L’ancrage factuel est la propriété d’une sortie de modèle d’être soutenue par un matériel source spécifique ou un contexte fourni. En IA générative, il aide à évaluer si une réponse est traçable à des preuves plutôt qu’à une génération non étayée du modèle.
A property of a model whose output is based on (is "grounded on") specific source material. For example, suppose you provide an entire physics textbook as input ("context") to a large language model. Then, you prompt that large language model with a physics question. If the model's response reflects information in that textbook, then that model is grounded on that textbook. Note that a grounded model is not always a factual model. For example, the input physics textbook could contain mistakes.
A side-by-side comparison of Grounding and Groundedness. Understand the difference between the process of connecting outputs to sources and the property of an output being supported by evidence.
A side-by-side comparison of Factuality and Groundedness. Understand how the terms differ, when each applies, and what the distinction means for AI governance, system design, or assurance evidence.