Abliteration

Editing a model’s weights to suppress its refusal behaviour, rather than retraining it or prompting around it. The name is a portmanteau of ablate + obliterate. It presumes that refusal is carried by an identifiable direction in activation space — find it, remove it from the weights, and the model stops declining.

The technique only exists where weights are downloadable, which makes it a property of the open-weight tier specifically. A closed model’s refusals can be routed around (claude-refusals-and-fallback) but not edited out.

The trade it is named for

The standard complaint about abliteration is collateral damage: removing the direction wholesale costs coherence and general capability, because the direction was never carrying refusal alone. Every refinement of the method is an attempt to pay less of that cost.

scotoma is one such attempt — a Jacobian-lens projection that keeps ~22% of the abliteration magnitude and re-applies it at 1.5× across a layer band, on the theory that the discarded 78% was never refusal. Whether that works is unverified; the model ships with no evals.

The interpretability seam

Abliteration is downstream of mechanistic interpretability: it is what you do once you believe you have located a behaviour inside the network. The instruments are shared — the scotoma edit uses Anthropic’s Jacobian lens, a technique built for auditing models, to modify one. A tool for finding out what a model is doing is also a tool for changing it, and the second use needs no permission from whoever built the first.

Gaps

The corpus has no primary source on the technique itself. Everything above traces to a single derivative model card (gemma-4-31b-it-scotoma, T3, self-reported) — the refusal-direction literature it rests on is unread here. Treat the mechanism as sketched, not established.

Related

gemma-4-31b-it-scotoma · open-weight-models · claude-refusals-and-fallback · readyart · gemma-4