LLM Categorization Questions: Levels of Groundedness in Questions for Large Language Models

Miquel Noguer i Alonso
DOI: https://doi.org/10.2139/ssrn.4798403
2024-01-01
SSRN Electronic Journal
Abstract:This paper explores the nuanced concept of groundedness in questions posed to large language models (LLMs) and provides a detailed classification of these questions, ranging from fully grounded to context-dependent. Groundedness in this context refers to the extent to which the questions are based on verifiable facts, hypothetical scenarios, or require specific contextual information for accurate responses. By categorizing questions into distinct levels—fully grounded, partially grounded, hypothetical or speculative, opinion-based, and context-dependent—this study highlights the diverse strategies required to tailor responses appropriately to each type of inquiry. This classification not only aids in enhancing the response accuracy of LLMs but also serves as a critical framework for training and refining AI systems to better interpret and interact with human users. Through a series of examples and analyses, the document examines the challenges and methodologies associated with processing each category of question, emphasizing the importance of precise model training and adaptive response generation. Ultimately, this work contributes to the broader field of AI by providing insights that can improve the practical applications of language models in various real-world settings, from educational tools to customer service and beyond, ensuring that LLMs are more responsive and effective in their interactions
What problem does this paper attempt to address?