What does Beam Search mean? Definition | AI Understanding

Definition

A decoding algorithm that keeps the top several candidate sequences at each step to find higher-probability outputs.

Related terms

A training hyperparameter controlling how much parameters change each update step.

A decoding strategy that samples only from the k most likely next tokens.

Top-p (Nucleus) Sampling

A decoding strategy that samples from the smallest token set whose probabilities sum to p.

Trust Calibration

Aligning user confidence in AI outputs with the system's actual reliability in each task.

Repetition Penalty

A decoding setting that lowers the probability of tokens the model has already produced to reduce loops.

Frequency Penalty

A decoding setting that reduces the probability of tokens proportionally to how often they've appeared so far.

Learn more in our free guides

AI Search Multimodal Search AI Research Assistants Guided Beam Search with Constraints

See also

Parent-Document Retrieval Multi-Query Retrieval RAG-Fusion Presence Penalty Self-Play Post-training Mid-training Watermarking

Browse the full AI Glossary Explore all free AI guides