluations on multiple datasets, including our new benchmark, confirm ASGDRO's robustness against distribution shifts, highlighting the limitations of existing methods.

Seyun Bae, Hoyoon Byun, Changdae Oh et al.

Multimodal machine learning has been widely studied for the development of general intelligence. Recently, the Perceiver and Perceiver IO, show competitive results for diverse dataset domains and tasks. However, recent works, Perceiver and Perceiver IO, have focused on heterogeneous modalities, including image, text, and there are few research works for graph structured datasets. A graph has an adjacency matrix different from other datasets such as text and image, and it is not trivial to handle the topological information. In this study, we provide a Graph Perceiver IO (GPIO), the Perceiver IO for the graph structured dataset. We keep the main structure of the GPIO as the Perceiver IO because the Perceiver IO already handles the diverse dataset well, except for the graph structured dataset. The GPIO is a general method that handles diverse datasets, such as graph-structured data, text, and images, by leveraging positional encoding and output query smoothing. Compared to graph neural networks (GNNs), GPIO requires lower complexity and can efficiently incorporate global and local information, which is also empirically validated through experiments. Furthermore, we propose GPIO+ for the multimodal few-shot classification that incorporates both images and graphs simultaneously. GPIO achieves higher benchmark accuracy than GNNs across multiple tasks, including graph classification, node classification, and multimodal text classification, while also attaining superior AP and AUC in link prediction. Additionally, GPIO+ outperforms GNNs in multimodal few-shot classification. Our GPIO(+) can serve as a general architecture for handling various modalities and tasks.

Hyeondey Kim, Jinwoo Nam, Minjae Lee et al.

Knowledge tracing plays a pivotal role in intelligent tutoring systems. This task aims to predict the probability of students answering correctly to specific questions. To do so, knowledge tracing systems should trace the knowledge state of the students by utilizing their problem-solving history and knowledge about the problems. Recent advances in knowledge tracing models have enabled better exploitation of problem solving history. However, knowledge about problems has not been studied, as well compared to students' answering histories. Knowledge tracing algorithms that incorporate knowledge directly are important to settings with limited data or cold starts. Therefore, we consider the problem of utilizing skill-to-skill relation to knowledge tracing. In this work, we introduce expert labeled skill-to-skill relationships. Moreover, we also provide novel methods to construct a knowledge-tracing model to leverage human experts' insight regarding relationships between skills. The results of an extensive experimental analysis show that our method outperformed a baseline Transformer model. Furthermore, we found that the extent of our model's superiority was greater in situations with limited data, which allows a smooth cold start of our model.

Keon-Hee Park, Hakyung Lee, Kyungwoo Song et al.

With the advancement of deep neural networks in computer vision, artificial intelligence (AI) is widely employed in real-world applications. However, AI still faces limitations in mimicking high-level human capabilities, such as novel category discovery, for practical use. While some methods utilizing offline continual learning have been proposed for novel category discovery, they neglect the continuity of data streams in real-world settings. In this work, we introduce Online Continuous Generalized Category Discovery (OCGCD), which considers the dynamic nature of data streams where data can be created and deleted in real time. Additionally, we propose a novel method, DEAN, Discovery via Energy guidance and feature AugmentatioN, which can discover novel categories in an online manner through energy-guided discovery and facilitate discriminative learning via energy-based contrastive loss. Furthermore, DEAN effectively pseudo-labels unlabeled data through variance-based feature augmentation. Experimental results demonstrate that our proposed DEAN achieves outstanding performance in proposed OCGCD scenario.

Keon-Hee Park, Kyungwoo Song, Gyeong-Moon Park

Few-Shot Class Incremental Learning (FSCIL) is a task that requires a model to learn new classes incrementally without forgetting when only a few samples for each class are given. FSCIL encounters two significant challenges: catastrophic forgetting and overfitting, and these challenges have driven prior studies to primarily rely on shallow models, such as ResNet-18. Even though their limited capacity can mitigate both forgetting and overfitting issues, it leads to inadequate knowledge transfer during few-shot incremental sessions. In this paper, we argue that large models such as vision and language transformers pre-trained on large datasets can be excellent few-shot incremental learners. To this end, we propose a novel FSCIL framework called PriViLege, Pre-trained Vision and Language transformers with prompting functions and knowledge distillation. Our framework effectively addresses the challenges of catastrophic forgetting and overfitting in large models through new pre-trained knowledge tuning (PKT) and two losses: entropy-based divergence loss and semantic knowledge distillation loss. Experimental results show that the proposed PriViLege significantly outperforms the existing state-of-the-art methods with a large margin, e.g., +9.38% in CUB200, +20.58% in CIFAR-100, and +13.36% in miniImageNet. Our implementation code is available at https://github.com/KHU-AGI/PriViLege.

