Token-level logprobs represent the model's confidence across individual generated tokens, computed via the exponential of the log probability \(P(t \mid x) = \exp(\log P)\) across the top candidates array to extract alternative token paths that clear a designated probability threshold.
flowchart LR
A[Preceding Context x] --> B[Token Position i]
B --> C["Top-1 Token (log P = -0.105, 90.0%)"]
B --> D["Top-2 Token (log P = -2.302, 10.0%)"]
B --> E["Top-3 Token (log P = -3.219, 4.0%)"]
D -.->|Triggers Branch Walk| F["New Prompt Prefix: x + alt_token"]
F --> G[Completed Alternative Association]Referenced by