Efficient Summarization with Lightweight LLMs through Sparse Input Activation and Adaptive Prompting.
Disciplines
Other Computer Sciences
Abstract (300 words maximum)
Large Language Models are powerful tools that are designed to read, reason and generate in human language. They help us by making knowledge, problem-solving and communication easier through natural language understanding and generation. However, their performance heavily depends on the effectiveness of the prompt given and the size of the model. Light-weight LLMs(sub-10B parameters) struggle with summarizing complex technical documents like research publications due to technical jargon, lengthy citations, and mathematical notations, often leading to hallucinations that reduce reliability and accuracy.
Our project introduces the hybrid framework that combines NLP driven pre-processing techniques, sparse input activation and prompt engineering to enhance the summarization capacity of the small-scale models by minimizing hallucinations and maximizing factual accuracy. The pipeline starts with cleaning and segmenting full articles into sections, removing references, citations, and formulas. On this normalized output, salient sentence extraction and keyphrase extraction is performed as part of the sparse input activation module. The activated input and optimized prompt is fed to LLMs of different scales and the summary is benchmarked with respect to different prompting strategies and scale of the models.
This workflow significantly enhances the factual accuracy and reliability of summaries generated by Light-weight LLMs, making them competitive for complex scientific and technical summarization tasks.
Use of AI Disclaimer
no
Academic department under which the project should be listed
CCSE – Computer Science
Primary Investigator (PI) Name
Md Abdullah Al Hafiz Khan
Efficient Summarization with Lightweight LLMs through Sparse Input Activation and Adaptive Prompting.
Large Language Models are powerful tools that are designed to read, reason and generate in human language. They help us by making knowledge, problem-solving and communication easier through natural language understanding and generation. However, their performance heavily depends on the effectiveness of the prompt given and the size of the model. Light-weight LLMs(sub-10B parameters) struggle with summarizing complex technical documents like research publications due to technical jargon, lengthy citations, and mathematical notations, often leading to hallucinations that reduce reliability and accuracy.
Our project introduces the hybrid framework that combines NLP driven pre-processing techniques, sparse input activation and prompt engineering to enhance the summarization capacity of the small-scale models by minimizing hallucinations and maximizing factual accuracy. The pipeline starts with cleaning and segmenting full articles into sections, removing references, citations, and formulas. On this normalized output, salient sentence extraction and keyphrase extraction is performed as part of the sparse input activation module. The activated input and optimized prompt is fed to LLMs of different scales and the summary is benchmarked with respect to different prompting strategies and scale of the models.
This workflow significantly enhances the factual accuracy and reliability of summaries generated by Light-weight LLMs, making them competitive for complex scientific and technical summarization tasks.