Sungjun Lim, Heedong Kim, Andrew Lee et al.

Mechanistic interpretability aims to explain a model's behavior by identifying causally responsible internal structures. Dictionary-based explainers such as sparse autoencoders and transcoders are a primary tool, but their faithfulness under out-of-distribution (OOD) shift has received little systematic attention. We show that distribution shift rotates the subspace that the model actively uses, misaligning the explainer's dictionary trained on in-distribution (ID) activations. We formalize this misalignment as the faithfulness gap, a geometric distance between the ID dictionary and the OOD-active subspace, and show that it controls OOD faithfulness degradation. To reduce this gap, we propose the Geometry-Adaptive Explainer (GAE), which realigns the explainer's dictionary with the OOD-active subspace while preserving the original feature structure. This requires only unlabeled OOD activations and no gradient updates. We prove that GAE improves over the unadapted ID explainer, with excess loss bounded quadratically by the second-moment shift. Empirically, GAE even matches or surpasses all training-based baselines in causal faithfulness across multiple models and OOD settings.

Kangjun Noh, Baekryun Seong, Hoyoon Byun et al.

Large Language Models (LLMs) have great success in natural language processing tasks such as response generation. However, their use in tabular data has been limited due to their inferior performance compared to traditional machine learning models (TMLs) such as XGBoost. We find that the pre-trained knowledge of LLMs enables them to interpret new variables that appear in a test without additional training, a capability central to the concept of Out-of-Variable (OOV). From the findings, we propose a Language-Based-Classifier (LBC), a classifier that maximizes the benefits of LLMs to outperform TMLs on OOV tasks. LBC employs three key methodological strategies: 1) Categorical changes to adjust data to better fit the model's understanding, 2) Advanced order and indicator to enhance data representation to the model, and 3) Using verbalizer to map logit scores to classes during inference to generate model predictions. These strategies, combined with the pre-trained knowledge of LBC, emphasize the model's ability to effectively handle OOV tasks. We empirically and theoretically validate the superiority of LBC. LBC is the first study to apply an LLM-based model to OOV tasks. The source code is at https://github.com/sksmssh/LBCforOOVGen

YongTaek Lim, Suho Kang, Yewon Kim et al.

Large-scale pre-trained models such as CLIP excel in transferability and robust generalization across diverse datasets. However, adapting these models to new datasets or domains is computationally costly, especially in low-resource or few-shot settings, and existing prompt-learning methods often lack interpretability. We introduce Efficient Prompt Learning with Knowledge Graph (EPLKG), which uses a knowledge graph to curate diverse, interpretable prompts and, where KG coverage is limited, augments this bank with LLM-generated human-readable visual descriptions. EPLKG operates entirely on cached CLIP image and text embeddings and employs a lightweight Gumbel-Softmax module to select a single prompt per image-class pair, enabling low-memory, fast training. Across 11 benchmarks, EPLKG reduces per-image training time by up to 45 percent and peak GPU memory by around 30 to 40 percent compared to strong prompt-learning baselines, while keeping the average base-new harmonic-mean accuracy within 2 percentage points, thereby improving the efficiency of model adaptation without sacrificing competitive performance or interpretability.

Subeen Park, Joowang Kim, Hakyung Lee et al.

Deep learning models achieve strong performance across various domains but often rely on spurious correlations, making them vulnerable to distribution shifts. This issue is particularly severe in subpopulation shift scenarios, where models struggle in underrepresented groups. While existing methods have made progress in mitigating this issue, their performance gains are still constrained. They lack a rigorous theoretical framework connecting the embedding space representations with worst-group error. To address this limitation, we propose Spurious Correlation-Aware Embedding Regularization for Worst-Group Robustness (SCER), a novel approach that directly regularizes feature representations to suppress spurious cues. We show theoretically that worst-group error is influenced by how strongly the classifier relies on spurious versus core directions, identified from differences in group-wise mean embeddings across domains and classes. By imposing theoretical constraints at the embedding level, SCER encourages models to focus on core features while reducing sensitivity to spurious patterns. Through systematic evaluation on multiple vision and language, we show that SCER outperforms prior state-of-the-art studies in worst-group accuracy. Our code is available at \href{https://github.com/MLAI-Yonsei/SCER}{https://github.com/MLAI-Yonsei/SCER}.

