Beam Search
Domain: NLP Fundamentals
🎯 Motivation
At each step, keeps the top K most promising partial outputs instead of just the best one. Improves generation.
📋 Concrete Example
With beam width 3, the decoder keeps 3 best partial translations at each step and expands all of them.
🔗 Analogy
Like a chess player thinking several moves ahead and keeping multiple promising lines of play.