Streaming Knowledge Compilation: Proactive Materiality-Scored Pinning for Time-Evolving LLM Wikis

The paper presents a novel approach called Streaming Knowledge Compilation, which enables the maintenance of a dynamic knowledge base for LLM wikis that adapts to evolving information landscapes. It introduces a materiality scoring system to proactively pin relevant documents before queries are made, achieving an $O(\sqrt{T\log K})$ regret bound, with empirical results demonstrating significant performance improvements in finance and Wikipedia domains. This method is crucial for practitioners as it enhances the efficiency of LLMs in handling real-time data changes, ensuring more accurate and contextually relevant responses.