Heeyoung Lee, Hoyoon Byun, Changdae Oh et al.

Accessing machine learning models through remote APIs has been gaining prevalence following the recent trend of scaling up model parameters for increased performance. Even though these models exhibit remarkable ability, detecting out-of-distribution (OOD) samples remains a crucial safety concern for end users as these samples may induce unreliable outputs from the model. In this work, we propose an OOD detection framework, MixDiff, that is applicable even when the model's parameters or its activations are not accessible to the end user. To bypass the access restriction, MixDiff applies an identical input-level perturbation to a given target sample and a similar in-distribution (ID) sample, then compares the relative difference in the model outputs of these two samples. MixDiff is model-agnostic and compatible with existing output-based OOD detection methods. We provide theoretical analysis to illustrate MixDiff's effectiveness in discerning OOD samples that induce overconfident outputs from the model and empirically demonstrate that MixDiff consistently enhances the OOD detection performance on various datasets in vision and text domains.

Kangjun Noh, Seongchan Lee, Ilmun Kim et al.

Ensuring factuality is essential for the safe use of Large Language Models (LLMs) in high-stakes domains such as medicine and law. Conformal inference provides distribution-free guarantees, but existing approaches are either overly conservative, discarding many true-claims, or rely on adaptive error rates and simple linear models that fail to capture complex group structures. To address these challenges, we reformulate conformal inference in a multiplicative filtering setting, modeling factuality as a product of claim-level scores. Our method, Multi-LLM Adaptive Conformal Inference (MACI), leverages ensembles to produce more accurate factuality-scores, which in our experiments led to higher retention, while validity is preserved through group-conditional calibration. Experiments show that MACI consistently achieves user-specified coverage with substantially higher retention and lower time cost than baselines. Our repository is available at https://github.com/MLAI-Yonsei/MACI

Suho Kang, Jungyang Park, Joonseo Ha et al.

Foundation models (FMs) have achieved significant success across various tasks, leading to research on benchmarks for reasoning abilities. However, there is a lack of studies on FMs performance in exceptional scenarios, which we define as out-of-distribution (OOD) reasoning tasks. This paper is the first to address these cases, developing a novel dataset for evaluation of FMs across multiple modalities, including graphic novels, calligraphy, news articles, and lyrics. It includes tasks for instance classification, character recognition, token prediction, and text generation. The paper also proposes prompt engineering techniques like Chain-of-Thought (CoT) and CoT+Few-Shot to enhance performance. Validation of FMs using various methods revealed improvements. The code repository is accessible at: https://github.com/MLAI-Yonsei/ExceptionalBenchmark

Hankyeol Lee, Gawon Seo, Kyounggyu Lee et al.

Vision-language models (VLMs) often struggle to generate accurate and detailed captions for high-resolution images since they are typically pre-trained on low-resolution inputs (e.g., 224x224 or 336x336 pixels). Downscaling high-resolution images to these dimensions may result in the loss of visual details and the omission of important objects. To address this limitation, we propose a novel pipeline that integrates vision-language models, large language models (LLMs), and object detection systems to enhance caption quality. Our proposed pipeline refines captions through a novel, multi-stage process. Given a high-resolution image, an initial caption is first generated using a VLM, and key objects in the image are then identified by an LLM. The LLM predicts additional objects likely to co-occur with the identified key objects, and these predictions are verified by object detection systems. Newly detected objects not mentioned in the initial caption undergo focused, region-specific captioning to ensure they are incorporated. This process enriches caption detail while reducing hallucinations by removing references to undetected objects. We evaluate the enhanced captions using pairwise comparison and quantitative scoring from large multimodal models, along with a benchmark for hallucination detection. Experiments on a curated dataset of high-resolution images demonstrate that our pipeline produces more detailed and reliable image captions while effectively minimizing hallucinations.

Taero Kim, Hoyoon Byun, Youngjun Choi et al.

