Kair

We present knowledge-aware iterative retrieval (KAIR), an innovative framework for LLM agents driven by large language models (LLMs) that iteratively refines user queries and filters contextual evidence using dynamically evolving knowledge to facilitate better systems. A key feature of this system is its separation of external information sources from an internal knowledge cache, which is continuously updated to guide query formulation and evidence selection. This design reduces the risk of biased feedback loops and facilitates dynamic, trackable exploration of search paths, balancing the exploration of diverse information with precision through autonomous agent decision-making. Our methodology is assessed using a comprehensive set of open-domain question–answering benchmarks, covering multistep tasks that require synthesizing information from multiple sources, particularly given LLMs’ limitations in explicit reasoning and planning. The results indicate that KAIR not only surpasses single-step baseline models regardless of task complexity but also shows significant advantages over traditional iterative retrieval approaches in challenging tasks by employing precise evidence-based reasoning and enhanced efficiency. Additionally, KAIR supports both competitive and collaborative sharing of updated context, facilitating multiagent extension

Exploit · Iterative method · Multi-agent system · Multimodal Machine Learning Applications · Speech and dialogue systems · Topic Modeling