acceptodds
Under review as a conference paper at ICLR 2027

Turning Abliteration into Model Self-Destruction in Open-Weight LLMs

Abstract

Open-weight LLMs can be easily un-censored through abliteration, which is a method that works by removing the refusal direction from the residual stream. In this paper, we contribute an abliteration defense that makes the attack backfire. We ensure removing the refusal direction leads to either model collapse or persistent refusal, while leaving the clean model intact. During alignment we supervise the model on simulated abliteration in random subspaces of the refusal direction, training the model to emit a collapse token with strong abliteration and maintain refusal under gentle partial abliteration. On Llama-3.1-8B, abliteration attack-success (HarmBench) falls from 0.72 to 0.00, and we show the effect generalizes across multiple model families and to out-of-distribution and capability-preserving adaptive abliteration. We also provide a discussion of the field, as well as future work that can build off of our method.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
open-weight safety, abliteration, refusal, tamper-resistant safeguards, adaptive attacks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_g2jyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a defense against abliteration attacks on open-weight LLMs. The main idea is to train models so that removing their refusal directions either preserves refusal or causes the model to collapse, rather than producing a capable, non-refusing model.

The authors introduce a training objective with three terms: standard SFT, refusal preservation under partial ablation, and EOS-token prediction under stronger ablation. During training, they simulate ablation over randomly sampled directions, layers, and strengths.

They evaluate the defense on eight models across the Llama, Qwen, and Gemma families. Standard abliteration achieves 0.00 attack success rate across all defended models, although adaptive attacks remain effective on some models, particularly Qwen-2.5-72B.

Critical strengths
  1. New/Interesting approach to abliteration defense: I think the idea of making strong abliteration cause model collapse rather than simply trying to preserve refusal is interesting. The distinction between maintaining refusal under weaker interventions and deliberately inducing collapse under stronger interventions is a potentially useful defense objective.

  2. Evaluation across multiple models and attack settings: The authors evaluate eight models from three families, ranging from 7B to 72B parameters. I like that they include adaptive attacks and report cases where the defense fails, particularly Qwen-2.5-72B. The observation that some models collapse under abliteration while others retain their refusal behavior is also interesting.

Critical weaknesses
  1. Limited novelty relative to existing abliteration defenses. My main concern is that the proposed approach seems closely related to existing work, particularly ReFAT (Yu et al., ICLR 2025) and Abliteration-Resistant Tuning (ART; Kuo et al., 2026). Both already study training models against simulated refusal-feature ablation.

    The main difference appears to be the additional collapse objective, which trains models to emit EOS after sufficiently strong ablation, along with the use of randomized refusal subspaces. I think this is a reasonable extension, but the paper does not establish how much of the robustness comes from these additions versus the previously studied adversarial-training approach.

    I also think Decoy Direction Optimization (DDO; Muhamed et al., 2026) is an important related work that is missing. DDO addresses the same refusal-direction ablation threat model, but takes a different approach by modifying the attacker's estimated refusal direction rather than training the model to collapse under ablation.

    I would particularly like to see comparisons against ReFAT, DDO and ART, as well as an ablation removing the collapse loss while keeping the refusal-preservation loss. Without these comparisons, it is difficult to determine the contribution of the proposed objective.

  2. The attack evaluation does not adequately show robustness against adaptive abliteration. I also have a more substantial concern about the evaluation protocol. Appendix H states that the defended models are evaluated using refusal-only attack selection, whereas the undefended controls use stronger fluency-gated selection. This seems important given that Section 5.3 demonstrates that attack selection alone can change measured ASR from 0.00 to 0.56 on an undefended model.

    This confounder makes the headline reduction in ASR difficult to interpret, since the defended and undefended models are not evaluated under the same attack-selection procedure.

    More generally, the adaptive attack only searches over a restricted family of ablation configurations and is described as a simplified version of Heretic. Since the attacker has full white-box access, it could also try to find interventions that avoid triggering the learned collapse response. It is unclear whether the defense would remain effective against such attacks.

