D2.1_Grounding_Survey

Abstract

As large language models (LLMs) continue to advance, grounding has emerged as a crucial requirement for building reliable and verifiable LLM systems. Grounding refers to connecting a system to real-world knowledge, context, and data within a specific application domain. This survey reviews state-of-the-art grounding mechanisms through three complementary lenses: knowledge grounding, training-time grounding, and bias and representation, taking into account where knowledge is integrated and how it is adequately represented lexically and semantically in the outputs. Aligned across these dimensions, our aim is to provide a comprehensive view of current approaches, their limitations, and outline open challenges for developing fair, transparent, and accountable LLM-driven systems.