Tag
This paper formalizes Streaming Knowledge Compilation for LLM wikis, introducing a materiality signal to proactively pin important documents from a streaming corpus under a token budget. It proves an O(√(T log K)) regret bound and validates the approach in finance and Wikipedia domains, showing that regret analysis is a reliable evaluation metric.
The paper introduces WiCER, an iterative algorithm for compiling domain knowledge into LLM Wiki systems to minimize information loss and catastrophic failure rates during knowledge distillation. It demonstrates that this approach improves upon full-context KV cache inference by preserving critical facts better than blind compilation methods.
The paper introduces Cognitive Agent Compilation (CAC), a framework that uses teacher LLMs to compile problem-solving knowledge into explicit, inspectable agents for educational applications. It aims to address the lack of controllability and explainability in standard LLMs by separating knowledge representation from policy and verification rules.