Pilsung Kang

Hankyeol Kim, Pilsung Kang

LLM confidence calibration is often evaluated by comparing two signals: token-probability scores and verbalized confidence. These signals are sometimes treated as direct readouts of model uncertainty, but their comparison depends on measurement choices that are rarely made explicit. In the main analysis, we hold the verbalized-confidence elicitation fixed: a single prompt template, probability scale, and output format. We then vary the measurement axes that define the verbalized-vs-token comparison: which answer string receives the token-probability score, how that score is read from the answer tokens, and under which conditioning context it is measured. We evaluate this design on four QA benchmarks across three open 7--8B base/Instruct model families, with larger Qwen2.5 variants as same-family robustness checks. The resulting comparison is sensitive to these choices: conditioning context changes the sign or magnitude of the ECE gap across settings, token readout produces smaller but still sign-moving changes, and changing the ECE estimator has little effect. Under the default generated-answer, bare-context protocol, Instruct settings are close to parity rather than showing a large calibration gain for verbalized confidence. In a separate supplied-answer analysis, surface-plausible wrong answers receive nearly the same confidence as supplied gold answers, suggesting that verbalized confidence also reflects answer plausibility and provenance rather than correctness alone. We argue that both confidence signals should be treated as protocol-dependent behavioral measurements, and provide a reporting checklist covering elicitation provenance, scored answer, token-probability readout, and conditioning context.

Yunseung Lee, Pilsung Kang

Image anomaly detection problems aim to determine whether an image is abnormal, and to detect anomalous areas. These methods are actively used in various fields such as manufacturing, medical care, and intelligent information. Encoder-decoder structures have been widely used in the field of anomaly detection because they can easily learn normal patterns in an unsupervised learning environment and calculate a score to identify abnormalities through a reconstruction error indicating the difference between input and reconstructed images. Therefore, current image anomaly detection methods have commonly used convolutional encoder-decoders to extract normal information through the local features of images. However, they are limited in that only local features of the image can be utilized when constructing a normal representation owing to the characteristics of convolution operations using a filter of fixed size. Therefore, we propose a vision transformer-based encoder-decoder model, named AnoViT, designed to reflect normal information by additionally learning the global relationship between image patches, which is capable of both image anomaly detection and localization. The proposed approach constructs a feature map that maintains the existing location information of individual patches by using the embeddings of all patches passed through multiple self-attention layers. The proposed AnoViT model performed better than the convolution-based model on three benchmark datasets. In MVTecAD, which is a representative benchmark dataset for anomaly localization, it showed improved results on 10 out of 15 classes compared with the baseline. Furthermore, the proposed method showed good performance regardless of the class and type of the anomalous area when localization results were evaluated qualitatively.

Jaehyuk Heo, YongGi Jeong, Sunwoo Kim et al.

We describe an approach used in the Generic Boundary Event Captioning challenge at the Long-Form Video Understanding Workshop held at CVPR 2022. We designed a Rich Encoder-decoder framework for Video Event CAptioner (REVECA) that utilizes spatial and temporal information from the video to generate a caption for the corresponding the event boundary. REVECA uses frame position embedding to incorporate information before and after the event boundary. Furthermore, it employs features extracted using the temporal segment network and temporal-based pairwise difference method to learn temporal information. A semantic segmentation mask for the attentional pooling process is adopted to learn the subject of an event. Finally, LoRA is applied to fine-tune the image encoder to enhance the learning efficiency. REVECA yielded an average score of 50.97 on the Kinetics-GEBC test data, which is an improvement of 10.17 over the baseline method. Our code is available in https://github.com/TooTouch/REVECA.

Joonghoon Kim, Saeran Park, Kiyoon Jeong et al.

This paper describes the DSBA submissions to the Prompting Large Language Models as Explainable Metrics shared task, where systems were submitted to two tracks: small and large summarization tracks. With advanced Large Language Models (LLMs) such as GPT-4, evaluating the quality of Natural Language Generation (NLG) has become increasingly paramount. Traditional similarity-based metrics such as BLEU and ROUGE have shown to misalign with human evaluation and are ill-suited for open-ended generation tasks. To address this issue, we explore the potential capability of LLM-based metrics, especially leveraging open-source LLMs. In this study, wide range of prompts and prompting techniques are systematically analyzed with three approaches: prompting strategy, score aggregation, and explainability. Our research focuses on formulating effective prompt templates, determining the granularity of NLG quality scores and assessing the impact of in-context examples on LLM-based evaluation. Furthermore, three aggregation strategies are compared to identify the most reliable method for aggregating NLG quality scores. To examine explainability, we devise a strategy that generates rationales for the scores and analyzes the characteristics of the explanation produced by the open-source LLMs. Extensive experiments provide insights regarding evaluation capabilities of open-source LLMs and suggest effective prompting strategies.

