Treewalker evaluates token-level log probabilities from an initial model response, identifies alternative candidate tokens that exceed a probability threshold, and runs completion prompts on those fragments to discover latent associations in parametric memory.

flowchart TD
    Root["Prompt: Products associated with brand"] --> Primary["Primary Path: Custom jerseys (85%)"]
    Root -.-> Branch["Alternative Token: Cycling... (15%)"]
    Branch --> Comp["Completion: Cycling windbreaker jackets"]