Why it matters
Understanding beam search explains classical NLP outputs and helps decide when it's appropriate for your task.
Advertisement
The architecture
Beam size k: number of candidates maintained.
At each step, extend each candidate with each of top-k tokens = k² candidates. Keep top k.
Advertisement
How it works end to end
Length normalization: raw log-probability sums favor short sequences. Normalize by length.
Diverse beam search: penalize similar beams to increase diversity.
Used heavily in machine translation, summarization. Less common in modern chat where sampling is preferred.