acceptodds
Withdrawn from ICLR 2027

Target-free Latent Safety Alignment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models (LLMs) remain highly vulnerable to jailbreak attacks that induce harmful behaviors and circumvent safety alignment. To defend against such attacks, adversarial training paradigms have been proposed to first simulate failure modes and then train the model to correct them, yielding promising improvements in safety alignment. However, these methods typically construct adversarial samples either by encouraging fixed harmful target completions or by performing targeted activation ablation derived from fixed benign–harmful data pairs. As a result, the generated adversarial samples tend to induce homogeneous harmful behaviors that poorly reflect the diversity of behaviors elicited by real-world jailbreak attacks. This behavior-level narrowness fundamentally limits their robustness. To address this issue, we propose a target-free adversarial training framework that generates adversarial samples in an unsupervised manner. By amplifying and diversifying behavior-level shifts in the model’s latent space, our approach produces semantically diverse adversarial samples that induce a wide range of harmful behaviors. This expanded behavioral coverage exposes more diverse failure modes and thereby improves safety alignment. To quantify this effect, we use semantic entropy as an output-level measure of adversarial behavioral diversity. Empirically, our method elicits diverse harmful behaviors in the target model, substantially mitigating behavioral narrowness and improving robustness to jailbreak attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
latent adversarial training, safety alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pCH8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the latent adversarial training and proposes Target-free Latent Safety Alignment to improve jailbreak robustness by increasing the behavioural diversity of latent adversaries. Unlike prior LAT methods that optimise individual interventions toward fixed harmful targets or refusal-ablation directions, the proposed method jointly learns multiple latent intervention directions and maximises the nuclear norm of their induced behaviour-shift matrix. The resulting adversaries are then used to align safety while preserving benign utility. Experiments across multiple jailbreak attacks show improved robustness with limited utility degradation

Critical strengths
  1. The motivation and method are clear, and the visualization is helpful.

  2. The authors provide sufficient evidence to support the effectiveness of their proposed method.


  3. This paper is well-written, making it easy to follow.

Critical weaknesses
  1. The claimed advantage over prior LAT in terms of behavioural diversity needs a stronger justification. The paper argues that prior LAT is behaviorally narrow because each optimisation step constructs a single intervention, whereas the proposed method jointly learns multiple directions. However, a single intervention per optimisation run does not necessarily imply a single-mode adversarial distribution: prior LAT could potentially generate diverse interventions through different prompts, initialisations, or repeated optimisation. A more controlled comparison under matched numbers of adversarial directions/samples would better establish whether the improvement comes from the proposed objective rather than simply generating more interventions.

  2. The connection between the target-free objective and harmful behavioural diversity is insufficiently established. Maximising the nuclear norm encourages strong and diverse latent shifts, but does not explicitly ensure that these shifts correspond to harmful behaviours. Indeed, not all optimised interventions produce harmful outputs. Moreover, the reported modes, such as multilingual, role-play, coding, and direct compliance, are qualitative post-hoc interpretations rather than independently annotated behavioural categories. Stronger evidence is needed to show that latent diversity corresponds to meaningful diversity of real-world jailbreak behaviours.

  3. The empirical evaluation could be broader. All evaluated models are small dense LLMs, leaving generalisation to larger (such as Qwen 3.5 27B) and structurally different models, particularly MoE architectures (such as gpt-oss-20b), unclear. Finally, evaluation on more realistic jailbreak datasets such as WildJailbreak and JailbreakBench, rather than primarily attack-generated variants of HarmBench, would provide stronger evidence of robustness under diverse real-world jailbreak distributions.

Questions for authors

Refer to weaknesses.

Reviewer_WHu5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper starts from the claim that current latent adversarial training (LAT) methods lack diversity, for their attacks elicit a narrow range of harmful behaviors. (Recall that LAT methods are anti-jailbreak techniques which consist in (i) adversarially perturbing a language model's activations to elicit harmful content from it, then (ii) fine-tuning said LM so that it refuses generating said content.)

