Grounding

The word has several senses. In cognitive science the symbol grounding problem (Harnad, 1990) asks how symbols acquire meaning through connection to perception; in vision-language research "visual grounding" means locating the image region a phrase refers to. In LLM applications it means source-grounded generation: the answer should be entailed by, and attributable to, text supplied at inference time rather than by the model's parametric memory. Rashkin et al. formalised this as "attributable to identified sources" (AIS): a statement is attributable if a generic reader, shown the source, would agree that the source supports it.

A grounded pipeline has three parts. Context construction assigns each retrieved passage a stable identifier and metadata (document, version, section, URL) and places them in the prompt, usually separated from instructions by delimiters. The instruction requires every claim to cite passage identifiers, forbids information not present in the passages, and specifies an abstention answer when coverage is insufficient. Verification then checks the output: parsing citations, confirming that cited identifiers exist, and testing support claim by claim, typically by decomposing the answer into atomic claims and running an entailment (NLI) model or an LLM judge against the cited passage. Some platforms expose this as a built-in "grounding check" or return per-sentence support scores; search-grounded APIs attach web results and citation spans to the response.

Evaluation distinguishes citation recall (is each claim supported by its citations?) from citation precision (is each citation needed and relevant?), the metrics used in the ALCE benchmark (Gao et al., 2023), and faithfulness from answer correctness. An answer can be perfectly faithful to an outdated or wrong source, and it can be correct yet ungrounded because the model answered from memory. Both matter, but only faithfulness is under the application's control once retrieval is fixed.

Common failure modes include citations that point to a real passage that does not actually say the claim, blending of two sources into a statement neither makes, silent reliance on parametric knowledge when retrieval returns nothing relevant, weaker use of passages placed in the middle of a long context, and conflicts between sources resolved arbitrarily. Grounding also does not make sources trustworthy: a retrieved document can carry prompt injection, so passages must be treated as data rather than instructions. NIST AI 600-1 lists confabulation among the risks specific to generative AI, and grounding with source display and human review of cited passages is one of the practical controls for it, not a guarantee.