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.
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.
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.
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.
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.
Abstract, restated
Safety alignment makes LLMs refuse harmful queries; abliteration removes refusal by orthogonalising weights against refusal-related directions, and several open-source tools implement it. This study evaluates four tools (Heretic, DECCP, ErisForge, FailSpy) on sixteen instruction-tuned models from 7B to 14B parameters, measuring tool compatibility, refusal suppression, distribution shift and capability preservation. Single-pass methods preserve capabilities better on the benchmarked subset than the Bayesian-optimised method, which produced variable distribution shift; mathematical reasoning (GSM8K) is the most sensitive capability. Richard J. Young, arXiv:2512.13655 (v1 December 2025; v2 January 2026).
Section · §2.1–2.3
The edit and its variants
All tools implement a variant of weight orthogonalisation against a refusal direction r (unit norm), with an ablation strength α:
α = 1 is the full projection of Arditi et al.; other values partially remove the direction. Two variants appear in the tools: norm-preserving ablation (attributed to Lai), which removes the directional component but restores each weight row's original norm, and projected ablation, which first uses Gram–Schmidt to remove the harmless-mean component from r, so that the edit does not also delete what harmful and harmless prompts share.
The tools differ in how they choose r, layers and strength:
| Tool | Approach (as described in §2.3) | Coverage |
|---|---|---|
| Heretic | Bayesian optimisation (Optuna TPE, 50 trials) over layer ranges, ablation weights and direction indices; objective combines KL divergence on harmless prompts with refusal rate. | 16/16 |
| DECCP | Single pass; direction chosen by refusal-direction magnitude scanning; memory-efficient sharded processing; norm-preserving and projected variants. | 11/16 |
| ErisForge | Single pass; wraps decoder layers in ablation (or addition) modules with a weight parameter (1.0 in this study). | 9/16 |
| FailSpy abliterator | TransformerLens hooks; selects the layer with maximum refusal-direction magnitude. | 5/16 |
Check yourself
- Why might norm preservation reduce side effects?
- Plain projection shortens every row that had a component along r, changing activation scales seen by later layers and normalisation. Restoring the norm removes the direction while keeping magnitudes as the rest of the network expects.
Section · §1, §3 · Hypotheses
Design, metrics and hypotheses
Hypotheses. H1 (model dependence): outcomes are strongly model-dependent, with some models showing low KL and low residual refusals and others "coupled resistance". H2 (capability preservation): single-pass methods (DECCP, ErisForge) show smaller average capability loss than optimisation-based Heretic, largest on GSM8K. H3 (alignment sensitivity): a DPO-only model is more susceptible than RLHF-aligned ones.
Metrics. First-token KL divergence on 100 harmless prompts (below 0.1 "excellent", above 1.0 significant shift). Refusal rate on 100 harmful prompts by refusal-marker matching; ASR = 100% − refusal rate. MMLU (5-shot), GSM8K (5-shot, strict match), HellaSwag (10-shot, normalised). Effect sizes binned as minimal (|Δ| < 1 pp), moderate (1–5) and substantial (≥ 5).
Ethics (§3.4). The author argues the contribution is "systematic evaluation and comparison, not capability advancement", since the tools are already public, and does not release abliterated weights for models without explicitly permissive licences.
Section · §4.1 · Table 2 · H1
Model dependence: KL vs residual refusals
Heretic was run to completion on eight models. Residual refusals range from 2/100 to 54/100 and KL from 0.043 to 1.646. The two are strongly correlated (Pearson r = 0.87, p = 0.005, n = 8): models that resist refusal removal also suffer the largest distribution shift on harmless prompts. The figure recomputes the correlation from the table.
Check yourself
- Why does dropping two points change r so much?
- With eight points, two high-leverage outliers can create most of a correlation. The remaining six are much more weakly related, so H1's support rests substantially on two models.
Section · §4.2–4.4 · Tables 3–4 · H2, H3
Capability preservation and coverage
Three models (DeepSeek-7B, Mistral-7B, Yi-1.5-9B) had complete results for Heretic, DECCP and ErisForge. Average changes versus base (Table 4): Heretic −0.78 MMLU, −7.81 GSM8K, −0.56 HellaSwag; DECCP −0.61, −0.13, −0.25; ErisForge −0.12, −0.28, −0.08. GSM8K spread runs from +1.51 (DECCP on Yi) to −18.81 (Heretic on Yi; −26.5% relative). H2 is supported on this subset.
For H3, the paper says the result "is consistent with" the hypothesis but that "stronger conclusions require additional DPO-only models": Zephyr-7B-beta, the DPO-only model, had the fewest residual refusals (2/100), but one model cannot separate alignment method from other differences. Coverage: Heretic was the only tool to process the Mamba state-space model, as the others assume transformer blocks.
The paper's refusal heuristic is coarse. On a separate cross-check (UncensorBench, n = 900) it gave 72.2% non-refusal where a classifier gave 95.7%, so marker-based refusal rates likely understate how much refusal was removed.
Section · §5.3 · Lineage
Limitations and lineage
- Unmatched budgets. Tools ran at defaults, so differences mix algorithm with compute and refusal targets. Heretic optimises for low KL subject to refusal constraints and may simply push refusal further; a fair comparison needs Pareto curves at matched refusal rates, as the paper itself recommends.
- Single runs and small n. One run per configuration; only three models with complete cross-tool baselines; H1's correlation rests on eight points.
- Narrow evaluation. Three benchmarks; no instruction following, long-context or multi-turn behaviour; marker-based refusal detection; no models above 14B or mixture-of-experts.
Reference
Glossary
- Abliteration
- Orthogonalising weights against a refusal direction.
- KL divergence
- Here, shift in the first-token output distribution on harmless prompts.
- Single-pass
- Compute a direction once and apply it, no search.
- TPE
- Tree-structured Parzen estimator, a Bayesian optimisation method (Optuna).
- Norm-preserving
- Variant that restores each row's norm after removing the direction.
- Projected ablation
- Variant removing the harmless-mean component from r first.
- DPO
- Direct preference optimisation, an alignment method (see DPO).
- Coupled resistance
- Residual refusal that comes with high distribution shift.