Training-stage emergence of Protection Mode

Determine when Protection Mode becomes observable during the training of large language models.

Background

The paper identifies Protection Mode as the tendency to prioritize a low-risk person or animal over an object facing a greater immediate physical consequence. Its experiments show that models can exhibit this behavior either because they misjudge relative risks or because they state the correct risk ordering but nevertheless choose the living subject. Interventions further indicate that protection-priority reasoning can causally influence the final choice.

Although the paper compares pretrained Qwen3.5 Base models with post-trained counterparts and finds that Protection Mode is already observable before post-training, the authors do not establish at what point during training the behavior first emerges. The unresolved issue is therefore to locate or characterize the training stage at which Protection Mode becomes detectable.