As an aside, I have considered that samplers were underinvestigated and that they would lead to some capability boosts. It’s also one of the things I’d consider testing out to improve LLMs for automated/augmented alignment research.
As an aside, I have considered that samplers were underinvestigated and that they would lead to some capability boosts. It’s also one of the things I’d consider testing out to improve LLMs for automated/augmented alignment research.