Vol. 52, No. 8,
Aug. 2025
Digital Library
Efficient Memory Management Techniques for LLM Inference in Mobile System
Hyunjeong Shim, Seoyoung Ko, Wanju Doh, Jung Ho Ahn
http://doi.org/10.5626/JOK.2025.52.8.637
On-device LLMs have gained increased attention due to privacy and network latency issues associated with cloud-based LLMs. However, the memory management policies in mobile operating systems have limitations in efficiently handling memory resources during LLM inference. In this paper, we propose two techniques, Initial KV Cache Swap and Deferred Weight Reclamation, which leverage zRAM for preallocated KV cache and reduce storage I/O by deferring weight eviction, leading to enhanced LLM inference performance. Our proposed approach achieves up to a 27% reduction in memory usage compared to the default Linux kernel, optimizing LLM inference performance in memory-constrained mobile environments. Moreover, our approach yields greater memory savings as the number of candidate paths increases in inference techniques such as speculative decoding, demonstrating its effectiveness in supporting diverse LLM decoding techniques on mobile devices.
Controllable High-resolution Cloth Image Generation based on the Diffusion Model
http://doi.org/10.5626/JOK.2025.52.8.644
This paper presents a novel approach for generating high-resolution clothing images using ControlNet, which enables image manipulation guided by both images and prompts. The objective is to create clothing images that reflect the preferences of designers or users, customized with prompts and flat sketches. Given the challenges of obtaining flat sketches from designers, this study employs a pseudo flat sketch—referred to as the pseud flat sketch—derived from an edge extraction algorithm as input for ControlNet. Extensive experiments were conducted under various learning conditions based on pseudo flat sketches generated by DiffusionEdge, which was pre-trained on the BIPED and BSDS datasets, alongside prompt generation for clothing data using BLIP. The results indicate that the model, newly prompted by BLIP and DiffusionEdge pre-trained on BIPED, exhibits superior performance across four main evaluation metrics, demonstrating the feasibility of creating customized high-resolution clothing images that align with user prompts. Additionally, the experimental results on controllability reveal that the fine-tuned ControlNet can adjust colors and patterns in accordance with user prompts.
Are Early Layers of Encoder-based Large Language Models Effective in Code Classification?
Changsup Lee, Suhwan Ji, Hyeonseung Im
http://doi.org/10.5626/JOK.2025.52.8.654
Encoder-based models are used in code classification due to their ability to effectively represent data. A recently proposed methodology, EarlyBIRD, demonstrated that using the outputs from the early layers of encoder-based models can effectively perform the given task. However, this study only used the CodeBERT model and showed its effectiveness in specific tasks. In this paper, we apply EarlyBIRD to various tasks using the encoder-decoder-based CodeT5 model and discuss its effects. Experimental results showed a 13.79%p performance improvement when the language model was not pre-trained on the programming language of the task, but only a 0.41%p improvement when pre-trained on a similar language. Additionally, the performance of the encoder-decoder-based model without applying EarlyBIRD was similar to the best performance of encoder-based models with EarlyBIRD. It was also found that EarlyBIRD was not effective because it was difficult to pre-select which early layers should be used.
Temporal Pattern-Based Credit Default Prediction: Time-Series Data Imbalance Mitigation and Deep Learning Application
Taehyoung Kwon, Eungseon An, Doguk Kim
http://doi.org/10.5626/JOK.2025.52.8.660
Credit default cases are considerably rarer than non-default cases, leading to a significant class imbalance issue. This imbalance negatively impacts the performance of predictive models. To tackle this problem, this study introduces T-SMOTE, a time-series-based data augmentation technique. Unlike traditional SMOTE, T-SMOTE leverages the continuity of time-series data to generate samples that are closer to the boundaries, thereby enhancing model performance. However, the original T-SMOTE had a limitation in handling short time-series data, which was addressed by incorporating the Zero-Padding technique. Comparative experiments using data from American Express showed that T-SMOTE effectively mitigates the data imbalance problem. These findings suggest that advanced data augmentation technologies can create new opportunities for credit risk management in the financial industry.
Enhancing Automated Program Repair using Patch Lightweighting and Context Information
Eunseo Jung, Abdinabiev Aslan Safarovich, Byungjeong Lee
http://doi.org/10.5626/JOK.2025.52.8.670
Large Language Models LLMs play a crucial role in the Automated Program RepairAPR field. However, their effectiveness is constrained by token limitations. When the number of tokens exceeds the model’s capacity, it struggles to fully utilize its capabilities, often failing to correctly detect and fix bugs. This study proposed an approach that could leverage patch lightweighting and context information to overcome these constraints. By incorporating the most semantically similar method as a context method and applying patch lightweighting to long methods, we ensured that the methods remained within the LLM’s token limit. Through this approach, experimental results demonstrated that effective bug fixing could be achieved with fewer tokens, improving repair efficiency.
