Does it support beam search generation strategy?

[colinzhaoxp]

I have a question about beam search as the title described. In other words, Does LLaDA support Beam search, Top-K, Top-p generation strategy that supported by AR models ?

[nieshenx]

LLaDA supports it, but the details are different from those of ARMs. In our code, you can see the option to set the temperature, which you can replace with top-p or TOP-K accordingly.

[colinzhaoxp]

thanks for your replay!

As for the beam search, do you mean this line of code? https://github.com/ML-GSAI/LLaDA/blob/4aa0dd2402fb9fec1137648cf768b56103a4a849/generate.py#L84

I also find that LLaDA supports two remasking strategies: low_confidence and random. If I want to support beam search, do i need to change them? https://github.com/ML-GSAI/LLaDA/blob/4aa0dd2402fb9fec1137648cf768b56103a4a849/generate.py#L87-L94