The paper aims to address the issue above by (i) learning 512 vectors shared across all harmful prompts (as opposed to one vector and one harmful target), (ii) adding them to the residual stream at an early layer, and (iii) measuring the resulting change of the residual stream at a later layers. The adversary's goal then is to maximize the nuclear norm (i.e., the sum of the singular values) of the matrix whose 512 columns are the average changes obtained in (iii). (The intuition is that such a nuclear norm is large when each average change is large, when the change caused by one vector is consistent across prompts, and when the changes caused by different vectors are orthogonal.) The update the model using LoRA adapters. Over three ~7B models and seven attacks, the authors claim that: i. Table 1. Models are more robust to jailbreaks than 4 LAT baselines, over-refuse less than the baselines on two of the three models, and are roughly as "useful" as the baselines. ii. Fig. 2 and Table 2. As the adversary optimizes its vectors, the model complies more often and its answers become more diverse. Iii. Fig. 3. On one model they compare four adversaries, i.e., no adversary, one targeted vector, the 64 most dissimilar of their 512 vectors, and all 512. The latter two yield similar results, and both outperform the other two baselines.

Critical strengths

i. The adversarial method is simple and its underlying rationale is clear. The method is also cheap (i.e., one objective instead of, e.g., harmful completions or contrastive pairs). ii. The reported improvements over the four baselines are quite significant and consistent.

Critical weaknesses

i. It appears that a very similar paper was published at ICML 2026 ("Steering beyond the support..."), and the submission does not cite or compare against it.

  • They are equivalent in their first half (the adversary) and differ in the second half (where the ICML paper employs steering, this paper employs LoRA fine-tuning). The experiments, too, are the same (e.g., same models, overlapping attacks, same judge, same utility benchmarks, same LAT baseline, ...). A comparison with this paper is notably missing, and the novelty of the current paper is therefore significantly overstated.
  • Moreover, the two papers seem to numerically agree on many data (cf. the respective Tables 1 and 3), yet they differ on the baselines, which look worse in the current submission. As such, the soundness of the current paper cannot be assumed.

ii. Some claims are overstated. For instance:

  • Fig. 3 does not necessarily show that diversity among vectors is what yields robustness. For instance, training for 512 non-diverse vectors could produce similar results, or training against 512 random ones.
  • The "full-V NLL" is the refusal loss under all 512 learned vectors, so it is measured on the method's own adversarial set (and this disproportionally favors variants trained on it). By contrast, the SR panel is a fair comparison.
  • The premise that prior LAT adversaries are narrow is, indeed, only assumed. For instance, the paper does not report the baseline semantic entropy.
Questions for authors

Can you clarify the absent comparison with recent work?

Reviewer_TXyiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the behavior narrowing problem in latent adversarial training (LAT) by proposing a new training framework. The method learns a set of latent intervention vectors and uses nuclear norm maximization to encourage both strong and diverse behavioral shifts across interventions, combined with a harmful consistency loss and a benign KL consistency loss for LoRA fine-tuning. Experiments on several open-source models demonstrate its effectiveness.

Critical strengths
  1. The problem is well-motivated: behavior narrowing in LAT is a practical concern in safety alignment, and approaching it from a diversity perspective is reasonable.

  2. The technical design has some novelty: using nuclear norm to characterize the diversity of intervention vectors is more elegant than simple random perturbations or fixed templates.

  3. The experimental evaluation is fairly comprehensive, covering multiple mainstream models and showing improvements on some challenging benchmarks.

Critical weaknesses
  1. The title does not accurately reflect the method. The term "Target-Free" may lead readers to believe that the method operates without any target at all, whereas the approach still optimizes toward generating harmful behaviors, which simply replaces explicit harmful text targets with an implicit diversity objective. A more precise title highlighting "generalization enhancement" or "diversity regularization" would be more appropriate.

  2. The generalization boundary is unclear. The paper claims that K intervention vectors can achieve diverse coverage, but these vectors are learned entirely from the training data. There is no theoretical or experimental guarantee for out-of-distribution attacks (e.g., encoded instructions, multi-turn induction). Explicitly discussing the scope and limitations would strengthen the paper.

  3. The contribution is incremental rather than a paradigm shift. The nuclear norm regularization is a useful patch to LAT, but the overall framework remains within the existing adversarial training paradigm. The narrative could be more modest, positioning the work as "a regularization method to improve diversity in LAT."

  4. Theoretical depth is limited. The approach is largely technical, lacking deeper analysis on why diversity leads to better defense, which reduces the paper's insightfulness.

Questions for authors
  1. How was the value of K chosen? At what point does performance saturate as K increases? Is there any theoretical guidance for selecting K?

  2. All evaluated attacks come from known semantic libraries. How would the method perform against out-of-distribution adversarial inputs (e.g., base64-encoded instructions, poetry-style malicious prompts)? Are there plans for such OOD tests?

  3. Does the training process require a harmful dataset to initialize the intervention vectors? If harmful data were completely removed, could effective interventions be learned solely through nuclear norm maximization?

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

Related papers

Discussion (0)

Sign in to comment.