It is now common knowledge in the NLP community that beam search only works for situations where the output space is very constrained, specifically, in neural m

It is now common knowledge in the NLP community that beam search only works for situations where the output space is very constrained, specifically, in neural machine translation. In more open ended generation such as summarization, question answering and story generation, beam search leads to poor and repetitive outputs. Different (stochastic) sampling methods lead to more interesting, diverse and .. functional outputs.