Scaling large language models (LLMs) demands approaches that increase capacity without incurring excessive parameter growth or inference cost. Depth Up-Scaling (DUS) has emerged as a promising strategy by duplicating layers and applying Continual Pre-training (CPT), but its reliance on feed-forward networks (FFNs) limits efficiency and attainable gains. We introduce Memory-Infused Depth Up-Scaling (MIDUS), which replaces FFNs in duplicated blocks with a head-wise memory (HML) layer. Motivated by observations that attention heads have distinct roles both across and within layers, MIDUS assigns an independent memory bank to each head, enabling head-wise retrieval and injecting information into subsequent layers while preserving head-wise functional structure. This design combines sparse memory access with head-wise representations and incorporates an efficient per-head value factorization module, thereby relaxing the usual efficiency-performance trade-off. Across our CPT experiments, MIDUS exhibits robust performance improvements over strong DUS baselines while maintaining a highly efficient parameter footprint. Our findings establish MIDUS as a compelling and resource-efficient alternative to conventional FFN replication for depth up-scaling by leveraging its head-wise memory design.

Hankyeol Lee, Gawon Seo, Wonseok Choi et al.

The performance of vision-language models (VLMs), such as CLIP, in visual classification tasks, has been enhanced by leveraging semantic knowledge from large language models (LLMs), including GPT. Recent studies have shown that in zero-shot classification tasks, descriptors incorporating additional cues, high-level concepts, or even random characters often outperform those using only the category name. In many classification tasks, while the top-1 accuracy may be relatively low, the top-5 accuracy is often significantly higher. This gap implies that most misclassifications occur among a few similar classes, highlighting the model's difficulty in distinguishing between classes with subtle differences. To address this challenge, we introduce a novel concept of comparative descriptors. These descriptors emphasize the unique features of a target class against its most similar classes, enhancing differentiation. By generating and integrating these comparative descriptors into the classification framework, we refine the semantic focus and improve classification accuracy. An additional filtering process ensures that these descriptors are closer to the image embeddings in the CLIP space, further enhancing performance. Our approach demonstrates improved accuracy and robustness in visual classification tasks by addressing the specific challenge of subtle inter-class differences.

Haeji Jung, Changdae Oh, Jooeon Kang et al.

Approaches to improving multilingual language understanding often struggle with significant performance gaps between high-resource and low-resource languages. While there are efforts to align the languages in a single latent space to mitigate such gaps, how different input-level representations influence such gaps has not been investigated, particularly with phonemic inputs. We hypothesize that the performance gaps are affected by representation discrepancies between these languages, and revisit the use of phonemic representations as a means to mitigate these discrepancies. To demonstrate the effectiveness of phonemic representations, we present experiments on three representative cross-lingual tasks on 12 languages in total. The results show that phonemic representations exhibit higher similarities between languages compared to orthographic representations, and it consistently outperforms grapheme-based baseline model on languages that are relatively low-resourced. We present quantitative evidence from three cross-lingual tasks that demonstrate the effectiveness of phonemic representations, and it is further justified by a theoretical analysis of the cross-lingual performance gap.

Seonggyun Lee, Sungjun Lim, Seojin Park et al.

The field of preference optimization has made outstanding contributions to the alignment of language models with human preferences. Despite these advancements, recent methods still rely heavily on substantial paired (labeled) feedback data, leading to substantial resource expenditures. To address these challenges, we study the problem of Semi-Supervised Preference Optimization (SSPO) in which the idea is to learn from both a small number of pairwise preference labels and a large pool of unpaired samples simultaneously. Our key theoretical contribution proves the existence of an optimal reward threshold capable of separating winning and losing responses with high probability, which enables a principled pseudo-labeling of unpaired data. By leveraging these pseudo-labels, SSPO effectively distills latent preferences from large-scale unpaired data, thus maintaining human alignment while drastically reducing acquisition costs. Extensive experiments across datasets validate this remarkable data efficiency; for instance, SSPO trained with Llama3-8B-Instruct on just 1% of UltraFeedback consistently surpasses strong baselines trained on 10% of UltraFeedback.

Youngjun Choi, Joonseong Kang, Sungjun Lim et al.

Data valuation has become central in the era of data-centric AI. It drives efficient training pipelines and enables objective pricing in data markets by assigning a numeric value to each data point. Most existing data valuation methods estimate the effect of removing individual data points by evaluating changes in model validation performance under in-distribution (ID) settings, as opposed to out-of-distribution (OOD) scenarios where data follow different patterns. Since ID and OOD data behave differently, data valuation methods based on ID loss often fail to generalize to OOD settings, particularly when the validation set contains no OOD data. Furthermore, although OOD-aware methods exist, they involve heavy computational costs, which hinder practical deployment. To address these challenges, we introduce \emph{Eigen-Value} (EV), a plug-and-play data valuation framework for OOD robustness that uses only an ID data subset, including during validation. EV provides a new spectral approximation of domain discrepancy, which is the gap of loss between ID and OOD using ratios of eigenvalues of ID data's covariance matrix. EV then estimates the marginal contribution of each data point to this discrepancy via perturbation theory, alleviating the computational burden. Subsequently, EV plugs into ID loss-based methods by adding an EV term without any additional training loop. We demonstrate that EV achieves improved OOD robustness and stable value rankings across real-world datasets, while remaining computationally lightweight. These results indicate that EV is practical for large-scale settings with domain shift, offering an efficient path to OOD-robust data valuation.

