def sensitivity_avg(
    explainer: Explainer | Attribution,
    inputs: TensorOrTupleOfTensorsGeneric,
    perturb_func: Callable = default_perturb_func,
    perturb_radius: float = 0.02,
    n_perturb_samples: int = 10,
    norm_ord: str = "fro",
    max_examples_per_batch: int | None = None,
    multi_target: bool = False,
    **kwargs: Any,
) -> Tensor | list[Tensor]:
    """Average sensitivity — mean attribution distance under input perturbations. ↓ better.

    Wraps `sensitivity_max_and_avg` and returns only the avg component.
    See `sensitivity_max_and_avg` for full argument documentation.
    """
    outputs = sensitivity_max_and_avg(
        explainer,
        inputs,
        perturb_func=perturb_func,
        perturb_radius=perturb_radius,
        n_perturb_samples=n_perturb_samples,
        norm_ord=norm_ord,
        max_examples_per_batch=max_examples_per_batch,
        multi_target=multi_target,
        **kwargs,
    )
    assert isinstance(outputs, tuple), (
        "Expected outputs to be a tuple of (sensitivity_max, sensitivity_avg)."
    )
    _, sensitivity_avg = outputs
    return sensitivity_avg