Heejeong Choi, Pilsung Kang

Time-series representation learning can extract representations from data with temporal dynamics and sparse labels. When labeled data are sparse but unlabeled data are abundant, contrastive learning, i.e., a framework to learn a latent space where similar samples are close to each other while dissimilar ones are far from each other, has shown outstanding performance. This strategy can encourage varied consistency of time-series representations depending on the positive pair selection and contrastive loss. We propose a new time-series representation learning method by combining the advantages of self-supervised tasks related to contextual, temporal, and transformation consistency. It allows the network to learn general representations for various downstream tasks and domains. Specifically, we first adopt data preprocessing to generate positive and negative pairs for each self-supervised task. The model then performs contextual, temporal, and transformation contrastive learning and is optimized jointly using their contrastive losses. We further investigate an uncertainty weighting approach to enable effective multi-task learning by considering the contribution of each consistency. We evaluate the proposed framework on three downstream tasks: time-series classification, forecasting, and anomaly detection. Experimental results show that our method not only outperforms the benchmark models on these downstream tasks, but also shows efficiency in cross-domain transfer learning.

Takyoung Kim, Hoonsang Yoon, Yukyung Lee et al.

Dialogue state tracking (DST) aims to extract essential information from multi-turn dialogue situations and take appropriate actions. A belief state, one of the core pieces of information, refers to the subject and its specific content, and appears in the form of domain-slot-value. The trained model predicts "accumulated" belief states in every turn, and joint goal accuracy and slot accuracy are mainly used to evaluate the prediction; however, we specify that the current evaluation metrics have a critical limitation when evaluating belief states accumulated as the dialogue proceeds, especially in the most used MultiWOZ dataset. Additionally, we propose relative slot accuracy to complement existing metrics. Relative slot accuracy does not depend on the number of predefined slots, and allows intuitive evaluation by assigning relative scores according to the turn of each dialogue. This study also encourages not solely the reporting of joint goal accuracy, but also various complementary metrics in DST tasks for the sake of a realistic evaluation.

Yukyung Lee, Takyoung Kim, Hoonsang Yoon et al.

Dialogue State Tracking (DST) is critical for comprehensively interpreting user and system utterances, thereby forming the cornerstone of efficient dialogue systems. Despite past research efforts focused on enhancing DST performance through alterations to the model structure or integrating additional features like graph relations, they often require additional pre-training with external dialogue corpora. In this study, we propose DSTEA, improving Dialogue State Tracking via Entity Adaptive pre-training, which can enhance the encoder through by intensively training key entities in dialogue utterances. DSTEA identifies these pivotal entities from input dialogues utilizing four different methods: ontology information, named-entity recognition, the spaCy, and the flair library. Subsequently, it employs selective knowledge masking to train the model effectively. Remarkably, DSTEA only requires pre-training without the direct infusion of extra knowledge into the DST model. This approach resulted in substantial performance improvements of four robust DST models on MultiWOZ 2.0, 2.1, and 2.2, with joint goal accuracy witnessing an increase of up to 2.69% (from 52.41% to 55.10%). Further validation of DSTEA's efficacy was provided through comparative experiments considering various entity types and different entity adaptive pre-training configurations such as masking strategy and masking rate.

Yukyung Lee, Jaehee Kim, Doyoon Kim et al.

As the e-commerce market continues to expand and online transactions proliferate, customer reviews have emerged as a critical element in shaping the purchasing decisions of prospective buyers. Previous studies have endeavored to identify key aspects of customer reviews through the development of sentiment analysis models and topic models. However, extracting specific dissatisfaction factors remains a challenging task. In this study, we delineate the pain point detection problem and propose Painsight, an unsupervised framework for automatically extracting distinct dissatisfaction factors from customer reviews without relying on ground truth labels. Painsight employs pre-trained language models to construct sentiment analysis and topic models, leveraging attribution scores derived from model gradients to extract dissatisfaction factors. Upon application of the proposed methodology to customer review data spanning five product categories, we successfully identified and categorized dissatisfaction factors within each group, as well as isolated factors for each type. Notably, Painsight outperformed benchmark methods, achieving substantial performance enhancements and exceptional results in human evaluations.

Jaehyuk Heo, Pilsung Kang

Active learning (AL) aims to enhance model performance by selectively collecting highly informative data, thereby minimizing annotation costs. However, in practical scenarios, unlabeled data may contain out-of-distribution (OOD) samples, which are not used for training, leading to wasted annotation costs if data is incorrectly selected. Therefore, to make active learning feasible in real-world applications, it is crucial to consider not only the informativeness of unlabeled samples but also their purity to determine whether they belong to the in-distribution (ID). Recent studies have applied AL under these assumptions, but challenges remain due to the trade-off between informativeness and purity, as well as the heavy dependence on OOD samples. These issues lead to the collection of OOD samples, resulting in a significant waste of annotation costs. To address these challenges, we propose a novel query strategy, VLPure-AL, which minimizes cost losses while reducing dependence on OOD samples. VLPure-AL sequentially evaluates the purity and informativeness of data. First, it utilizes a pre-trained vision-language model to detect and exclude OOD data with high accuracy by leveraging linguistic and visual information of ID data. Second, it selects highly informative data from the remaining ID data, and then the selected samples are annotated by human experts. Experimental results on datasets with various open-set conditions demonstrate that VLPure-AL achieves the lowest cost loss and highest performance across all scenarios. Code is available at https://github.com/DSBA-Lab/OpenAL.

Hyeongwon Kang, Jeongseob Kim, Jinwoo Park et al.