    The reported adaptive results already show limitations: Qwen-2.5-72B reaches 0.77 ASR, while Qwen-2.5-7B and 32B remain partially vulnerable. The defense also has a substantial clean capability cost on Gemma-1-7B, where LAMBADA accuracy drops from 0.62 to 0.33. These findings suggest that neither robustness nor capability preservation is consistent across models.

Questions for authors
  1. On novelty and baselines: How does the proposed approach compare with ReFAT and ART, which already train models against simulated refusal-feature ablation? Could the authors include a baseline that only uses SFT and the refusal-preservation objective, without the collapse term? It would be nice to include additional baselines like DDO as well.

  2. On attack evaluation: Why are defended models evaluated with refusal-only attack selection while undefended controls use fluency-gated selection? Given the large differences reported in Section 5.3, could the authors rerun the defended evaluations with the stronger, matched attack-selection procedure?

  3. On adaptive attacks: Could an adaptive attacker explicitly optimize to avoid the learned EOS-collapse behavior while suppressing refusal? Have the authors evaluated a full Heretic-style attack or other adaptive weight-space interventions outside the training distribution?

  4. On capability preservation: LAMBADA accuracy drops substantially on Gemma-1-7B, and the best training configuration differs between models. How sensitive is the method to these hyperparameters? Could the authors also evaluate general capabilities using benchmarks such as MMLU or instruction-following tasks?

1 comment

We thank the reviewer for taking the time to read our paper and for providing us with feedback that will allow us to improve our work.

  1. ReFAT and ART are close to what we do, and we should say so. All of these methods train a model to keep refusing even when someone removes its “refusal direction”, and our own ablation shows that this part is what does most of the work on Llama-8B. However, unlike ReFAT, we focus on abliteration, and unlike ART, we add a random-subspace variant, and a “collapse” objective that makes the model emit EOS when it is attacked hard. This increases the robustness to more adaptive attacks. For DDO, we agree that we should mention it in the related works section. We will make that change, and we will run comparison experiments with ART, ReFAT, and DDO, as well as ablations.
  2. We agree about the confound. We re-ran the results for Llama-3.1-8B: | Model | Clean refusal | Clean ASR | Clean PPL | Clean LAMBADA | ASR: standard abliteration | ASR: adaptive attack | | Ours (defended Llama-3.1-8B) | 0.99 | 0.00 | 3.85 | 0.76 | 0.13 | 0.15 |

As for the adaptive attackers, we will include results with other types of adaptive attackers. We will also include experiments that study hyperparameter sensitivity. As for other benchmarks, we have run additional experiments:

BenchmarkSettingBase Llama-3.1-8B-InstructDefended (ours)Change
MMLU5-shot, acc0.6860.675−0.010
– humanities5-shot0.6470.628−0.019
– other5-shot0.7490.741−0.008
– social sciences5-shot0.7780.771−0.008
– STEM5-shot0.5910.588−0.003
HellaSwag0-shot, acc0.5720.377−0.195
ARC-Challenge0-shot, acc0.4800.329−0.151
Winogrande0-shot, acc0.6710.647−0.024
Reviewer_vXZXOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a training-time defense against abliteration, the attack that uncensors open-weight LLMs by removing the refusal direction from a model’s internals. Instead of trying to protect that direction, the authors fine-tune the model on simulated abliteration so that strong removal makes it collapse into emitting only end-of-sequence tokens, while gentle removal still leaves its refusals intact. Tested on eight models from the Llama, Qwen, and Gemma families (7B to 72B), standard abliteration success drops from 0.72 to 0.00. A stronger adaptive attack is mostly blocked, with Qwen-2.5-72B as a clear failure and Gemma-1-7B suffering a notable capability loss. The authors admit that the defense covers only this one narrow class of attack. The defense needs per-model tuning and their defended and undefended attack setups weren’t selected the same way. They frame the result modestly as showing that a specific modification can be made to yield a broken model rather than a capable uncensored one.

