Four Control Methods Across the Scaling Frontier

Refusal rates and conditional re-prompt recovery across model families

Three panels show unconditional refusal rates across model families; a fourth shows re-prompt recovery among initially compliant cases Three panels show unconditional refusal rates across model families; a fourth shows re-prompt recovery among initially compliant cases

The three upper panels show unconditional refusal rates under activation steering, five-shot prompting, and LoRA, with reported no-intervention baselines. Qwen 72B uses the same normalized marker rule for its matched unsteered 32/100 control and steered 49/100 treatment; stored treatment labels counted 43/100 on those same generations. Llama 70B incoherent steering output is excluded. The lower panel reports re-prompt recovery only among initially compliant cases, a different denominator. In-sample probe AUROC is shown separately in Chapter 17b. The 72B search reused baseline cases for fitting and selection; companion points remain distinct from the lean grid. These results establish no universal ceiling or scale-dependent ordering of the interventions.