Concept
Beam Search
Decoding that keeps the top-N highest-probability partial sequences at each step.
Definition
Beam search expands several candidate continuations in parallel, pruning all but the N best at every step. It produces higher-likelihood outputs than greedy decoding but is rarely used for open-ended chat because it tends to generate bland, repetitive text.
Common use cases
- Translation
- Summarisation
- Constrained decoding