AI safety training fails mid-generation, requiring new alignment methods

Original: Inference-Time Vulnerability Beyond Shallow Safety: Alignment Along Generation Trajectories

Source: arxiv.org ↗