Sungjun Lim, Kangjun Noh, Youngjun Choi et al.

Text embeddings are essential components in modern NLP pipelines. While numerous embedding models have been proposed, their performance varies across domains. This variability motivates the use of ensemble techniques to combine complementary strengths. However, most existing ensemble methods operate on deterministic embeddings and fail to account for model-specific uncertainty, limiting their robustness and reliability in downstream applications. To address these limitations, we propose Uncertainty-driven Embedding Convolution (UEC). UEC first transforms deterministic embeddings into probabilistic ones in a post-hoc manner. It then computes adaptive ensemble weights based on embedding uncertainty, grounded in a Bayes-optimal solution under a surrogate loss. Additionally, UEC employs an uncertainty-aware similarity function that directly incorporates uncertainty into the similarity scoring, providing a theoretically grounded and efficient surrogate to distributional distances. Extensive experiments on diverse benchmarks demonstrate that UEC consistently improves both performance and robustness by leveraging principled uncertainty modeling.

Joonseong Kang, Soojeong Lee, Subeen Park et al.

In-Context Learning (ICL) enables Large Language Models (LLMs) to perform tasks without parameter updates by conditioning on a few demonstrations provided in the prompt. Despite its success, ICL suffers from several limitations, including sensitivity to demonstration order, context length constraints, and computational inefficiency. To address these challenges, task vector-based approaches compress task information into a single vector. However, these methods typically construct task vectors from fixed sets of demonstrations and reuse them across input queries, without conditioning on the specific input. This limitation can lead models to struggle with effective adaptation when the input query is not well aligned with the underlying demonstrations, consequently degrading their generalization performance on unseen tasks. To overcome this limitation, we propose Adaptive Task Vectors (ATV), a simple and effective framework that dynamically generates task vectors conditioned on each input query. ATV employs a small language model to generate task vectors, which are then transformed to match the target LLM's architecture and applied to guide its output generation. In contrast to ICL and previous vector-based approaches, which rely on fixed demonstration sets and their corresponding vectors, ATV dynamically generates task vectors tailored to each specific input query and task. Consequently, ATV demonstrates strong performance and generalization capabilities, even for unseen tasks. Furthermore, we provide a theoretical analysis indicating that ATV is expressively equivalent to LoRA under equal rank budgets and more expressive than Prefix-Tuning, thereby offering formal support for its representational advantage.

Sungjun Lim, Jeyoon Yeom, Sooyon Kim et al.

Bayesian neural networks (BNNs) estimate the posterior distribution of model parameters and utilize posterior samples for Bayesian Model Averaging (BMA) in prediction. However, despite the crucial role of flatness in the loss landscape in improving the generalization of neural networks, its impact on BMA has been largely overlooked. In this work, we explore how posterior flatness influences BMA generalization and empirically demonstrate that (1) most approximate Bayesian inference methods fail to yield a flat posterior and (2) BMA predictions, without considering posterior flatness, are less effective at improving generalization. To address this, we propose Flat Posterior-aware Bayesian Model Averaging (FP-BMA), a novel training objective that explicitly encourages flat posteriors in a principled Bayesian manner. We also introduce a Flat Posterior-aware Bayesian Transfer Learning scheme that enhances generalization in downstream tasks. Empirically, we show that FP-BMA successfully captures flat posteriors, improving generalization performance.

Dongjun Kim, Seungjae Shin, Kyungwoo Song et al.

Recent advances in diffusion models bring state-of-the-art performance on image generation tasks. However, empirical results from previous research in diffusion models imply an inverse correlation between density estimation and sample generation performances. This paper investigates with sufficient empirical evidence that such inverse correlation happens because density estimation is significantly contributed by small diffusion time, whereas sample generation mainly depends on large diffusion time. However, training a score network well across the entire diffusion time is demanding because the loss scale is significantly imbalanced at each diffusion time. For successful training, therefore, we introduce Soft Truncation, a universally applicable training technique for diffusion models, that softens the fixed and static truncation hyperparameter into a random variable. In experiments, Soft Truncation achieves state-of-the-art performance on CIFAR-10, CelebA, CelebA-HQ 256x256, and STL-10 datasets.

