Techniques & Methods
Beam Search
Beam search explores multiple token sequences simultaneously, keeping the top-k (beam width) most probable partial sequences at each step. Unlike greedy decoding which commits to one token at a time, beam search can recover from locally suboptimal choices by considering alternatives.
Larger beam widths improve sequence quality but increase computation linearly. Beam search is widely used in machine translation and summarization but can produce generic, repetitive outputs for open-ended generation where diversity is preferred.
Authority Links
Related Terms
Techniques & Methods
Greedy Algorithms
Algorithms that make the locally optimal choice at each step to find a global solution.
Techniques & Methods
Decoding Rules
Guidelines and algorithms that control how language models translate internal representations into output tokens.
Techniques & Methods
Sequence Generation
Process where models produce sequences—such as words or tokens—based on learned patterns.
Techniques & Methods
Machine Translation
Software that automatically translates text or speech between languages.