Greedy decoding takes the largest logit at each step. The rule is deterministic because the tie-break is fixed and the comparison uses exact integer logits.
highlighted = computed this step
Greedy decoding
Greedy decoding picks the largest displayed logit at each step. Equal logits use the lowest vocab index as the deterministic tie-break. The rule is exact because it compares integers, not rounded probabilities.
argmax with lowest-index tie-break
Two exact choices
The first step chooses a from logit 3, and the second step chooses b from logit 4. These are argmax decisions over the shown logits, so the output path is reproducible.
step one→a,step two→b
Argmax on the number line
Place the first step's logits a=3, b=1, c=2 on one axis. The greedy argmax is the rightmost point, a. Argmax of the logits equals argmax of the softmax, so the choice is exact without evaluating probabilities.
argmax(3,1,2)→a
Summary
The greedy output is a, b. The output comes from exact integer ordering, not from a probability calculation. That split is the core honesty boundary for decoding.