LLM: Hämtning kontra Parametriskt Minne

Large language models (\textit{LLM}) have made remarkable progress but still suffer from factual inconsistencies when it comes to general knowledge questions, often producing incorrect or unsupported information. This thesis addresses that limitation by comparing two paradigms for open-domain question answering: a retrieval-augmented generation (\textit{RAG}) system that supplements a small \textit{LLM} with an external document retriever, and a large standalone \textit{LLM} that relies solely on knowledge stored in its model weights. The central question is whether a much smaller retriever-backed model can achieve factual accuracy comparable to or better than a state-of-the-art large standalone \textit{LLM} in answering general knowledge questions. To investigate this, a \textit{RAG} pipeline was implemented using a relatively small \textit{LLM} (billions of parameters) for generation, and a large (millions of documents) retrieval corpus. Its performance on answer accuracy was then evaluated against leading large \textit{LLM}s (trillion+ parameters) on a curated dataset of 1,000 open-domain question \& answer (\textit{QA}) pairs. Each system’s rate of factually correct answers were measured and analyzed where their responses differ. This study thus explores the trade-off between “memory” (internal parametric knowledge) and “retrieval” (external knowledge lookup) in LLMs, aiming to understand how much external retrieval can compensate for a smaller model’s limited internal knowledge in terms of factual accuracy.