Study of Effective Valid Page Tracking Methods in Mobile Flash Storage without DRAM

Jaesun No, Jinkyu Jeong

http://doi.org/10.5626/JOK.2025.52.4.357

Mobile systems, commonly use flash-based storage devices like Universal Flash Storage (UFS). These devices are designed with a small form factor and operate under limited power and budget constraints, often lacking large internal DRAM. As a result, they rely on small SRAM to run the Flash Translation Layer (FTL). This limitation makes it difficult to manage metadata, such as address mapping tables and the Valid Page Bitmap related to garbage collection (GC), within SRAM. Managing the Valid Page Bitmap in flash memory poses challenges due to performance degradation from significant metadata I/O overhead. This paper proposes an efficient method for tracking valid pages within specific blocks by managing L2P segment bitmaps per block. This approach minimizes the metadata access overhead during valid page tracking, leading to improved performance. Evaluation results indicate up to an 83% reduction in latency for finding valid pages compared to existing methods, particularly with a 128KB I/O unit.

Beyond Traditional Search: SIMD-Optimized Correction for Learned Index

Yeojin Oh, Nakyeong Kim, Jongmoo Choi, Seehwan Yoo

http://doi.org/10.5626/JOK.2025.52.4.363

To address the limitations of traditional indexing techniques, this study examines the search performance of machine learning-based Learned Indexes, focusing on the read-only RMI and the modifiable ALEX We propose a SIMD-based optimization technique to minimize the overhead incurred during the correction phase, which accounts for over 80% of the total search time. Learned Indexes operate in two phases: prediction and correction. In our experiments with RMI, we found that when the error range is large, the SIMD Branchless Binary Search capable of quickly narrowing down the search range outperforms other methods. In contrast. when the error range is small, the model prediction-based SIMD Linear Search demonstrates superior performance. For ALEX, which maintains a relatively constant error range, the straightforward SIMD Linear Search proved to be the most efficient compared to more complex search techniques. These results underscore the importance of choosing the right search algorithm based on the dataset’s error range, index size, and density to achieve optimal performance.

Reinforcement Learning with the Law of Diminishing Marginal Utility: Efficient and Equitable Resource Allocation in Multi-Agent Systems

Yunsu Lee, Byoung-Tak Zhang

http://doi.org/10.5626/JOK.2025.52.4.374

The law of diminishing marginal utility is an economic theory stating that as additional units of a good are consumed, the utility gained from each additional unit is decreased. We incorporated the law of diminishing marginal utility into multi-agent reinforcement learning for resource allocation, demonstrating that optimal distribution could emerge without direct communication among agents. This approach aligns with market principles, where individual self-Ainterested actions can lead to maximization of total utility. Experimental results in a grid-world environment showed that when two agents competed for two resources, applying the law of diminishing marginal utility led to a more equitable and Pareto-optimal allocation of resources.

Enhancing Molecular Understanding in LLMs through Multimodal Graph-SMILES Representations

http://doi.org/10.5626/JOK.2025.52.4.379

Recent advancements in large language models (LLMs) have shown remarkable performace across various tasks, with increasing focus on multimodal research. Notably, BLIP-2 can enhance performance by efficiently aligning images and text using a Q-Former, aided by an image encoder pre-trained on multimodal data. Inspired by this, the MolCA model extends BLIP-2 to the molecular domain to improve performance. However, the graph encoder in MolCA is pre-trained on unimodal data, necessitating updates during model training, which is a limitation. Therefore, this paper replaced it with a graph encoder pre-trained on multimodal data and frozen while training the model. Experimental results showed that using the graph encoder pre-trained on multimodal data generally enhanced performance. Additionally, unlike the graph encoder pre-trained on unimodal data, which performed better when updated, the graph encoder pre-trained on multimodal data achieved superior results across all metrics when frozen.

Enhancing Passage Selection and Answer Generation in FiD Systems Using Relevance Gating

Seung-ho Choi, Shihyun Park, Minsang Kim, Chansol Park, Junho Wang, Ji-Yoon Kim, Bong-Su Kim

http://doi.org/10.5626/JOK.2025.52.4.385

In this paper, we proposed a novel approach to enhance the performance of the Fusion-in-Decoder (FiD) model in open-domain question answering systems. The FiD model operates by independently encoding multiple passages and then combining them during the decoding stage to generate answers. However, this method has the drawback of not filtering out passages containing unnecessary information, thereby placing an excessive burden on the decoder. To address this issue, we introduced a Relevance Gate inspired by the forget gate of Long Short-Term Memory (LSTM). This gate can evaluate the relevance of each passage in parallel, selectively transmitting information to the decoder, thereby significantly improving the accuracy and efficiency of answer generation. Additionally, we applied a new activation function suitable for open-domain question answering systems instead of the sigmoid function to ensure the model's stability.

Efficient Large Language Model Based Passage Re-Ranking Using Single Token Representations

Jeongwoo Na, Jun Kwon, Eunseong Choi, Jongwuk Lee

http://doi.org/10.5626/JOK.2025.52.4.395

In information retrieval systems, document re-ranking involves reordering a set of candidate documents based on evaluation of their relevance to a given query. Leveraging extensive natural language understanding capabilities of large language models(LLMs), numerous studies on document re-ranking have been conducted, demonstrating groundbreaking performance. However, studies utilizing large language models focus solely on improving reranking performance, resulting in degraded efficiency due to excessively long input sequences and the need for repetitive inference. To address these limitations, we propose ListT5++, a novel model that represents the relevance between a query and a passage using single token embedding and significantly improves the efficiency of LLM-based reranking through a single-step decoding strategy that minimizes the decoding process. Experimental results showed that ListT5++ could maintain accuracy levels comparable to existing methods while reducing inference latency by a factor of 29.4 relative to the baseline. Moreover, our approach demonstrates robust characteristics by being insensitive to th initial ordering of candidate documents, thereby ensuring high practicality in real-time retrieval environments.

