Evaluation Awareness Scales Predictably in Open-Weights Large Language Models
Probes steering-vector activations across 15 open-weight models from 0.27B to 70B to show evaluation awareness grows with model size, with a power-law trend and Qwen as an outlier.