KARLA Enhances LLM Factual Accuracy with Knowledge Base Retrieval

A novel method named KARLA (Knowledge-base Augmented Retrieval for Language Models) has been introduced to significantly enhance the factual accuracy and transparency of Large Language Models (LLMs). This approach enables an LLM to dynamically pull in factual information from an external knowledge base as it generates tokens. The core innovation lies in training the LLM to produce special tokens that act as triggers for queries to the knowledge base. This mechanism offers several key advantages: it allows factual knowledge within the LLM's output to be updated simply by modifying the knowledge base, eliminating the need for costly LLM retraining. Furthermore, it provides clear traceability, linking generated facts back to their source in the knowledge base, thereby improving explainability and transparency. Crucially, KARLA also demonstrates that smaller LLMs can achieve factual accuracy comparable to much larger models when augmented with this retrieval capability. Experimental results confirm that the method improves factual grounding in both short-form and long-form content generation.