KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/kmitl/1

Browse

Settings

Subject: search.filters.subject.Large Language Models

Search Results

Now showing 1 - 4 of 4
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Assessing the Effects of Corrupted Parameters in a Large Language Model: A Case Study of LLAMA 3.2 1B
    (2026-01-21)
    Chawuthai, Rathachai
    ;
    Thongsawaeng, Anon
    ;
    Perdio, John Paul Layug
    ;
    Zaw, Kaung Khant
    ;
    Kraichoke, Phalat
    This study explores the effects of parameter corruption in a large language model (LLM) by altering its weights and evaluating performance. Experiments involve corrupting different layers and matrix types, including Self-Attention and Feed-Forward components, with performance assessed using BERT and ROUGE scores. Testing the Llama-3.2-1B-Instruct model was performed on the GLUE-QNLI dataset. Results show that increased corruption leads to greater degradation, with Feed-Forward matrices having a stronger impact especially in the Down matrices. According to the study, later layers are more important for performance than those that come before them. These results shed light on possible future chip implementations of LLM, which may help guide the design of fault-tolerant systems by taking vulnerable parameter placement into account.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Benchmarking Next-Generation Frontier Models for Document-Level Arabic-Thai Medical Translation: A Reliability Study of LLM-as-a-Judge
    (2026-01-01)
    Lertsuksakda, Rathawut
    ;
    Netisopakul, Ponrudee
    ;
    Boonsang, Siridech
    This research targets Arabic-to-Thai medical translation, a low-resource and linguistically distant pair critical for public health. We utilize the OpenWHO dataset, a resource comprising 293 documents. This dataset is shielded from webcrawling to minimize training data contamination and ensure rigorous zero-shot evaluation. We conduct a comparative analysis between 2026-era frontier models including GPT-5.2, Gemini 3 Pro, and Claude Opus 4.5 and industry-standard NMT represented by Google Translate. Using an LLM-as-a-Judge framework with an AI jury of efficient reasoning models such as GPT-5.1, Gemini 3 Flash, and Claude Sonnet 4.5, we performed 3,516 evaluations of fidelity, fluency, and cultural appropriateness. Results demonstrate that frontier models outperform traditional NMT across all dimensions. While high exact agreement suggests LLM-as-a-Judge frameworks are promising for scalable evaluation, human validation reveals persistent opportunities for improving AI-human alignment, necessitating targeted oversight for safety-critical applications.
      1
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Application of Large Language Models for Aspect-Based Sentiment Analysis on Social Media Data: A Case Study of the Thai Telecommunications Industry
    (2026-01-01)
    Limseesawan, Krittapas
    ;
    Netisopakul, Ponrudee
    ;
    Chotipant, Supannada
    ;
    Voravuthikunchai, Winn
    ;
    Sirivorachodphokin, Thitirat
    Social media is a valuable source of consumer opinion data, particularly in Thailand's highly competitive telecommunications market among AIS, TRUE, and DTAC. This study compares three LLM-based approaches for Aspect-Based Sentiment Analysis (ABSA) on 3,508 Thai-language messages from Pantip.com and YouTube.com: (1) Prompt Engineering, (2) Domain Adaptation & Fine-Tuning, and (3) Multi-Agent Debate Framework. Results show that Domain Adaptation & Fine-Tuning achieves the best accuracy-latency trade-off, with Qwen2.5-1.5B-Instruct exceeding 87% average accuracy in under 4 seconds per message. Multi-Agent Debate achieves the highest Category accuracy (82.82%) but at a latency cost of 10-16 seconds per message. Gemini-2.5-Flash-Lite provides the best speed-accuracy balance for Prompt Engineering without additional training. Critically, small open-source models (1B-1.5B parameters) subjected to domain-specific fine-tuning can approach or surpass proprietary large models on this task, suggesting domain alignment may outweigh raw parameter scale for narrow, well-defined ABSA tasks.
      1
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Context-Aware Prompting for Japanese-Thai Literary Translation in a Low-Resource Setting
    (2026-01-01)
    Sikkhamarn, Korawit
    ;
    Vittayakorn, Sirion
    Japanese-Thai literary translation remains underexplored, particularly in low-resource settings where parallel data and evaluation resources are limited. This paper investigates whether context-aware prompting can improve Japanese-Thai literary translation using large language models. To support this study, we construct JTLitCorpus-80, a Japanese-Thai literary parallel corpus derived from 80 publicly accessible online translated web novels. Using this corpus, we compare eight prompt-based translation settings, including machine translation, post-editing, and context-aware prompting variants with and without dictionary guidance. Results on a 20-title test split show that direct context-aware prompting consistently outperforms machine translation and non-context-aware baselines, while post-editing provides only limited gains once the initial draft is already generated by the same model. We further show that prompt-derived supervision can improve a smaller open model through parameter-efficient adaptation. MQM-based human evaluation supports the automatic-metric trends, indicating that the strongest context-aware setting provides the best overall translation quality. These findings suggest that context-aware prompting is a practical strategy for Japanese-Thai literary translation in low-resource conditions.