Beam Search
Beam search is a heuristic decoder that keeps the k best partial sequences at each step instead of greedily taking only the top token — the classic seq2seq inference method.
Definition
At each time, expand every beam prefix by the vocabulary, score (usually log-prob, sometimes length-normalized), and keep the top k. Greedy is k = 1. Sampling (top-p, temperature) is a different family used by chat models.
Why it matters
Translation and many structured generators still use it. Chat LMs usually sample because beam search is bland and repetitive on open-ended text.
Case
A beam-10 NMT decode lifts BLEU versus greedy. The same beam on a joke prompt produces the safest joke, not the funniest.
How to read it
Report k, length penalty, and whether you blocked repeats. Comparing a sampled chat model to a beam translator is not an eval.