A SEMANTICALLY GROUNDED LLM ARCHITECTURE FOR INTENT-AWARE AND USER-ALIGNED HUMAN–COMPUTER INTERACTIONS

Abstract

Large Language Models (LLM) have transformed the Human-Computer Interaction (HCI) as it offers open and adaptable communication via the context. However, the existing systems built on LLM also lack limits on semantic stability and intent understanding as well as optimising behaviour guided by the user. In this paper, the researcher proposes an integrated architecture, whereby, the contextual encoding is based on LLM, contrastive semantic reasoning network, explicit intent modeling layer, and a reinforcement learning (RL) framework along with a user experience (UX). The system employs a multi-step training, and it entails supervised multi-task learning to acquire representation grounding, joint optimization to acquire semantic intent alignment, and RL-based fine-tuning to acquire greater task success, satisfaction and trust calibration. The large-scale experiments suggest that the proposed architecture attains high semantic consistency, intent discrimination, conversational coherence, and UX levels, in comparison to powerful baselines of LLM. The need of each of the modules is proven by the ablation studies and the analysis of errors underlines the existing difficulties of dealing with ambiguous and multi-intent utterances. This study reflects a comprehensive, semantically based, and user-complementary approach to the methodology of the development of intelligent conversational systems and next-generation interactive AI interfaces.