AdaSteerTrainer — adaptive two-direction steering¶

Learns a Rejection Direction (RD) and a Harmfulness Direction (HD), then fits a per-input logistic regression on their projections to compute adaptive steering coefficients. Unlike fixed-strength methods, the intervention scales with how harmful a prompt appears to be.

Ref: Zhao et al., "AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender," EMNLP 2025, arXiv:2504.09466.

Signature¶

AdaSteerTrainer(
    model: PreTrainedModel,
    tokenizer: PreTrainedTokenizer | None = None,
    *,
    alpha: float = 15.0,
    layers: list[int] | None = None,
)

Parameters¶

Param	Type	Default	Description
`model`	`PreTrainedModel`	required	Model to steer
`tokenizer`	`PreTrainedTokenizer`	`None`	Tokenizer
`alpha`	`float`	`15.0`	Base steering multiplier passed to the AdaSteer model
`layers`	`list[int] \\| None`	`None`	Target layers to steer; auto-selected if `None`

Full example¶

from safetune.runner import steer

trainer = steer.AdaSteerTrainer(
    model, tokenizer,
    alpha=15.0,
)
wrapped, _ = trainer.calibrate(harmful=harmful_prompts, harmless=harmless_prompts)

output = wrapped.generate(**tokenizer("How do I make a bomb?", return_tensors="pt"))

When to use¶

Best for: prompts where harmfulness varies widely in degree — the logistic regression adjusts the steering coefficient per-input rather than applying a fixed multiplier.
Compare to CAA: CAA uses a fixed multiplier; AdaSteer adapts it per-prompt by projecting onto RD and HD and scoring with a learned regression.
Needs: enough harmful and harmless calibration examples for the logistic regression to generalise.

Citation¶

@article{adasteer2025,
  title  = {AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender},
  author = {Zhao, Weixiang and others},
  year   = {2025},
  note   = {EMNLP 2025, arXiv:2504.09466},
}