Technology

beam search

Beam search is a heuristic graph search algorithm: it efficiently prunes the search space by retaining only the top *k* (beam width) most promising partial solutions at each step.

Beam search is the go-to heuristic for sequence generation tasks, balancing speed and solution quality. It operates as a pruned breadth-first search: at each level, it expands all successor nodes but only keeps a fixed number, *k* (the beam width), of the highest-scoring candidates. This process drastically reduces memory and time complexity, making it essential for computationally intensive systems like Neural Machine Translation (NMT) and large-scale speech recognition. While it sacrifices guaranteed optimality—unlike an exhaustive search—it consistently delivers high-probability, near-optimal sequences, proving its value in real-world AI applications.