Definition
Plain language
How hard you push on the model's internals when nudging it toward a behavior.
As stated in the literature
Scaling coefficient applied to an injected direction; behavior typically varies monotonically with it up to a breakdown point where generation degrades into repetition or incoherence.
Also called: steering strength, injection strength
Why it matters: Reporting intensity matters because results at gentle settings and results near the breakdown point mean very different things, and the strongest-looking effects are often just incoherence.
For example, a small push might make a model's answers slightly gloomier, a bigger push might make it refuse the task, and too big a push makes it repeat the same phrase forever.
Heard on the show
“They commit to using the lowest steering intensity that produces a measurable response.”Episode 267 — A Pain Axis, a Relief Button, and the Control the Paper Skipped