Safety Alignment Should be Made More Than Just a Few Tokens Deep

Investigates safety alignment concentrated in the first output tokens. Case studies and regularized fine-tuning objectives explore deeper alignment and its persistence under subsequent fine-tuning.

Investigates safety alignment concentrated in the first output tokens. Case studies and regularized fine-tuning objectives explore deeper alignment and its persistence under subsequent fine-tuning.

When reading the original paper, identify its problem, inputs, outputs, and methodological assumptions.

Investigates safety alignment concentrated in the first output tokens. Case studies and regularized fine-tuning objectives explore deeper alignment and its persistence under subsequent fine-tuning.

Check the baselines, datasets, evaluation protocol, and ablations in the original paper. This directory overview does not provide an independent evaluation or experimental results table.