- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes an interesting idea to defend against prefill jailbreak attacks by restoring the model's weakened refusal state. Its main design, J-Anchor, uses the model's pre-attack safety state as a request-specific anchor and selectively restores refusal signals in J-space during prefilling and decoding, without requiring additional training. In the evaluation, the authors test J-Anchor on five open-weight LLMs under both fixed and deep prefill attacks. Compared to existing baselines, J-Anchor achieves strong defense performance, reducing attack success rates to nearly zero in most settings while largely preserving benign utility and reasoning capabilities. Overall, the results are promising and demonstrate a favorable trade-off between safety and utility.
Critical strengths
- The paper is well-written and easy to follow, with clear motivation and problem formulation.
- The paper presents a novel approach to addressing refusal-state erosion in language models.
- The empirical results demonstrate the effectiveness of the proposed method in improving safety under prefill attacks.
Critical weaknesses
- Lack of ablation studies to justify key design choices in J-Anchor.
- Lack of comparison with more well-known strong baselines
- The small judge model can be a concern and may introduce some bias in the evaluation.
Questions for authors
Overall, I like this paper. The motivation is clear, and I find the idea of tracking refusal-state erosion and restoring refusal based on the model's pre-attack state quite interesting. The empirical results are also strong, especially on deep prefill attacks. But I have a few questions and suggestions.
- More component-level ablations would be helpful.
While the paper includes ablations on intervention directions and representation spaces, I feel that some of the key design choices in J-Anchor are not sufficiently justified.
For example:
(1) How much does the request-level gate actually help? What happens if it is removed?
(2) Is the request-specific anchor really necessary? Would a fixed global refusal target work similarly?
(3) Do we really need to intervene during both prefilling and decoding? What happens if we only intervene during one of these stages?
(4) How much does the harmfulness signal contribute to setting the restoration target? Would using only the refusal signal achieve similar results?
I think these ablations would help clarify which components are actually driving the improvements.
- It would be nice to include a few more baselines.
I think two relevant works are worth considering:
Improving Alignment and Robustness with Circuit Breakers (NeurIPS 2024)
Safety Alignment Should Be Made More Than Just a Few Tokens Deep (ICLR 2025)
Both explore alternative ways to make safety behavior more robust beyond the initial refusal tokens. Circuit Breakers, in particular, directly intervenes in harmful representations to disrupt unsafe generation, while the second paper studies shallow safety alignment and proposes ways to make safety behavior persist deeper into generation.
I wonder how these approaches would perform under the same prefill attacks. It would be interesting to see whether they can achieve similar robustness, and how they compare in terms of benign utility.
- The judge model used for evaluation may introduce some bias.
The paper uses HarmBench-Llama-2-13B-cls for ASR evaluation and WildGuard for over-refusal evaluation. While these are standard safety evaluators, I wonder whether the results would remain consistent when evaluated using a stronger, independent LLM judge (e.g., a GPT-class model)[Fine-tuning aligned language models compromises safety, even when users do not intend to! ICLR 2024], or through a small-scale human evaluation.
My concern is that since the defense directly manipulates safety-related representations, it might change how unsafe responses are expressed, potentially making them harder for certain safety classifiers to detect. This could lead to some judge-specific evaluation bias.
5 comments
W2: Comparison with Stronger Baselines
Thank you for the suggestions. Our original baseline selection was designed to cover representative defense mechanisms and methods directly targeting assistant-prefill attacks. Specifically, Refusal Vector provides a fixed activation-steering baseline, while CAST and AlphaSteer represent conditional and input-dependent steering approaches. We also include ADA-RK and ADA-LP, which are specifically designed to defend against deep-prefill attacks. Together, these baselines enable comparisons across different steering strategies and defenses developed for the same threat model.
To address your concerns, we extend our evaluation to Circuit Breakers (NeurIPS 2024) [1] and Safety Alignment Should Be Made More Than Just a Few Tokens Deep (ICLR 2025) [2] on Llama-3.1-8B-Instruct and Qwen3.5-4B, using the same fixed and deep-prefill attack settings. We report ASR on StrongREJECT (SR), AdvBench (AB), and HarmBench (HB), and ORR on XSTest (XS) and OKTest (OK).
