Beam Search
A decoding algorithm commonly used in sequence generation tasks. At each time step, it maintains the K highest-probability candidate sequences, where K is the beam width. Compared with greedy decoding, beam search explores a broader candidate space and generally improves output quality.
Timeline (last 90 days)
Beam Search maintains k candidate sequences with the highest scores at each decoding step, striking a balance between greedy decoding and exhaustive search, and can avoid local optima traps.
For translation tasks, beam search remains the preferred choice, while Top-p is more suitable for creative generation scenarios.
Beam Search maintains the K highest-scoring candidate sequences at each time step; K is called the beam width, and a larger beam width typically improves quality but increases computational overhead linearly.
In practice, setting the beam width to between 4 and 10 is a relatively common choice.
All Facts (4)
Beam Search maintains k candidate sequences with the highest scores at each decoding step, striking a balance between greedy decoding and exhaustive search, and can avoid local optima traps.
50%UnverifiedFor translation tasks, beam search remains the preferred choice, while Top-p is more suitable for creative generation scenarios.
50%UnverifiedBeam Search maintains the K highest-scoring candidate sequences at each time step; K is called the beam width, and a larger beam width typically improves quality but increases computational overhead linearly.
50%UnverifiedIn practice, setting the beam width to between 4 and 10 is a relatively common choice.
50%