Effectiveness of an aligned source model alone

Determine whether an aligned source model, without a paired source model trained to pursue the unwanted goal, can provide an effective steering signal for value transplant.

Background

The cross-family experiment uses both an honest and a cheating Qwen3-8B source model to compute a value difference that steers a GPT-OSS-20B host. Because such paired source models may not exist for other goals, the authors leave unresolved whether an aligned source model by itself is sufficient.