Dongjun Kim, Kyungwoo Song, Seungjae Shin et al.

A simulation is useful when the phenomenon of interest is either expensive to regenerate or irreproducible with the same context. Recently, Bayesian inference on the distribution of the simulation input parameter has been implemented sequentially to minimize the required simulation budget for the task of simulation validation to the real-world. However, the Bayesian inference is still challenging when the ground-truth posterior is multi-modal with a high-dimensional simulation output. This paper introduces a regularization technique, namely Neural Posterior Regularization (NPR), which enforces the model to explore the input parameter space effectively. Afterward, we provide the closed-form solution of the regularized optimization that enables analyzing the effect of the regularization. We empirically validate that NPR attains the statistically significant gain on benchmark performances for diverse simulation tasks.

Hyemi Kim, Seungjae Shin, JoonHo Jang et al.

The problem of fair classification can be mollified if we develop a method to remove the embedded sensitive information from the classification features. This line of separating the sensitive information is developed through the causal inference, and the causal inference enables the counterfactual generations to contrast the what-if case of the opposite sensitive attribute. Along with this separation with the causality, a frequent assumption in the deep latent causal model defines a single latent variable to absorb the entire exogenous uncertainty of the causal graph. However, we claim that such structure cannot distinguish the 1) information caused by the intervention (i.e., sensitive variable) and 2) information correlated with the intervention from the data. Therefore, this paper proposes Disentangled Causal Effect Variational Autoencoder (DCEVAE) to resolve this limitation by disentangling the exogenous uncertainty into two latent variables: either 1) independent to interventions or 2) correlated to interventions without causality. Particularly, our disentangling approach preserves the latent variable correlated to interventions in generating counterfactual examples. We show that our method estimates the total effect and the counterfactual effect without a complete causal graph. By adding a fairness regularization, DCEVAE generates a counterfactual fair dataset while losing less original information. Also, DCEVAE generates natural counterfactual images by only flipping sensitive information. Additionally, we theoretically show the differences in the covariance structures of DCEVAE and prior works from the perspective of the latent disentanglement.

Yoon-Yeong Kim, Kyungwoo Song, JoonHo Jang et al.

Active learning effectively collects data instances for training deep learning models when the labeled dataset is limited and the annotation cost is high. Besides active learning, data augmentation is also an effective technique to enlarge the limited amount of labeled instances. However, the potential gain from virtual instances generated by data augmentation has not been considered in the acquisition process of active learning yet. Looking ahead the effect of data augmentation in the process of acquisition would select and generate the data instances that are informative for training the model. Hence, this paper proposes Look-Ahead Data Acquisition via augmentation, or LADA, to integrate data acquisition and data augmentation. LADA considers both 1) unlabeled data instance to be selected and 2) virtual data instance to be generated by data augmentation, in advance of the acquisition process. Moreover, to enhance the informativeness of the virtual data instances, LADA optimizes the data augmentation policy to maximize the predictive acquisition score, resulting in the proposal of InfoMixup and InfoSTN. As LADA is a generalizable framework, we experiment with the various combinations of acquisition and augmentation methods. The performance of LADA shows a significant improvement over the recent augmentation and acquisition baselines which were independently applied to the benchmark datasets.

Dongjun Kim, Kyungwoo Song, YoonYeong Kim et al.

Bayesian inference without the likelihood evaluation, or likelihood-free inference, has been a key research topic in simulation studies for gaining quantitatively validated simulation models on real-world datasets. As the likelihood evaluation is inaccessible, previous papers train the amortized neural network to estimate the ground-truth posterior for the simulation of interest. Training the network and accumulating the dataset alternatively in a sequential manner could save the total simulation budget by orders of magnitude. In the data accumulation phase, the new simulation inputs are chosen within a portion of the total simulation budget to accumulate upon the collected dataset. This newly accumulated data degenerates because the set of simulation inputs is hardly mixed, and this degenerated data collection process ruins the posterior inference. This paper introduces a new sampling approach, called Neural Proposal (NP), of the simulation input that resolves the biased data collection as it guarantees the i.i.d. sampling. The experiments show the improved performance of our sampler, especially for the simulations with multi-modal posteriors.

