TRUST-LLM: A Modular Trustworthy Architecture for Reliable Large Language Model Systems

TRUST-LLM is a trust-centric reference architecture for Large Language Models (LLMs) that integrates intent understanding, prompt risk assessment, governance, retrieval, reasoning verification, hallucination detection, confidence estimation, explainability, runtime monitoring, and a composite trust metric into a unified modular framework. The paper formally specifies a 12-module architecture, introduces a mathematically defined trust score, presents algorithms, deployment architecture, threat model, and evaluation methodology, and provides a reference design for trustworthy LLM systems intended for future empirical validation.