Beam Search
A heuristic sequence-search algorithm that keeps only a fixed number of highest-scoring partial hypotheses at each decoding step.
Search-Algorithm Context
Beam search keeps only a fixed number of high-scoring partial sequence hypotheses at each decoding step. Increasing beam width explores more alternatives but also raises latency and memory cost; length normalization and language-model scores can materially change which path survives.
Optimality Boundary
Beam search does not guarantee the global optimum. It explores more possibilities than greedy decoding but prunes aggressively compared with exhaustive sequence search.