Context-Induced Activation Drift: Long benign context passively decouples RLHF alignment without adversarial prompts (Mechanistic Interpretability + Ablation) [D]

TL;DR: We observed that feeding a long, benign, thematically coherent context prefix ($L \in [100, 3000]$ tokens) into google/gemma-3-1b-it causes a massive passive shift in internal activations ($\Delta h_2 \approx 3434$) at deep layers ($\sim 85%$ depth). This leads to a logit decoupling ($D_{KL}