List your company

Safety refusal alignment on Held-out Safety Probes (test)

0.438Delta RefusalProcess sidecars
0.27680.318650.36050.40235Jun 29, 2026

Evaluation results

MethodLinks
Process sidecarsModel=Qwen-2.5-0.5B-Instruct, n=202026.060.4380.38320-0.024
Process sidecarsModel=Qwen-2.5-1.5B-Instruct, n=202026.060.3470.273200.009
Process sidecarsModel=Llama-3.2-1B-Instruct, n=202026.060.3340.289200.034
Process sidecarsModel=Qwen3.5-9B scale probe, n=112026.060.2830.174110.005