One direction, two effects: a target-and-control audit of abliteration's side effects
Year: 2026
Type: preprint
Abstract: Abliteration removes a language model's refusal direction from its weights, and the resulting uncensored checkpoints are widely run locally. A 2026 study reported that the surgery also shifts disposition: abliterated models become more optimistic, write longer justifications and use fewer uncertainty words. The claim is repeated as a property of the technique, and no study of abliteration has tested whether the change comes from removing the refusal direction or from moving the weights. I ran that control on Qwen3-4B, Qwen3-8B and Llama-3.1-8B. For each model I built the aligned baseline, a Heretic abliteration, and three random orthogonal edits with per-matrix Frobenius norms matched to the abliteration at 0.5, 1.0 and 1.5 times its size. Refusal moves only with the directional edit: the abliteration takes it from 94-100 per 100 to 0-2, while every random edit leaves it at 94-99. The side effects decompose into a generic component that scales with edit size and a direction-specific component. The specific component is negative for confidence on all three models (-0.24, -0.15, -0.12, every interval excluding zero), so abliteration makes a model state higher confidence and that is not an artifact of weight movement. The specific component for verbosity changes sign between models (-0.35, +0.39, -0.59), which explains why verbosity claims conflict in the literature. Abliteration is effectively rank-one per matrix: the leading component carries 98% of the edit norm, removes refusals on its own, and carries both side effects, while the residual tail moves nothing measurable.
Author: Peddy, Jackson
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: verbosity, Qwen3, Llama 3.1, heretics
Open Access status: green
OpenAlex ID: W7227080095
