Constitutional adapters are low-rank adapters or steering vectors trained on synthetic corpora that follow a written constitution, then reduced by subtracting a matched control object trained on value-free content. They are trained without any harmful requests or jailbreaks in the training data.
The adapters improve jailbreak defense and measured alignment, with the largest gains at long context lengths and against multi-turn attacks, where they beat prompted and steered baselines. They transfer zero-shot from a base model to its instruct checkpoint and can be scaled at inference time to trade defense against benign compliance. This makes them a lightweight, portable, tunable option for API deployments.