Hallucination
Fluent output that is not supported by the model's input or by fact.
Обзор
A hallucination is output that reads as confident and specific but is not grounded in the source or in fact. It solves nothing; it is the failure mode that makes a language model unsafe to use without checking. It exists as a named problem because it is not a bug that can be fixed by more training data: a model trained to predict plausible text has no separate mechanism for knowing what is true, which is why grounding and citation are engineering responses to it rather than model fixes.
Граф технологий
Как Hallucination связана с остальным стеком.
Применяется к
- TruthfulQA benchmark Benchmarks TruthfulQA used for Hallucination High ✓ source
- Grounding technique Techniques Grounding solves Hallucination High ✓ source
- RAG technique Techniques RAG solves Hallucination Medium ✓ source