The integration of Large Language Models (LLMs) such as ChatGPT into academic research has garnered interest due to their potential to enhance efficiency in literature reviews and early-stage research. However, persistent challenges, including risks of "hallucinations," inaccuracies, and a lack of transparency, hinder their broader acceptance in academic contexts. To address these issues, this paper introduces the COMPACTION Framework (Comprehensive Organized Method for Prompt-based Analysis and Condensation of Textual Information into an Optimized Notation), a structured approach for utilizing LLMs in literature research.

Grounded in the Design Science Research (DSR) paradigm, the COMPACTION Framework integrates Large Language Model (LLM) processing with the IMRAD structure (Introduction, Methods, Results, and Discussion), a widely accepted format for scientific communication. The framework was operationalized through a prototype prompt that enables LLMs to analyze and summarize scientific articles into concise, structured, one-page summaries.

The methodology followed a three-phase approach. First, a detailed analysis of the IMRAD structure and related literature identified essential components for effective prompt design, focusing on key elements such as research gaps, methods, and results. Second, a prototype prompt was developed iteratively using the CLEAR guideline (Conciseness, Logic, Explicitness, Adaptability, Reflexivity). Iterative testing and refinement ensured the prompt’s adaptability to diverse article structures while maintaining clarity and precision. Third, the prototype was evaluated using peer-reviewed articles from top-tier journals. LLM-generated summaries were assessed for accuracy, coherence, and alignment with the original texts and compared to manually created summaries. The evaluation also addressed potential issues such as hallucinations or misinterpretations.

Key findings demonstrate that the COMPACTION Framework enhances transparency, integrity, and alignment of LLM outputs with academic standards. The iterative design process refined the framework to produce outputs that meet expectations of academic rigor while minimizing deviations from source content. The framework’s adaptability across disciplines and potential for further refinement with evolving LLM technologies underscore its utility and relevance.

While the COMPACTION Framework represents a step toward integrating LLMs into academic workflows, limitations remain, including difficulties in capturing the nuance of complex arguments and datasets, and the need for broader validation across diverse academic fields. Future research should extend the evaluation of the COMPACTION Framework to additional disciplines, focusing on its adaptability to complex academic contexts. Advancing prompt engineering techniques is essential, including dynamic prompts tailored to diverse fields and mechanisms to minimize hallucinations. As one of the most advanced AI techniques, Retrieval-Augmented Generation (RAG) offers the potential to provide reliable and up-to-date external knowledge. Integrating RAG into the framework could enhance its capacity to synthesize structured and unstructured information, bridging the gap between academic rigor and real-time, context-specific knowledge retrieval.

Keywords: Large Language Models, Prompt Engineering, IMRAD, Research Integrity, Artificial Intelligence in Research.