Critical strengths
  1. Most of the existing defenses try to make the refusal direction harder to remove. This paper assumes it will be removed and trains the model to break or keep refusing when that happens.

  2. The authors report their failures. Qwen-2.5-72B is broken by the adaptive attack, Gemma-1-7B loses a lot of capability, and the settings have to be tuned per model.

  3. The attack-selection finding is useful. Figure 2 shows that picking an attack by refusal rate alone can select one that just wrecks the model, giving a fake 0.00 ASR even with no defense. Adding a fluency requirement moves it to 0.56.

Critical weaknesses
  1. It only defends against one very specific attack. The defense is trained against exactly the thing it’s tested on, a linear edit to the refusal direction. It says nothing about fine-tuning, which is already the more common way to strip safety from open models. Even the attacks it does face weren’t pushed very hard: the adaptive attack is a simplified version, and nobody tried going after the collapse mechanism itself. Meanwhile the defense already cracks on Qwen. It feels less like tamper resistance and more like patching one door and leaving the rest of the house open.

  2. It doesn’t feel like a general method. It needs different settings per model, and the fix that helped one Gemma model made another one worse. Nobody understands why it fails on Qwen or why large models behave differently. Each model was trained once, on small test sets, with no comparison to existing defenses like TAR, AntiDote or circuit breakers. It is difficult to tell whether this technique is effective in general or a result that happens to work on a few hand-tuned models. And the draft has some loose ends (broken appendix references, inconsistent counts of how many models it works on).

  3. The highlight result isn’t a fair representation of the method's capability. The 0.72 → 0.00 number compares an attacker who was carefully tuned against the undefended model with one who was picked more sloppily against the defended model. So the main claim, that the defense stops abliteration, is basically unproven until someone reruns it fairly.

Questions for authors
  1. The paper fails to cite many relevant existing papers in the related work. The following paper is an important baseline which is missing in the comparison section. Some relevant papers: a) Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attack b) AntiDote: Bi-level Adversarial Training for Tamper-Resistant LLMs

Check the Weakness section

1 comment

We thank the reviewer for taking the time to read our paper and for providing us with feedback that will allow us to improve our work.

  1. We agree that our method only defends against abliteration. Fine-tuning is not in the scope of the paper, so we will state that as a limitation.

  2. We agree that we need to compare to other defenses. We will include TAR, AntiDote, and Circuit Breakers. For the purposes of this rebuttal, we have run experiments for some of the other defenses: | Model | Clean refusal | Clean ASR | Clean PPL | Clean LAMBADA | ASR: standard abliteration | ASR: adaptive attack | |---|---|---|---|---|---|---| | Llama-3-8B-Instruct (undefended) | 0.99 | 0.00 | 7.15 | 0.72 | 0.84 | 0.79 | | Circuit Breakers (Llama-3-8B-RR) | 0.45 | 0.00 | 7.17 | 0.73 | 0.00 | 0.00 | | TAR (Llama-3-8B-TAR-Refusal) | 0.80 | 0.09 | 17.85 | 0.51 | 0.25 | 0.35 | | Llama-3.1-8B-Instruct (undefended) | 0.73 | 0.00 | 6.09 | 0.73 | 0.86 | 0.89 | | Ours (defended Llama-3.1-8B) | 0.99 | 0.00 | 3.85 | 0.76 | 0.13 | 0.15 |

  3. We agree that we should run more hyperparameter sweeps. We will do so, and report the results in the next version of the paper.

Reviewer_wxR6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a defense against abliteration attacks, which aim to remove safety alignment from LLMs by eliminating refusal directions. The method simulates abliteration during fine-tuning and trains the model to maintain safe responses under these interventions. The reported results are promising, but the evaluation is limited in scope.

Critical strengths
  1. The paper addresses an interesting and important problem: preserving LLM safety under attacks that directly modify model representations.