Recent studies have explored large language models for time-series anomaly detection, yet existing approaches often rely on a single general-purpose model to directly infer anomaly indices or intervals, limiting controllability, interpretability, and reliability for complex anomaly patterns. We propose SAGE (Specialized Analyzer Group for Expert-like Detection), a multi-agent framework for structured anomaly diagnosis in univariate time series. It decomposes anomaly analysis into four specialized Analyzers for point, structural, seasonal, and pattern anomalies. Each Analyzer applies family-specific numerical tools and diagnostic visualizations to generate evidence, while an evidence-grounded Detector consolidates the evidence into confidence-scored anomaly records with intervals and candidate types. A Supervisor then converts these structured records into analyst-facing diagnostic reports. SAGE further constructs synthetic in-context examples from normal-reference training segments, without using real anomalous segments or anomaly-type labels as in-context examples. Across three benchmarks, SAGE achieves the best average performance among strong ML/DL and language-model-based baselines. Ablation studies and human evaluation further show that the proposed framework improves detection reliability and the practical usefulness of diagnostic outputs.

Gunho No, Yukyung Lee, Hyeongwon Kang et al.

As the IT industry advances, system log data becomes increasingly crucial. Many computer systems rely on log texts for management due to restricted access to source code. The need for log anomaly detection is growing, especially in real-world applications, but identifying anomalies in rapidly accumulating logs remains a challenging task. Traditional deep learning-based anomaly detection models require dataset-specific training, leading to corresponding delays. Notably, most methods only focus on sequence-level log information, which makes the detection of subtle anomalies harder, and often involve inference processes that are difficult to utilize in real-time. We introduce RAPID, a model that capitalizes on the inherent features of log data to enable anomaly detection without training delays, ensuring real-time capability. RAPID treats logs as natural language, extracting representations using pre-trained language models. Given that logs can be categorized based on system context, we implement a retrieval-based technique to contrast test logs with the most similar normal logs. This strategy not only obviates the need for log-specific training but also adeptly incorporates token-level information, ensuring refined and robust detection, particularly for unseen logs. We also propose the core set technique, which can reduce the computational cost needed for comparison. Experimental results show that even without training on log data, RAPID demonstrates competitive performance compared to prior models and achieves the best performance on certain datasets. Through various research questions, we verified its capability for real-time detection without delay.

Kiyoon Jeong, Woojun Lee, Woongchan Nam et al.

This report presents the ECO (Ensembled Clip score and cOnsensus score) pipeline from team DSBA LAB, which is a new framework used to evaluate and rank captions for a given image. ECO selects the most accurate caption describing image. It is made possible by combining an Ensembled CLIP score, which considers the semantic alignment between the image and captions, with a Consensus score that accounts for the essentialness of the captions. Using this framework, we achieved notable success in the CVPR 2024 Workshop Challenge on Caption Re-ranking Evaluation at the New Frontiers for Zero-Shot Image Captioning Evaluation (NICE). Specifically, we secured third place based on the CIDEr metric, second in both the SPICE and METEOR metrics, and first in the ROUGE-L and all BLEU Score metrics. The code and configuration for the ECO framework are available at https://github.com/DSBA-Lab/ECO .

Pilsung Kang

Barren plateaus (BPs) are usually described by the exponential suppression of gradient variance, but the mechanism by which gradient signal disappears remains unclear. We show that this phenomenon can be understood as destructive interference among termwise gradient contributions. To make this perspective operational, we introduce a diagnostic framework based on the cancellation ratio $R_k$, the effective term count $N_{\mathrm{eff},k}$, and the interference-quality measure $B_{\mathrm{eff},k}=R_k\sqrt{N_{\mathrm{eff},k}}$. Under a random-sign model, $B_{\mathrm{eff},k}$ remains near a stable baseline, defining a random-sign cancellation regime. For the transverse-field Ising model (TFIM), we find that the hardware-efficient ansatz (HEA) remains close to this regime across system sizes and depths, whereas the Hamiltonian variational ansatz (HVA) systematically escapes it. In particular, HVA exhibits larger $B_{\mathrm{eff},k}$ not merely because $N_{\mathrm{eff},k}$ is larger, but because $R_k$ also remains systematically larger despite the broader term participation. This pattern indicates improved sign organization rather than simple term suppression. We further establish an exact identity that connects the proposed interference diagnostics directly to the standard variance-based theory of BPs. These results position destructive interference as a mechanistic interpretation of BP-like behavior in the regimes studied here, but they do not imply that BPs and destructive interference are universally interchangeable across all architectures and settings.

Jaehee Kim, Pilsung Kang

Modern LLMs are increasingly accessed via black-box APIs, requiring users to transmit sensitive prompts, outputs, and fine-tuning data to external providers, creating a critical privacy risk at the API boundary. We introduce AlienLM, a deployable API-only privacy layer that protects text by translating it into an Alien Language via a vocabulary-scale bijection, enabling lossless recovery on the client side. Using only standard fine-tuning APIs, Alien Adaptation Training (AAT) adapts target models to operate directly on alienized inputs. Across four LLM backbones and seven benchmarks, AlienLM retains over 81\% of plaintext-oracle performance on average, substantially outperforming random-bijection and character-level baselines. Under adversaries with access to model weights, corpus statistics, and learning-based inverse translation, recovery attacks reconstruct fewer than 0.22\% of alienized tokens. Our results demonstrate a practical pathway for privacy-preserving LLM deployment under API-only access, substantially reducing plaintext exposure while maintaining task performance.

