Nucleus sampling with beam?

[lenbrocki]

I am a bit confused about how exactly the inference works when I choose nucleus sampling, in particular why there is a beam size as parameter. My understanding was that nucleus sampling does not do beam search, but instead it is a version of random sampling(so randomly choose on each decoding step). Does the implementation of nucleus sampling in ParlAI somehow combine this with beam search?

[klshuster]

When using sampling, you can think of the beam_size parameter as more of a best_of_n parameter; if you specify e.g. --beam-size 5 with nucleus sampling, ParlAI will sample 5 generations in parallel and output the one with the highest score at the end.

[lenbrocki]

I see, thanks for the answer!