  2. The proposed training objective offers a promising approach to improving robustness against abliteration attacks.

Critical weaknesses
  1. Some claims are stronger than the evidence supports. For example, the use of “ensure” in line 014 implies a guarantee that cannot be established through a limited set of empirical evaluations. The authors should qualify this language and align their claims with the scope of the evidence.

  2. The method appears to substantially degrade clean-model performance. In Table 1, the score for Gemma-1-7B drops from 0.62 before applying the method to 0.33 afterward. Qwen-2.5-7B also exhibits a performance decrease of approximately 5%. These results appear inconsistent with the abstract’s claim of “leaving the clean model intact.” The authors should explicitly discuss this trade-off and assess whether the safety gains justify the loss in utility, particularly for practical deployment.

  3. The evaluation of model utility is too narrow. Table 1 relies on a small dataset, which provides limited evidence about the method’s effect on general model capabilities. The authors should evaluate performance across a broader range of established benchmarks covering diverse tasks.

  4. The presentation contains redundant results. Table 1 and Figure 1 appear to convey the same information. The authors should consider retaining only one or revising the figure to provide additional insight.

  5. The experimental validation is insufficient. The limited range of baselines and models makes it difficult to assess the method’s comparative effectiveness and generalizability. Additional comparisons with relevant defenses and evaluations across more model families and sizes would strengthen the paper.

Questions for authors

see above

1 comment

We thank the reviewer for taking the time to read our paper and for providing us with feedback that will allow us to improve our work. With that said:

  1. We agree with the reviewer that some claims in our paper are stronger than the evidence supports. We will go through the paper and change such claims. For example, instead of “We ensure that removing the refusal direction leads to either model collapse or persistent refusal, while leaving the clean model intact”, we will say “Our method attempts to ensure that removing the refusal direction leads to either model collapse or persistent refusal, while leaving the clean model intact”
  2. We agree with the reviewer that our method does not leave the clean model intact. We will modify the paper accordingly and discuss the safety-utility trade-off.
  3. We agree that our capabilities experiments are limited. We have widened it and will run the experiments with all models: | Benchmark | Setting | Base Llama-3.1-8B-Instruct | Defended (ours) | Change | |---|---|---|---|---| | MMLU | 5-shot, acc | 0.686 | 0.675 | −0.010 | | – humanities | 5-shot | 0.647 | 0.628 | −0.019 | | – other | 5-shot | 0.749 | 0.741 | −0.008 | | – social sciences | 5-shot | 0.778 | 0.771 | −0.008 | | – STEM | 5-shot | 0.591 | 0.588 | −0.003 | | HellaSwag | 0-shot, acc | 0.572 | 0.377 | −0.195 | | ARC-Challenge | 0-shot, acc | 0.480 | 0.329 | −0.151 | | Winogrande | 0-shot, acc | 0.671 | 0.647 | −0.024 |
  4. We agree that Table 1 and Figure 1 are redundant. We will keep Table 1.
  5. We agree that using more baselines is important. However, we would like to note that we validate our method with 8 models from three different families, and sizes ranging from 7B to 70B. In terms of baselines, we have expanded our evaluation for Llama 8B, and will run the experiments on the rest of the models. | Model | Clean refusal | Clean ASR | Clean PPL | Clean LAMBADA | ASR: standard abliteration | ASR: adaptive attack | |---|---|---|---|---|---|---| | Llama-3-8B-Instruct (undefended) | 0.99 | 0.00 | 7.15 | 0.72 | 0.84 | 0.79 | | Circuit Breakers (Llama-3-8B-RR) | 0.45 | 0.00 | 7.17 | 0.73 | 0.00 | 0.00 | | TAR (Llama-3-8B-TAR-Refusal) | 0.80 | 0.09 | 17.85 | 0.51 | 0.25 | 0.35 | | Llama-3.1-8B-Instruct (undefended) | 0.73 | 0.00 | 6.09 | 0.73 | 0.86 | 0.89 | | Ours (defended Llama-3.1-8B) | 0.99 | 0.00 | 3.85 | 0.76 | 0.13 | 0.15 |

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.