Ground NLP
verb
Definition: To anchor a model’s output in externally provided context so that it uses the necessary information from that context and does not introduce unsupported content beyond it; in recent LLM research, a model is considered truly grounded when it fully uses the required knowledge from the provided context and stays within its limits [Lee et al. 2024]
Example in context: “Our framework tunes LLMs to self-ground the claims in their responses and provide accurate citations to retrieved documents.” [Ye et al. 2024]
Related terms: grounding, grounded generation, retrieval augmentation, factuality, citation grounding