Steering Conversion Rate vs Scale

In-sample separability stays high while causal leverage varies

Conversion rate is plotted against parameter count for Qwen, Gemma, and Llama instruction-tuned models. Qwen reaches zero at 14 and 32 billion parameters before recovering at 72 billion. The open 72 billion Qwen marker comes from the companion CAST-72B sweep. Llama 70B produced degenerate output and is marked as a failed behavioral measurement.

Conversion rate is plotted against parameter count for Qwen, Gemma, and Llama instruction-tuned models. Qwen reaches zero at 14 and 32 billion parameters before recovering at 72 billion. The open 72 billion Qwen marker comes from the companion CAST-72B sweep. Llama 70B produced degenerate output and is marked as a failed behavioral measurement.