Jinwoo Park, Hyeongwon Kang, Seung Hun Han et al.

Time series anomaly detection is a critical task across various industrial domains. However, capturing temporal dependencies and multivariate correlations within patch-level representation learning remains underexplored, and reliance on single-scale patterns limits the detection of anomalies across different temporal ranges. Furthermore, focusing on normal data representations makes models vulnerable to distribution shifts at inference time. To address these limitations, we propose Codebook-based Online-adaptive Multi-scale Embedding for Time-series anomaly detection (COMET), which consists of three key components: (1) Multi-scale Patch Encoding captures temporal dependencies and inter-variable correlations across multiple patch scales. (2) Vector-Quantized Coreset learns representative normal patterns via codebook and detects anomalies with a dual-score combining quantization error and memory distance. (3) Online Codebook Adaptation generates pseudo-labels based on codebook entries and dynamically adapts the model at inference through contrastive learning. Experiments on five benchmark datasets demonstrate that COMET achieves the best performance in 36 out of 45 evaluation metrics, validating its effectiveness across diverse environments.

Hyeongwon Kang, Jinwoo Park, Seunghun Han et al.

Detecting anomalies in time-series data is critical in domains such as industrial operations, finance, and cybersecurity, where early identification of abnormal patterns is essential for ensuring system reliability and enabling preventive maintenance. However, most existing methods are reactive: they detect anomalies only after they occur and lack the capability to provide proactive early warning signals. In this paper, we propose FATE (Forecasting Anomalies with Time-series Ensembles), a novel unsupervised framework for detecting Precursors-of-Anomaly (PoA) by quantifying predictive uncertainty from a diverse ensemble of time-series forecasting models. Unlike prior approaches that rely on reconstruction errors or require ground-truth labels, FATE anticipates future values and leverages ensemble disagreement to signal early signs of potential anomalies without access to target values at inference time. To rigorously evaluate PoA detection, we introduce Precursor Time-series Aware Precision and Recall (PTaPR), a new metric that extends the traditional Time-series Aware Precision and Recall (TaPR) by jointly assessing segment-level accuracy, within-segment coverage, and temporal promptness of early predictions. This enables a more holistic assessment of early warning capabilities that existing metrics overlook. Experiments on five real-world benchmark datasets show that FATE achieves an average improvement of 19.9 percentage points in PTaPR AUC and 20.02 percentage points in early detection F1 score, outperforming baselines while requiring no anomaly labels. These results demonstrate the effectiveness and practicality of FATE for real-time unsupervised early warning in complex time-series environments.

Pilsung Kang

Classical approaches often treat interaction as engineered product terms or as emergent patterns in flexible models, offering little control over how synergy or antagonism arises. We take a quantum-inspired view: following the Born rule (probability as squared amplitude), \emph{coherent} aggregation sums complex amplitudes before squaring, creating an interference cross-term, whereas an \emph{incoherent} proxy sums squared magnitudes and removes it. In a minimal linear-amplitude model, this cross-term equals the standard potential-outcome interaction contrast \(Δ_{\mathrm{INT}}\) in a \(2\times 2\) factorial design, giving relative phase a direct, mechanism-level control over synergy versus antagonism. We instantiate this idea in a lightweight \emph{Interference Kernel Classifier} (IKC) and introduce two diagnostics: \emph{Coherent Gain} (log-likelihood gain of coherent over the incoherent proxy) and \emph{Interference Information} (the induced Kullback-Leibler gap). A controlled phase sweep recovers the identity. On a high-interaction synthetic task (XOR), IKC outperforms strong baselines under paired, budget-matched comparisons; on real tabular data (\emph{Adult} and \emph{Bank Marketing}) it is competitive overall but typically trails the most capacity-rich baseline in paired differences. Holding learned parameters fixed, toggling aggregation from incoherent to coherent consistently improves negative log-likelihood, Brier score, and expected calibration error, with positive Coherent Gain on both datasets.

Hankyeol Kim, Pilsung Kang

Physics-Informed Neural Networks (PINNs) are effective methods for solving inverse problems and discovering governing equations from observational data. However, their performance degrades significantly under complex measurement noise and gross outliers. To address this issue, we propose the Noise-Adaptive Physics-Informed Neural Network (naPINN), which robustly recovers physical solutions from corrupted measurements without prior knowledge of the noise distribution. naPINN embeds an energy-based model into the training loop to learn the latent distribution of prediction residuals. Leveraging the learned energy landscape, a trainable reliability gate adaptively filters data points exhibiting high energy, while a rejection cost regularization prevents trivial solutions where valid data are discarded. We demonstrate the efficacy of naPINN on various benchmark partial differential equations corrupted by non-Gaussian noise and varying rates of outliers. The results show that naPINN significantly outperforms existing robust PINN baselines, successfully isolating outliers and accurately reconstructing the dynamics under severe data corruption.

