Knowledge and Aptitude Augmented Generation: Adaptive Multi-Turn Interaction in LLM Systems

Artificial intelligence (AI) systems have significantly advanced in handling knowledge-intensive tasks. However, a persistent gap remains between AI interactions and human behavior due to the static nature of most AI models, which lack the ability to adapt dynamically to evolving contexts. This paper introduces the Knowledge and Aptitude Augmented Generation (KAAG) framework, inspired by Dynamic Bayesian Networks (DBNs) used in robotics. By integrating knowledge and aptitude, and utilizing dynamic instruction generation, KAAG enables real-time adaptation to interaction states, improving the alignment between AI and users in multi-turn conversations. The framework includes a Gamified Interaction Model (GIM) to guide interactions toward convergence, with experiments showing enhanced Controlled Steerability compared to standard LLMs and Retrieval-Augmented Generation (RAG) systems.