loading
loading
A decoding strategy that explores several candidate sequences at once and keeps the most probable.
Instead of greedily picking the single best next token, beam search tracks the top-k partial sequences ('beams') and expands them in parallel, returning whichever full sequence has the highest overall probability. It was standard in older translation and summarization systems where there's one 'correct' output. For modern open-ended chat it's mostly abandoned — it produces bland, repetitive text and costs more compute, so sampling with temperature won. Know it mainly to understand why nobody uses it for LLMs anymore.
Plainly
Think of Beam Search as a simple recipe for doing the work better. A decoding strategy that explores several candidate sequences at once and keeps the most probable.
In practice
Use it when you need a repeatable method instead of guessing from vibes. In practice, define the owner, input, output, and failure mode before you rely on it.