Layout Code Generation using Large Multimodal Models
Yangsoo Choi, Jeongwoo Na, Dongcheol Lee, Jongwuk Lee
http://doi.org/10.5626/JOK.2025.52.8.677
GUI layout generation entails the analysis and organization of user interface components into structured formats. This paper introduces a novel method that leverages Large Multimodal Models (LMMs) to transform GUI layout images into structured code. The proposed framework enables LMMs to effectively comprehend both the visual and structural attributes of GUI images and produce the corresponding layout code without requiring additional training. The method begins by extracting feature vectors from an input image, followed by retrieving similar examples and applying visual and spatial augmentation techniques to create few-shot prompts. Importantly, it selects augmented examples that are least similar to the input image, encouraging the model to generalize and better capture the semantic relationship between the image and its associated code. Experimental results indicate that our approach outperforms existing text-based prompting methods in both quantitative and qualitative evaluations. This work offers a practical and effective strategy for GUI code generation using LMMs and underscores the potential of multimodal prompting in layout generation tasks.
A Study on Genetic Algorithm-Based Optimization of Multi-Regional Weather Data for Solar Power Generation Forecasting
http://doi.org/10.5626/JOK.2025.52.8.688
The intermittency of solar power generation is one of the factors that make stable power supply challenging. To overcome the intermittency issue caused by weather conditions, various solar power generation forecasting methods have been developed using statistical modeling techniques and deep learning. However, existing methods have limitations in improving prediction accuracy as they fail to account for the correlations and spatial dynamics inherent in meteorological data. In this study, we proposed a solar power generation forecasting method that enhances prediction accuracy by leveraging a long short-term memory (LSTM) model to extract temporal patterns from meteorological data across multiple cities and optimizing the importance of each city's weather patterns using a genetic algorithm. We evaluated the performance of the proposed method in three major solar power generation regions in South Korea and confirmed that it outperformed existing methods while effectively identifying key regions closely related to solar power generation.
Energy-adaptive Data Loss Recovery in Energy Harvesting Wireless Sensor Networks
http://doi.org/10.5626/JOK.2025.52.8.700
Wireless sensor networks often face limited lifespans due to constrained battery capacity. Specifically, the multi-hop transmission mechanism causes energy depletion in nodes located near the sink, as data converges at these nodes, making them susceptible to energy exhaustion and transmission failures. To overcome the lifetime limitations of sensor nodes, energy harvesting techniques have been employed as a promising solution. However, harvested energy cannot be stored beyond the battery capacity, leading to surplus energy that goes unused. In this paper, we propose a novel scheme to address data loss by effectively utilizing this excess harvested energy. Our method retransmits lost data during transmission errors by leveraging surplus energy that exceeds the battery threshold. Additionally, if extra energy remains, the scheme performs redundant transmissions to enhance data reliability. Simulation results demonstrate that our proposed scheme successfully collects more data compared to conventional approaches, highlighting its effectiveness in recovering from transmission errors.
Enhancing Autonomous Driving Safety and Ride Comfort using Vehicle-to-Vehicle Communications
http://doi.org/10.5626/JOK.2025.52.8.709
Advancements in autonomous driving technologies- have led to ongoing efforts to enhance performance through vehicle-to-vehicle (V2V) communication. However, current V2V integration methods do not fully leverage new data types, such as brake and steering wheel data, and often replace existing sensor data with V2V data. This paper proposes two architectures designed to maximize the benefits of V2V communication for autonomous emergency braking (AEB) and adaptive cruise control (ACC) systems: V2V-empowered AEB with brake and steering wheel data (VEBS) and V2V-empowered ACC with brake and steering wheel data (VCBS). These architectures integrate sensor data with V2V communication, utilizing brake and steering wheel values obtained via V2V to enable more accurate predictions and faster responsiveness. Simulation results demonstrate that VEBS can avoid collisions and increase stopping distance by up to 243%, thereby enhancing safety. VCBS not only improves safety but also reduces average deceleration by up to 54%, contributing to enhanced ride comfort.
Search
Journal of KIISE
- ISSN : 2383-630X(Print)
- ISSN : 2383-6296(Electronic)
- KCI Accredited Journal
Editorial Office
- Tel. +82-2-588-9240
- Fax. +82-2-521-1352
- E-mail. chwoo@kiise.or.kr