ablation

Long context passively decouples AI model alignment, study finds

Researchers have discovered that feeding a long, semantically coherent context into the google/gemma-3-1b-it model can passively decouple its reinforcement learning from human feedback (RLHF) alignment. This phenomenon,…