demo

Three ways to commit to a sequence

Greedy is fastest but myopic. Beam search keeps K hypotheses alive in parallel. Sampling adds creativity. See the candidate paths each strategy produces.