Yukyung Lee, Joonghoon Kim, Jaehee Kim et al.

Existing LLM-as-a-Judge approaches for evaluating text generation suffer from rating inconsistencies, with low agreement and high rating variance across different evaluator models. We attribute this to subjective evaluation criteria combined with Likert scale scoring in existing protocols. To address this issue, we introduce CheckEval, a checklist-based evaluation framework that improves rating reliability via decomposed binary questions. Through experiments with 12 evaluator models across multiple datasets, we first demonstrate that CheckEval strongly correlates with human judgments. More importantly, CheckEval dramatically improves the average agreement across evaluator models by 0.45 and reduces the score variance. CheckEval scores furthermore have the benefit of being more interpretable because it decomposes evaluation criteria into traceable binary decisions, allowing analyses of specific attributes driving quality judgments.

Yukyung Lee, Soonwon Ka, Bokyung Son et al.

Large Language Models (LLMs) have impacted the writing process, enhancing productivity by collaborating with humans in content creation platforms. However, generating high-quality, user-aligned text to satisfy real-world content creation needs remains challenging. We propose WritingPath, a framework that uses explicit outlines to guide LLMs in generating goal-oriented, high-quality text. Our approach draws inspiration from structured writing planning and reasoning paths, focusing on reflecting user intentions throughout the writing process. To validate our approach in real-world scenarios, we construct a diverse dataset from unstructured blog posts to benchmark writing performance and introduce a comprehensive evaluation framework assessing the quality of outlines and generated texts. Our evaluations with various LLMs demonstrate that the WritingPath approach significantly enhances text quality according to evaluations by both LLMs and professional writers.

Pilsung Kang

Bell's theorem reveals a profound conflict between quantum mechanics and local realism, a conflict we reinterpret through the modern lens of causal inference. We propose and computationally validate a framework where quantum entanglement acts as a "super-confounding" resource, generating correlations that violate the classical causal bounds set by Bell's inequalities. This work makes three key contributions: First, we establish a physical hierarchy of confounding (Quantum > Classical) and introduce Confounding Strength (CS) to quantify this effect. Second, we provide a circuit-based implementation of the quantum $\mathcal{DO}$-calculus to distinguish causality from spurious correlation. Finally, we apply this calculus to a quantum machine learning problem, where causal feature selection yields a statistically significant 11.3% average absolute improvement in model robustness. Our framework bridges quantum foundations and causal AI, offering a new, practical perspective on quantum correlations.

Pilsung Kang

Distinguishing correlation from causation is a fundamental challenge in machine intelligence, often representing a critical barrier to building robust and trustworthy systems. While Pearl's $\mathcal{DO}$-calculus provides a rigorous framework for causal inference, a parallel challenge lies in its physical implementation. Here, we apply and experimentally validate a quantum algorithmic framework for performing causal interventions. Our approach maps causal networks onto quantum circuits where probabilistic links are encoded by controlled-rotation gates, and interventions are realized by a structural remodeling of the circuit -- a physical analogue to Pearl's ``graph surgery''. We demonstrate the method's efficacy by resolving Simpson's Paradox in a 3-qubit model, and show its scalability by quantifying confounding bias in a 10-qubit healthcare simulation. Critically, we provide a proof-of-principle experimental validation on an IonQ Aria quantum computer, successfully reproducing the paradox and its resolution in the presence of real-world noise. This work establishes a practical pathway for quantum causal inference, offering a new computational tool to address deep-rooted challenges in algorithmic fairness and explainable AI (XAI).

Jaehyuk Heo, Pilsung Kang

Recent advances in image anomaly detection have extended unsupervised learning-based models from single-class settings to multi-class frameworks, aiming to improve efficiency in training time and model storage. When a single model is trained to handle multiple classes, it often underperforms compared to class-specific models in terms of per-class detection accuracy. Accordingly, previous studies have primarily focused on narrowing this performance gap. However, the way class information is used, or not used, remains a relatively understudied factor that could influence how detection thresholds are defined in multi-class image anomaly detection. These thresholds, whether class-specific or class-agnostic, significantly affect detection outcomes. In this study, we identify and formalize the requirements that a multi-class image anomaly detection model must satisfy under different conditions, depending on whether class labels are available during training and evaluation. We then re-examine existing methods under these criteria. To meet these challenges, we propose Hierarchical Coreset (HierCore), a novel framework designed to satisfy all defined requirements. HierCore operates effectively even without class labels, leveraging a hierarchical memory bank to estimate class-wise decision criteria for anomaly detection. We empirically validate the applicability and robustness of existing methods and HierCore under four distinct scenarios, determined by the presence or absence of class labels in the training and evaluation phases. The experimental results demonstrate that HierCore consistently meets all requirements and maintains strong, stable performance across all settings, highlighting its practical potential for real-world multi-class anomaly detection tasks.

