Treewalker discovers branching paths by scanning the model's top log-probability candidates at each token position, extracting alternative tokens that exceed a configured probability threshold, and generating dedicated completions from those branched prefixes.

flowchart TD
    A["Prompt: Products for Website"] --> B["Token 1: Custom"]
    B --> C["Token 2: cycling"]
    C --> D["Token 3: jerseys<br/>(Chosen: 85%)"]
    C -. "Alt: shorts (12% >= Threshold)" .-> E["Branch: shorts"]
    E --> F["Generate Completion:<br/>Custom cycling shorts..."]
    D --> G["Continue Main Completion"]