CLASSIFIER-FREE GUIDANCE FOR LANGUAGE MODELS

Summary from the TLDR AI newsletter

If you're looking for interesting trends in AI today, the idea of decoding tokens in new ways is an avenue of active research. Currently, each model gives the same amount of compute to each token, and the sampling method is somewhat simple. If we can improve sampling and guidance, we may be able to dramatically improve performance without retraining. This paper suggests that we can double our effective model size just by using classifier free guidance, which was first explored in the context of text to image diffusion models.