What strikes me is the finding that controllability decreases with longer reasoning --- suggests CoT monitoring gets more reliable in complex, multi-step tasks

What strikes me is the finding that controllability decreases with longer reasoning --- suggests CoT monitoring gets more reliable in complex, multi-step tasks where scheming would be hardest to catch from outputs alone. The question is whether this holds as models get better at instruction following generally.