Association probes query language models without live web search across three directions—E2B, B2E, and Q2B—to map what brands and concepts exist within their ungrounded parametric memory.

flowchart LR
    subgraph Probes[Association Probes]
        E[Entity or Topic] -->|E2B| B1[Brand Mentions]
        B[Brand Name] -->|B2E| E1[Associated Concepts]
        Q[Search Query] -->|Q2B| B2[Brand Mentions]
    end
    Probes --> M[(Parametric Memory)]

Referenced by