Pilsung Kang

Parameterized quantum circuits (PQCs) have recently emerged as promising components for enhancing the expressibility of neural architectures. In this work, we introduce QFFN-BERT, a hybrid quantum-classical transformer where the feedforward network (FFN) modules of a compact BERT variant are replaced by PQC-based layers. This design is motivated by the dominant parameter contribution of FFNs, which account for approximately two-thirds of the parameters within standard Transformer encoder blocks. While prior studies have primarily integrated PQCs into self-attention modules, our work focuses on the FFN and systematically investigates the trade-offs between PQC depth, expressibility, and trainability. Our final PQC architecture incorporates a residual connection, both $R_Y$ and $R_Z$ rotations, and an alternating entanglement strategy to ensure stable training and high expressibility. Our experiments, conducted on a classical simulator, on the SST-2 and DBpedia benchmarks demonstrate two key findings. First, a carefully configured QFFN-BERT achieves up to 102.0% of the baseline accuracy, surpassing its classical counterpart in a full-data setting while reducing FFN-specific parameters by over 99%. Second, our model exhibits a consistent and competitive edge in few-shot learning scenarios, confirming its potential for superior data efficiency. These results, supported by an ablation study on a non-optimized PQC that failed to learn, confirm that PQCs can serve as powerful and parameter-efficient alternatives to classical FFNs when co-designed with foundational deep learning principles.

Kiyoon Jeong, Jaehyuk Heo, Junyeong Son et al.

Zero-shot anomaly detection (ZSAD) in images is an approach that can detect anomalies without access to normal samples, which can be beneficial in various realistic scenarios where model training is not possible. However, existing ZSAD research has shown limitations by either not considering domain adaptation of general-purpose backbone models to anomaly detection domains or by implementing only partial adaptation to some model components. In this paper, we propose HeadCLIP to overcome these limitations by effectively adapting both text and image encoders to the domain. HeadCLIP generalizes the concepts of normality and abnormality through learnable prompts in the text encoder, and introduces learnable head weights to the image encoder to dynamically adjust the features held by each attention head according to domain characteristics. Additionally, we maximize the effect of domain adaptation by introducing a joint anomaly score that utilizes domain-adapted pixel-level information for image-level anomaly detection. Experimental results using multiple real datasets in both industrial and medical domains show that HeadCLIP outperforms existing ZSAD techniques at both pixel and image levels. In the industrial domain, improvements of up to 4.9%p in pixel-level mean anomaly detection score (mAD) and up to 3.0%p in image-level mAD were achieved, with similar improvements (3.2%p, 3.1%p) in the medical domain.

Jaewon Cheon, Pilsung Kang

The growing size of large language models has created significant computational inefficiencies. To address this challenge, sparse activation methods selectively deactivates non-essential parameters during inference, reducing computational costs in FFNN layers. While existing methods focus on non-linear gating mechanisms, we hypothesize that the sparsity of the FFNN layer lies globally in the form of a linear combination over its internal down projection matrix. Based on this insight, we propose two methods: M-COUNTDOWN, leveraging indirect coefficients, and D-COUNTDOWN, utilizing direct coefficients of the linear combination. Experimental results demonstrate that D-COUNTDOWN can omit 90% of computations with performance loss as low as 5.5% ideally, while M-COUNTDOWN provides a predictor-free solution with up to 29.4% better performance preservation compared to existing methods. Our specialized kernel implementations effectively realize these theoretical gains into substantial real-world acceleration.

Yookyung Kho, Jaehee Kim, Pilsung Kang

Recently, prompt-based fine-tuning has garnered considerable interest as a core technique for few-shot text classification task. This approach reformulates the fine-tuning objective to align with the Masked Language Modeling (MLM) objective. Leveraging unlabeled data, prompt-based self-training has shown greater effectiveness in binary and three-class classification. However, prompt-based self-training for multi-class classification has not been adequately investigated, despite its significant applicability to real-world scenarios. Moreover, extending current methods to multi-class classification suffers from the verbalizer that extracts the predicted value of manually pre-defined single label word for each class from MLM predictions. Consequently, we introduce a novel, efficient verbalizer structure, named Mapping-free Automatic Verbalizer (MAV). Comprising two fully connected layers, MAV serves as a trainable verbalizer that automatically extracts the requisite word features for classification by capitalizing on all available information from MLM predictions. Experimental results on five multi-class classification datasets indicate MAV's superior self-training efficacy.

Heejeong Choi, Subin Kim, Pilsung Kang

As large-scale time-series data can easily be found in real-world applications, multivariate time-series anomaly detection has played an essential role in diverse industries. It enables productivity improvement and maintenance cost reduction by preventing malfunctions and detecting anomalies based on time-series data. However, multivariate time-series anomaly detection is challenging because real-world time-series data exhibit complex temporal dependencies. For this task, it is crucial to learn a rich representation that effectively contains the nonlinear temporal dynamics of normal behavior. In this study, we propose an unsupervised multivariate time-series anomaly detection model named RAE-MEPC which learns informative normal representations based on multi-resolution ensemble and predictive coding. We introduce multi-resolution ensemble encoding to capture the multi-scale dependency from the input time series. The encoder hierarchically aggregates the temporal features extracted from the sub-encoders with different encoding lengths. From these encoded features, the reconstruction decoder reconstructs the input time series based on multi-resolution ensemble decoding where lower-resolution information helps to decode sub-decoders with higher-resolution outputs. Predictive coding is further introduced to encourage the model to learn the temporal dependencies of the time series. Experiments on real-world benchmark datasets show that the proposed model outperforms the benchmark models for multivariate time-series anomaly detection.

