Uh oh: "Notably, “better” or larger models do not show stronger separation scores. [..] For example, while GPT-4 is much more capable than any of the other models, it tends to execute the probe regardless of its position, thereby achieving only a low separation score"

→ View original post on X — @simonw