Parametric memory recall evaluates what large language models remember about a brand natively without live web search, utilizing zero-shot prompts and token-level confidence scores to uncover unprompted brand associations and factual hallucination gaps.
flowchart TD
A[Ungrounded Prompt] --> B[LLM Parametric Inference]
B --> C[Token Logprobs & Output]
C --> D[Extracted Brand Facts]
E[(Canonical Brand Facts)] --> F[Discrepancy Analysis]
D --> F
F --> G[Confidence Scores]
F --> H[Hallucinations & Gaps]