Yukyung Lee, Jina Kim, Pilsung Kang

The system log generated in a computer system refers to large-scale data that are collected simultaneously and used as the basic data for determining errors, intrusion and abnormal behaviors. The aim of system log anomaly detection is to promptly identify anomalies while minimizing human intervention, which is a critical problem in the industry. Previous studies performed anomaly detection through algorithms after converting various forms of log data into a standardized template using a parser. Particularly, a template corresponding to a specific event should be defined in advance for all the log data using which the information within the log key may get lost. In this study, we propose LAnoBERT, a parser free system log anomaly detection method that uses the BERT model, exhibiting excellent natural language processing performance. The proposed method, LAnoBERT, learns the model through masked language modeling, which is a BERT-based pre-training method, and proceeds with unsupervised learning-based anomaly detection using the masked language modeling loss function per log key during the test process. In addition, we also propose an efficient inference process to establish a practically applicable pipeline to the actual system. Experiments on three well-known log datasets, i.e., HDFS, BGL, and Thunderbird, show that not only did LAnoBERT yield a higher anomaly detection performance compared to unsupervised learning-based benchmark models, but also it resulted in a comparable performance with supervised learning-based benchmark models.

Jounghee Kim, Pilsung Kang

Wav2vec 2.0 is an end-to-end framework of self-supervised learning for speech representation that is successful in automatic speech recognition (ASR), but most of the work on the topic has been developed with a single language: English. Therefore, it is unclear whether the self-supervised framework is effective in recognizing other languages with different writing systems, such as Korean which uses the Hangul having a unique writing system. In this paper, we present K-Wav2Vec 2.0, which is a modified version of Wav2vec 2.0 designed for Korean automatic speech recognition by exploring and optimizing various factors of the original Wav2vec 2.0. In fine-tuning, we propose a multi-task hierarchical architecture to reflect the Korean writing structure. Moreover, a joint decoder is applied to alleviate the problem of words existing outside of the vocabulary. In pre-training, we attempted the cross-lingual transfer of the pre-trained model by further pre-training the English Wav2vec 2.0 on a Korean dataset, considering limited resources. Our experimental results demonstrate that the proposed method yields the best performance on both Korean ASR datasets: Ksponspeech (a large-scale Korean speech corpus) and Clovacall (a call-based dialog corpus). Further pre-training is also effective in language adaptation, leading to large improvements without additional data.

Takyoung Kim, Yukyung Lee, Hoonsang Yoon et al.

The primary purpose of dialogue state tracking (DST), a critical component of an end-to-end conversational system, is to build a model that responds well to real-world situations. Although we often change our minds from time to time during ordinary conversations, current benchmark datasets do not adequately reflect such occurrences and instead consist of over-simplified conversations, in which no one changes their mind during a conversation. As the main question inspiring the present study, "Are current benchmark datasets sufficiently diverse to handle casual conversations in which one changes their mind after a certain topic is over?" We found that the answer is "No" because DST models cannot refer to previous user preferences when template-based turnback utterances are injected into the dataset. Even in the the simplest mind-changing (turnback) scenario, the performance of DST models significantly degenerated. However, we found that this performance degeneration can be recovered when the turnback scenarios are explicitly designed in the training set, implying that the problem is not with the DST models but rather with the construction of the benchmark dataset.

Junghoon Lee, Jounghee Kim, Pilsung Kang

Language models (LMs) pretrained on a large text corpus and fine-tuned on a downstream text corpus and fine-tuned on a downstream task becomes a de facto training strategy for several natural language processing (NLP) tasks. Recently, an adaptive pretraining method retraining the pretrained language model with task-relevant data has shown significant performance improvements. However, current adaptive pretraining methods suffer from underfitting on the task distribution owing to a relatively small amount of data to re-pretrain the LM. To completely use the concept of adaptive pretraining, we propose a back-translated task-adaptive pretraining (BT-TAPT) method that increases the amount of task-specific data for LM re-pretraining by augmenting the task data using back-translation to generalize the LM to the target task domain. The experimental results show that the proposed BT-TAPT yields improved classification accuracy on both low- and high-resource data and better robustness to noise than the conventional adaptive pretraining method.

Youngbin Ro, Yukyung Lee, Pilsung Kang

In this paper, we propose Multi$^2$OIE, which performs open information extraction (open IE) by combining BERT with multi-head attention. Our model is a sequence-labeling system with an efficient and effective argument extraction method. We use a query, key, and value setting inspired by the Multimodal Transformer to replace the previously used bidirectional long short-term memory architecture with multi-head attention. Multi$^2$OIE outperforms existing sequence-labeling systems with high computational efficiency on two benchmark evaluation datasets, Re-OIE2016 and CaRB. Additionally, we apply the proposed method to multilingual open IE using multilingual BERT. Experimental results on new benchmark datasets introduced for two languages (Spanish and Portuguese) demonstrate that our model outperforms other multilingual systems without training data for the target languages.

