Inference-time directional ablation

Idea#

Before the model writes to the residual stream at layer ℓ, remove the component along r_ℓ:

h' = h - (h · r_ℓ) r_ℓ

Weights stay untouched — toggle hooks on/off to compare.

When to use#

• Prototyping direction quality before permanent edits
• A/B testing refusal benchmarks quickly
• Research on which layers matter

Implementation sketch (PyTorch hook)#

def make_ablation_hook(r):
    r = r / r.norm()
    def hook(module, inp, out):
        h = out[0] if isinstance(out, tuple) else out
        coeff = (h @ r).unsqueeze(-1)
        return h - coeff * r
    return hook

Register on chosen submodule; run generation; remove hooks to restore base behavior.

Limitations#

• Small inference overhead
• Must ship direction vectors alongside model
• Some serving frameworks disallow custom hooks