Yohan Jung, Kyungwoo Song, Jinkyoo Park

A spectral mixture (SM) kernel is a flexible kernel used to model any stationary covariance function. Although it is useful in modeling data, the learning of the SM kernel is generally difficult because optimizing a large number of parameters for the SM kernel typically induces an over-fitting, particularly when a gradient-based optimization is used. Also, a longer training time is required. To improve the training, we propose an approximate Bayesian inference for the SM kernel. Specifically, we employ the variational distribution of the spectral points to approximate SM kernel with a random Fourier feature. We optimize the variational parameters by applying a sampling-based variational inference to the derived evidence lower bound (ELBO) estimator constructed from the approximate kernel. To improve the inference, we further propose two additional strategies: (1) a sampling strategy of spectral points to estimate the ELBO estimator reliably and thus its associated gradient, and (2) an approximate natural gradient to accelerate the convergence of the parameters. The proposed inference combined with two strategies accelerates the convergence of the parameters and leads to better optimal parameters.

Kyungwoo Song, Yohan Jung, Dongjun Kim et al.

\textit{Attention} computes the dependency between representations, and it encourages the model to focus on the important selective features. Attention-based models, such as Transformer and graph attention network (GAT), are widely utilized for sequential data and graph-structured data. This paper suggests a new interpretation and generalized structure of the attention in Transformer and GAT. For the attention in Transformer and GAT, we derive that the attention is a product of two parts: 1) the RBF kernel to measure the similarity of two instances and 2) the exponential of $L^{2}$ norm to compute the importance of individual instances. From this decomposition, we generalize the attention in three ways. First, we propose implicit kernel attention with an implicit kernel function instead of manual kernel selection. Second, we generalize $L^{2}$ norm as the $L^{p}$ norm. Third, we extend our attention to structured multi-head attention. Our generalized attention shows better performance on classification, translation, and regression tasks.

Dongjun Kim, Weonyoung Joo, Seungjae Shin et al.

Generative Adversarial Network (GAN) can be viewed as an implicit estimator of a data distribution, and this perspective motivates using the adversarial concept in the true input parameter estimation of black-box generators. While previous works on likelihood-free inference introduces an implicit proposal distribution on the generator input, this paper analyzes theoretic limitations of the proposal distribution approach. On top of that, we introduce a new algorithm, Adversarial Likelihood-Free Inference (ALFI), to mitigate the analyzed limitations, so ALFI is able to find the posterior distribution on the input parameter for black-box generative models. We experimented ALFI with diverse simulation models as well as pre-trained statistical models, and we identified that ALFI achieves the best parameter estimation accuracy with a limited simulation budget.

Seungjae Shin, Kyungwoo Song, JoonHo Jang et al.

Recent research demonstrates that word embeddings, trained on the human-generated corpus, have strong gender biases in embedding spaces, and these biases can result in the discriminative results from the various downstream tasks. Whereas the previous methods project word embeddings into a linear subspace for debiasing, we introduce a \textit{Latent Disentanglement} method with a siamese auto-encoder structure with an adapted gradient reversal layer. Our structure enables the separation of the semantic latent information and gender latent information of given word into the disjoint latent dimensions. Afterwards, we introduce a \textit{Counterfactual Generation} to convert the gender information of words, so the original and the modified embeddings can produce a gender-neutralized word embedding after geometric alignment regularization, without loss of semantic information. From the various quantitative and qualitative debiasing experiments, our method shows to be better than existing debiasing methods in debiasing word embeddings. In addition, Our method shows the ability to preserve semantic information during debiasing by minimizing the semantic information losses for extrinsic NLP downstream tasks.

Mingi Ji, Weonyoung Joo, Kyungwoo Song et al.

Recent studies identified that sequential Recommendation is improved by the attention mechanism. By following this development, we propose Relation-Aware Kernelized Self-Attention (RKSA) adopting a self-attention mechanism of the Transformer with augmentation of a probabilistic model. The original self-attention of Transformer is a deterministic measure without relation-awareness. Therefore, we introduce a latent space to the self-attention, and the latent space models the recommendation context from relation as a multivariate skew-normal distribution with a kernelized covariance matrix from co-occurrences, item characteristics, and user information. This work merges the self-attention of the Transformer and the sequential recommendation by adding a probabilistic model of the recommendation task specifics. We experimented RKSA over the benchmark datasets, and RKSA shows significant improvements compared to the recent baseline models. Also, RKSA were able to produce a latent space model that answers the reasons for recommendation.

Kyungwoo Song, JoonHo Jang, Seung jae Shin et al.

