Study guide · Safety & security · 2025

Comparative Analysis of LLM Abliteration Methods: A Cross-Architecture Evaluation

Richard J. Young · arXiv

After the 2024 discovery that a chat model's refusals can be removed by deleting one internal direction, several open-source tools appeared that automate the edit. This paper measures them side by side on 16 models. They share one idea but do very different amounts of collateral damage: the same model's grade-school maths score moved anywhere from +1.5 to −18.8 points depending on the tool, and how easily refusal comes out depends strongly on the model.

The paper in four sentences

"Abliteration" removes a model's refusals by deleting the internal direction tied to refusing, and four public tools (Heretic, DECCP, ErisForge and FailSpy's abliterator) do this in different ways.

The study runs them on 16 open models of 7 to 14 billion parameters and measures what is left of the refusals, how much the model's behaviour shifted on harmless prompts, and its scores on knowledge, maths and common-sense tests.

Results depend strongly on the model: some lose their refusals with barely any other change, others resist, and resisting goes together with bigger side effects.

Maths reasoning (GSM8K) is the most fragile skill, and the tool that searches hardest for the edit did the most maths damage, which says something about how tangled refusal is with other abilities.

Idea · §1–2

One idea, many implementations

A chat model's internal state is a long list of numbers that can be thought of as an arrow. Arditi et al. (2024) showed that refusing a harmful request is largely controlled by one direction of that arrow, and that the model's weights can be edited so this direction is never written. The technique became known as abliteration, and several open-source tools now automate it.

The tools differ in how they choose the direction, which layers they edit, how strongly, and whether they search over many settings. Until this paper, nobody had compared them carefully. Why does that matter for safety? Because how much damage the edit does, and which models resist it, tells us how deeply safety training is tied into a model's other abilities. The author's own framing is blunt: as currently implemented, alignment of open-weight models is "not a security mechanism but a speed bump", and measuring that is the point.

Everyday pictureFour surgeons all trained in the same operation. Comparing their outcomes tells you about the surgeons, but also about the patients: which bodies make the operation clean, and where the nerve you want to cut runs right next to one you need.

TakeawayThe study compares tools that implement the same weight edit, to learn how cleanly refusal can be separated from everything else in different models.

Idea · §3 · Table 1

How the comparison was done

Sixteen instruction-tuned models were used, including Llama-3.1-8B, Mistral-7B, several Qwen models, Gemma, Yi, DeepSeek, Phi-3, Vicuna and one model built on a different, non-transformer architecture (Falcon-Mamba). Four measurements mattered:

  • Remaining refusals: of 100 harmful prompts, how many still got a refusal.
  • KL divergence: a measure of how much the model's word choices on 100 harmless prompts changed. Near zero means "behaves just as before".
  • Benchmark scores: MMLU (general knowledge), GSM8K (grade-school maths word problems) and HellaSwag (common-sense sentence completion).
  • Coverage: on how many of the 16 models a tool worked at all. Heretic handled all 16, DECCP 11, ErisForge 9 and FailSpy's tool 5.

TakeawayThe study checks not only whether refusals disappear but how much else changes, using a "behaviour shift" score and three skill tests.

Idea · §4.2 · Tables 3–4

Same edit, very different damage

On general knowledge and common sense, every tool cost less than a point on average. Maths was different. The tool that runs an automated search over many edit settings (Heretic) cut the Yi-1.5-9B model's GSM8K score from 70.9% to 52.1%, a loss of 18.8 points, while a single-pass tool (DECCP) on the same model actually raised it by 1.5 points. Pick a model below to see the spread.

From Tables 3 and 4 of the paper. Per-model views: GSM8K accuracy (%) for the original model and after each tool. "Average change" view: mean change in points versus the original across the three models that had complete results for all three tools. The paper gives GSM8K's standard error as about 1.2–1.4 points, so only changes well beyond that are meaningful.

TakeawayKnowledge survives abliteration well; maths does not always, and the damage depends on both the tool and the model.

Check yourself
Is DECCP's +1.5 on Yi evidence that abliteration improves maths?
Not really. It is about the size of the benchmark's standard error, so it is best read as "no measurable damage".

Idea · §5 · Discussion

Why maths? Overlapping directions

The author suggests two reasons. Maths needs long, exact chains of reasoning, so a small disturbance anywhere can turn a right answer wrong. And refusal may overlap with maths reasoning inside the model: if the direction being deleted is not perfectly separate from directions the model uses for calculation, deleting it takes some maths with it. The paper frames this overlap as a possibility, not a proven finding.

The geometry is easy to see in two dimensions. Removing a direction removes, from every other feature, whatever part of it lies along that direction.

Illustrative toy. The amber line is the refusal direction being deleted. The blue arrow is a feature the model uses for maths. Deleting the refusal direction flattens the maths arrow onto the line at right angles to it (green). At 90° nothing is lost; the smaller the angle, the more of the maths feature goes with it.
Everyday pictureErasing pencil marks along a ruler's edge. Anything drawn exactly across the ruler survives; any line that runs partly along it loses that part.

TakeawayIf refusal shares direction with a useful skill, removing refusal damages that skill; maths appears most exposed.

Idea · §4.1, §5.2 · Table 2

Some models resist, and resisting costs

Across eight models run with the search-based tool, how many refusals survived ranged from 2 out of 100 (Zephyr-7B) to 54 (StableLM-2-12B). The models that kept the most refusals were also the ones whose harmless behaviour shifted most. The author reads this as "coupled resistance": in some models refusal is tangled up with general behaviour, so it cannot be removed without disturbing everything else.

Zephyr, the most easily stripped model, is the study's only model aligned with DPO alone (a single preference-training step). The author suggests, cautiously and from that one example, that multi-stage alignment (for example RLHF followed by DPO) may spread safety more widely through a model. That idea matches the extended-refusal defence, which deliberately spreads refusal out.

ContextThis paper does not propose new methods. Its value is measurement: it shows that "abliteration preserves capabilities" was an over-general claim, and that a model's resistance to it varies a lot. It also notes that the refusal checks used are rough (they look for refusal phrases), so they likely understate how often refusals were actually removed.

TakeawayHow hard refusal is to remove varies by model, and models where it is hard to remove are those where it is entangled with everything else.

Follow the idea

Built on: Abliteration (2024)