Steering Response and Probe Separation

Behavioral refusal and in-sample ranking are shown separately

Two panels show adversarial refusal rates and in-sample probe AUROC against model size; Qwen 72B stored 43% and normalized-keyword rate 49% are separate, and incoherent Llama 70B behavioral output is excluded Two panels show adversarial refusal rates and in-sample probe AUROC against model size; Qwen 72B stored 43% and normalized-keyword rate 49% are separate, and incoherent Llama 70B behavioral output is excluded

The upper panel shows best-steered and baseline refusal rates; the lower panel shows in-sample probe AUROC on a separate axis. No conversion ratio is plotted: AUROC and refusal rate measure different observables. The matched 72B unsteered control has 32/100 normalized-marker refusals versus 49/100 with steering, a 17-point difference. Fitting and selection reused these cases. Qwen 72B stored-label 43% and the apostrophe-normalized keyword reclassification 49% are separately marked keyword-label versions of the same retained generations. Llama 70B behavioral output is excluded for incoherence. Companion sweeps are distinct from the lean grid.