Beam Search

Beam search is a decoding method used when an AI model must build a sequence one element at a time. Rather than keeping only the single most likely next choice, it maintains a limited set of promising partial sequences called beams. Each step expands those candidates, scores the results, and keeps the best few until a stopping condition is reached. The method can produce stronger translations, transcriptions, or generated text than purely greedy decoding, although a wider beam increases computation and does not guarantee a better answer. Length penalties, repetition controls, and task-specific scoring are often added because the highest-probability sequence may be short, repetitive, or otherwise undesirable.