Myeongjun Jang, Pilsung Kang

Sentence embedding is a significant research topic in the field of natural language processing (NLP). Generating sentence embedding vectors reflecting the intrinsic meaning of a sentence is a key factor to achieve an enhanced performance in various NLP tasks such as sentence classification and document summarization. Therefore, various sentence embedding models based on supervised and unsupervised learning have been proposed after the advent of researches regarding the distributed representation of words. They were evaluated through semantic textual similarity (STS) tasks, which measure the degree of semantic preservation of a sentence and neural network-based supervised embedding models generally yielded state-of-the-art performance. However, these models have a limitation in that they have multiple parameters to update, thereby requiring a tremendous amount of labeled training data. In this study, we propose an efficient approach that learns a transition matrix that refines a sentence embedding vector to reflect the latent semantic meaning of a sentence. The proposed method has two practical advantages; (1) it can be applied to any sentence embedding method, and (2) it can achieve robust performance in STS tasks irrespective of the number of training examples.

Myeongjun Jang, Pilsung Kang

Sentence embedding is an important research topic in natural language processing. It is essential to generate a good embedding vector that fully reflects the semantic meaning of a sentence in order to achieve an enhanced performance for various natural language processing tasks, such as machine translation and document classification. Thus far, various sentence embedding models have been proposed, and their feasibility has been demonstrated through good performances on tasks following embedding, such as sentiment analysis and sentence classification. However, because the performances of sentence classification and sentiment analysis can be enhanced by using a simple sentence representation method, it is not sufficient to claim that these models fully reflect the meanings of sentences based on good performances for such tasks. In this paper, inspired by human language recognition, we propose the following concept of semantic coherence, which should be satisfied for a good sentence embedding method: similar sentences should be located close to each other in the embedding space. Then, we propose the Paraphrase-Thought (P-thought) model to pursue semantic coherence as much as possible. Experimental results on two paraphrase identification datasets (MS COCO and STS benchmark) show that the P-thought models outperform the benchmarked sentence embedding methods.

Junhong Kim, Pilsung Kang

Keystroke dynamics-based user authentication (KDA) based on long and freely typed text is an enhanced user authentication method that can not only identify the validity of current users during login but also continuously monitors the consistency of typing behavior after the login process. Previous long and freely typed text-based KDA methods had difficulty incorporating the key sequence information and handling variable lengths of keystrokes, which in turn resulted in lower authentication performance compared to KDA methods based on short and fixed-length text. To overcome these limitations, we propose a recurrent neural network (RNN)-based KDA model. As the RNN model can process an arbitrary length of input and target sequences, our proposed model takes two consecutive keys as the input sequence and actual typing time for the corresponding key sequence as the target sequence. Based on experimental results involving 120 participants, our proposed RNN-KDA model yielded the best authentication performance for all training and test length combinations in terms of equal error rate (EER). It achieved a 5%-6% EER using only 10 test keystrokes while the EERs of other benchmark methods were above 20%. In addition, its performance steadily and more rapidly improves compared to the benchmark methods when the length of training keystrokes increases.

Myeongjun Jang, Seungwan Seo, Pilsung Kang

Sequence-to-sequence (Seq2seq) models have played an important role in the recent success of various natural language processing methods, such as machine translation, text summarization, and speech recognition. However, current Seq2seq models have trouble preserving global latent information from a long sequence of words. Variational autoencoder (VAE) alleviates this problem by learning a continuous semantic space of the input sentence. However, it does not solve the problem completely. In this paper, we propose a new recurrent neural network (RNN)-based Seq2seq model, RNN semantic variational autoencoder (RNN--SVAE), to better capture the global latent information of a sequence of words. To reflect the meaning of words in a sentence properly, without regard to its position within the sentence, we construct a document information vector using the attention information between the final state of the encoder and every prior hidden state. Then, the mean and standard deviation of the continuous semantic space are learned by using this vector to take advantage of the variational method. By using the document information vector to find the semantic space of the sentence, it becomes possible to better capture the global latent feature of the sentence. Experimental results of three natural language tasks (i.e., language modeling, missing word imputation, paraphrase identification) confirm that the proposed RNN--SVAE yields higher performance than two benchmark models.

Gichang Lee, Jaeyun Jeong, Seungwan Seo et al.

In order to maximize the applicability of sentiment analysis results, it is necessary to not only classify the overall sentiment (positive/negative) of a given document but also to identify the main words that contribute to the classification. However, most datasets for sentiment analysis only have the sentiment label for each document or sentence. In other words, there is no information about which words play an important role in sentiment classification. In this paper, we propose a method for identifying key words discriminating positive and negative sentences by using a weakly supervised learning method based on a convolutional neural network (CNN). In our model, each word is represented as a continuous-valued vector and each sentence is represented as a matrix whose rows correspond to the word vector used in the sentence. Then, the CNN model is trained using these sentence matrices as inputs and the sentiment labels as the output. Once the CNN model is trained, we implement the word attention mechanism that identifies high-contributing words to classification results with a class activation map, using the weights from the fully connected layer at the end of the learned CNN model. In order to verify the proposed methodology, we evaluated the classification accuracy and inclusion rate of polarity words using two movie review datasets. Experimental result show that the proposed model can not only correctly classify the sentence polarity but also successfully identify the corresponding words with high polarity scores.