Long Short-Term Memory (LSTM) infers the long term dependency through a cell state maintained by the input and the forget gate structures, which models a gate output as a value in [0,1] through a sigmoid function. However, due to the graduality of the sigmoid function, the sigmoid gate is not flexible in representing multi-modality or skewness. Besides, the previous models lack modeling on the correlation between the gates, which would be a new method to adopt inductive bias for a relationship between previous and current input. This paper proposes a new gate structure with the bivariate Beta distribution. The proposed gate structure enables probabilistic modeling on the gates within the LSTM cell so that the modelers can customize the cell state flow with priors and distributions. Moreover, we theoretically show the higher upper bound of the gradient compared to the sigmoid function, and we empirically observed that the bivariate Beta distribution gate structure provides higher gradient values in training. We demonstrate the effectiveness of bivariate Beta gate structure on the sentence classification, image classification, polyphonic music modeling, and image caption generation.

Kyungwoo Song, Wonsung Lee, Il-Chul Moon

Understanding politics is challenging because the politics take the influence from everything. Even we limit ourselves to the political context in the legislative processes; we need a better understanding of latent factors, such as legislators, bills, their ideal points, and their relations. From the modeling perspective, this is difficult 1) because these observations lie in a high dimension that requires learning on low dimensional representations, and 2) because these observations require complex probabilistic modeling with latent variables to reflect the causalities. This paper presents a new model to reflect and understand this political setting, NIPEN, including factors mentioned above in the legislation. We propose two versions of NIPEN: one is a hybrid model of deep learning and probabilistic graphical model, and the other model is a neural tensor model. Our result indicates that NIPEN successfully learns the manifold of the legislative bill texts, and NIPEN utilizes the learned low-dimensional latent variables to increase the prediction performance of legislators' votings. Additionally, by virtue of being a domain-rich probabilistic model, NIPEN shows the hidden strength of the legislators' trust network and their various characteristics on casting votes.

Kyungwoo Song, Mingi Ji, Sungrae Park et al.

A long user history inevitably reflects the transitions of personal interests over time. The analyses on the user history require the robust sequential model to anticipate the transitions and the decays of user interests. The user history is often modeled by various RNN structures, but the RNN structures in the recommendation system still suffer from the long-term dependency and the interest drifts. To resolve these challenges, we suggest HCRNN with three hierarchical contexts of the global, the local, and the temporary interests. This structure is designed to withhold the global long-term interest of users, to reflect the local sub-sequence interests, and to attend the temporary interests of each transition. Besides, we propose a hierarchical context-based gate structure to incorporate our \textit{interest drift assumption}. As we suggest a new RNN structure, we support HCRNN with a complementary \textit{bi-channel attention} structure to utilize hierarchical context. We experimented the suggested structure on the sequential recommendation tasks with CiteULike, MovieLens, and LastFM, and our model showed the best performances in the sequential recommendations.

Sungrae Park, Kyungwoo Song, Mingi Ji et al.

Successful application processing sequential data, such as text and speech, requires an improved generalization performance of recurrent neural networks (RNNs). Dropout techniques for RNNs were introduced to respond to these demands, but we conjecture that the dropout on RNNs could have been improved by adopting the adversarial concept. This paper investigates ways to improve the dropout for RNNs by utilizing intentionally generated dropout masks. Specifically, the guided dropout used in this research is called as adversarial dropout, which adversarially disconnects neurons that are dominantly used to predict correct targets over time. Our analysis showed that our regularizer, which consists of a gap between the original and the reconfigured RNNs, was the upper bound of the gap between the training and the inference phases of the random dropout. We demonstrated that minimizing our regularizer improved the effectiveness of the dropout for RNNs on sequential MNIST tasks, semi-supervised text classification tasks, and language modeling tasks.

Su-Jin Shin, Kyungwoo Song, Il-Chul Moon

The joint optimization of representation learning and clustering in the embedding space has experienced a breakthrough in recent years. In spite of the advance, clustering with representation learning has been limited to flat-level categories, which often involves cohesive clustering with a focus on instance relations. To overcome the limitations of flat clustering, we introduce hierarchically-clustered representation learning (HCRL), which simultaneously optimizes representation learning and hierarchical clustering in the embedding space. Compared with a few prior works, HCRL firstly attempts to consider a generation of deep embeddings from every component of the hierarchy, not just leaf components. In addition to obtaining hierarchically clustered embeddings, we can reconstruct data by the various abstraction levels, infer the intrinsic hierarchical structure, and learn the level-proportion features. We conducted evaluations with image and text domains, and our quantitative analyses showed competent likelihoods and the best accuracies compared with the baselines.