Can Frontier Models Autocomplete Safety Research?

We pose the following research question: how can we measure the "research taste" of language models in experiment planning? What parts of planning taste remain intrinsic to humans? TL;DR. A future where “tasteless autoresearch” improves capabilities but not safety is plausible and dangerous. We need rough tests of tasteful planning to see what is missing.We can start by masking part of a paper, sampling extensions from a language model, and comparing them against the masked experiments. This pro...