Hallucination Detection and Explanation Model for Enhancing the Reliability of LLM Responses

Sujeong Lee, Hayoung Lee, Seongsoo Heo, Wonik Choi

http://doi.org/10.5626/JOK.2025.52.4.404

Recent advancements in large language models (LLMs) have achieved remarkable progress in natural language processing. However, reliability issues persist due to hallucination, which remains a significant challenge. Existing hallucination research primarily focuses on detection, lacking the capability to explain the causes and context of hallucinations. In response, this study proposes a hallucination-specialized model that goes beyond mere detection by providing explanations for identified hallucinations. The proposed model was designed to classify hallucinations while simultaneously generating explanations, allowing users to better trust and understand the model’s responses. Experimental results demonstrated that the proposed model surpassed large-scale models such as Llama3 70B and GPT-4 in hallucination detection accuracy while consistently generating high-quality explanations. Notably, the model maintained stable detection and explanation performance across diverse datasets, showcasing its adaptability. By integrating hallucination detection with explanation generation, this study introduces a novel approach to evaluating hallucinations in language models.

Protein-Ligand Binding Affinity Prediction Using Protein Modality Alignme

Seungyong Lee, Sanghyun Park

http://doi.org/10.5626/JOK.2025.52.4.415

Identifying molecules with high binding affinity to a target protein for drug candidate discovery requires significant resources and time. Deep learning-based protein-ligand binding affinity prediction research plays a crucial role in addressing this challenge. Existing studies have utilized protein sequence and structural information along with ligand 2D structures. However, they have limitations in fully capturing complex interactions. Additionally, while sequence, structure, and surface information are used for protein modeling, previous approaches have struggled to incorporate their dependent relationships into the model. In this paper, we proposed a model that could inject these dependencies by aligning protein sequence, structure, and surface information based on sequence data. Furthermore, our model leverages both 2D structure of the ligand and its 3D representation using an SE(3)-invariant graph neural network. The proposed model outperformed existing baseline models. An ablation study demonstrated the importance of aligning different protein modalities and incorporating both 2D and 3D ligand information.

Enhancing Clustering Quality on Mixed-Type Multivariate Time Series Data of HVAC Simulations through Feature Summarization

Harin Seo, Young-Kyoon Suh

http://doi.org/10.5626/JOK.2025.52.4.424

Existing approaches for multivariate time series data clustering analysis often result in significant information loss, thereby reducing both clustering performance and interpretability. Moreover, most existing techniques primarily focus on numerical variables, making them less effective for real-world datasets that often include both numerical and categorical variables. To address these problems, this paper proposes a novel clustering technique for mixed-type multivariate time series data, enhancing interpretability by summarizing the data into representative features. The proposed technique is fundamentally different from existing methods in that it summarizes features to cluster mixed-type multivariate time series data. We evaluated the proposed method against existing techniques using three clustering evaluation metrics on two HVAC simulation datasets (MZVAV-1 and MZVAV-2-1). Experimental results showed that the proposed method outperformed existing techniques in clustering quality for over 61% of metric–cluster count combinations on MZVAV-1, and over 40% on MZVAV-2-1. These findings confirmed that the proposed technique could significantly improve clustering performance and interpretability for mixed-type time-series data.

An Inference Framework for Text-Based Sequential Recommendation Model Using Nearest Neighbor Mechanism

Junyoung Kim, Hyunsoo, Jongwuk Lee

http://doi.org/10.5626/JOK.2025.52.4.435

Sequential recommendation task aims to predict the next item to interact with based on users’ interaction history. Text-based recommendation models, which represent items as text, show improved performance in cold-start problems and zero-shot recommendation tasks. However, they suffer from textual bias and the lack of collaborative knowledge. To overcome these limitations, we propose a text-based recommendation model inference framework using the nearest neighbor mechanism. The proposed method leverages text-based recommendation models as a neighbor retriever model to search neighbors with similar preferences to the user and aggregate the neighbor information with existing recommendation results to improve recommendation performance. Experiments conducted on four datasets show that the proposed method consistently outperforms existing models, with performance improvement up to 25.27% on NDCG@50. Furthermore, the proposed method effectively complements collaborative knowledge and improves model explainability by providing recommendation rationale.

revention of Malware Installation in Dedicated Devices Built on General-Purpose Execution Environments

Doyeon Kim, Jione Choi, Kiseok Jeon, Wonjun Lee, Junghee Lee

http://doi.org/10.5626/JOK.2025.52.5.444

With digitalization of various industries, the demand for dedicated devices is increasing. Dedicated devices, such as digital banking branches, medical tablets, and educational tablets, are designed to perform specific tasks. Since they only run designated applications, they are them more secure with minimal the attack surface. Most of these devices are built on general-purpose execution environments like Android. Thus, they offer ease of development, usability, and high availability, contributing to their widespread adoption. At the same time, they may introduce new security vulnerabilities, necessitating security measures tailored to dedicated devices. This study analyed the vulnerabilities of dedicated devices operating in a general-purpose execution environment, evaluated the potential for vulnerabilities that could lead to malware installation, and proposed countermeasures. This research assumes that attackers do not have physical access to the device and that end users do not engage in malicious activities. The widely used Android environment was selected. Ten methods by which an attacker could remotely install malware on a Lenovo P11 device were identified. To mitigate these threats, a security mechanism optimized for dedicated devices was designed by implementing SELinux policies and installing a file integrity verification program.