Tree-walking analyzes the model's logprobs distribution at each token step to identify suppressed alternative tokens above a set probability threshold, then forces completion of those alternate paths to discover latent brand associations that were outranked in the primary output.
flowchart LR
P[Prompt] --> T1[Token Step 1<br/>Logprob Analysis]
T1 -->|Top Prob| B1[custom]
T1 -->|Above Threshold| B2[sublimation]
T1 -->|Above Threshold| B3[team]
B1 --> C1[jerseys]
B2 --> C2[uniforms]
B3 --> C3[apparel]
C1 --> O1[Primary Output]
C2 --> O2[Latent Association A]
C3 --> O3[Latent Association B]Referenced by