- Llama-3.1-8B-Instruct
| Method | Fixed SR ↓ | Fixed AB ↓ | Fixed HB ↓ | Deep500 SR ↓ | Deep500 AB ↓ | Deep500 HB ↓ | ORR XS ↓ | ORR OK ↓ |
|---|---|---|---|---|---|---|---|---|
| Circuit Breakers | 1.9 | 1.0 | 1.3 | 3.2 | 0.0 | 1.9 | 8.8 | 24.7 |
| Deep Safety Alignment | 0.3 | 1.0 | 0.6 | 60.1 | 56.0 | 66.0 | 10.0 | 20.7 |
| J-Anchor | 1.0 | 0.0 | 3.8 | 0.3 | 0.0 | 2.5 | 7.6 | 23.3 |
- Qwen3.5-4B
| Method | Fixed SR ↓ | Fixed AB ↓ | Fixed HB ↓ | Deep500 SR ↓ | Deep500 AB ↓ | Deep500 HB ↓ | ORR XS ↓ | ORR OK ↓ |
|---|---|---|---|---|---|---|---|---|
| Circuit Breakers | 1.0 | 0.0 | 0.0 | 4.2 | 1.0 | 0.6 | 9.2 | 29.0 |
| Deep Safety Alignment | 0.3 | 0.0 | 0.6 | 6.7 | 4.0 | 18.2 | 9.2 | 10.3 |
| J-Anchor | 0.3 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 | 6.4 | 17.0 |
J-Anchor remains competitive with both stronger baselines across the two models. Under deep prefill, J-Anchor achieves the lowest or joint-lowest ASR on all three datasets on Qwen, and on StrongREJECT and AdvBench on Llama. It also achieves the lowest XSTest ORR on both models. We have added these comparisons to the revised manuscript (Appendix D.4).
W3: Reliability of the Evaluation Judge
Our original evaluation follows established safety evaluation protocols, using HarmBench-Llama-2-13B-cls to assess attack success and WildGuard to identify refusals on benign requests. Both are widely used, task-specific safety evaluators, providing a consistent basis for comparing different defense methods. To assess robustness to judge choice, we additionally conduct LLM-as-judge evaluation (DeepSeek-V4-Flash [3]). We have included these results in a dedicated subsection in the revised manuscript (Appendix E).
The results are shown below. All values are percentages.
- Llama-3.1-8B-Instruct
| Method | Fixed SR↓ | Fixed AB↓ | Fixed HB↓ | Deep SR↓ | Deep AB↓ | Deep HB↓ | ORR XS↓ | ORR OK↓ | GSM8K | MATH500 |
|---|---|---|---|---|---|---|---|---|---|---|
| Base | 70.9 | 71.0 | 63.5 | 99.4 | 99.0 | 96.2 | 6.8 | 1.0 | 84.0 | 53.0 |
| Refusal Vector | 46.6 | 62.0 | 35.2 | 97.1 | 97.0 | 93.1 | 60.4 | 56.3 | 4.0 | 35.0 |
| CAST | 57.8 | 44.0 | 45.3 | 97.4 | 97.0 | 95.0 | 5.6 | 3.7 | 83.0 | 48.0 |
| AlphaSteer† | 21.7 | 15.0 | 15.7 | 98.4 | 96.0 | 95.0 | 7.2 | 2.3 | 89.0 | 52.0 |
| ADA-RK | 1.3 | 0.0 | 9.4 | 7.7 | 5.0 | 17.6 | 11.6 | 4.7 | 84.0 | 53.0 |
| ADA-LP† | 0.3 | 0.0 | 0.0 | 0.3 | 0.0 | 0.6 | 6.8 | 1.7 | 84.0 | 53.0 |
| J-Anchor (Ours) | 1.3 | 0.0 | 2.5 | 0.3 | 0.0 | 3.8 | 7.2 | 1.3 | 87.0 | 51.0 |
- Gemma-4-E4B-IT
| Method | Fixed SR↓ | Fixed AB↓ | Fixed HB↓ | Deep SR ↓ | Deep AB ↓ | Deep HB↓ | ORR XS↓ | ORR OK↓ | GSM8K | MATH500 |
|---|---|---|---|---|---|---|---|---|---|---|
| Base | 91.4 | 94.0 | 80.5 | 99.0 | 99.0 | 96.9 | 9.2 | 3.0 | 94.0 | 90.0 |
| Refusal Vector | 88.5 | 88.0 | 64.8 | 95.5 | 98.0 | 86.2 | 77.6 | 45.0 | 85.0 | 85.0 |
| CAST | 68.1 | 77.0 | 45.9 | 95.2 | 92.0 | 79.2 | 16.0 | 4.0 | 88.0 | 82.0 |
| AlphaSteer† | 87.2 | 92.0 | 79.2 | 99.0 | 99.0 | 97.5 | 14.4 | 2.3 | 94.0 | 87.0 |
| ADA-RK | 0.0 | 0.0 | 0.0 | 0.3 | 0.0 | 3.1 | 18.8 | 14.0 | 94.0 | 90.0 |
| ADA-LP† | 8.3 | 4.0 | 5.0 | 11.2 | 8.0 | 10.1 | 12.8 | 9.0 | 85.0 | 87.0 |
| J-Anchor (Ours) | 0.0 | 0.0 | 0.6 | 0.3 | 0.0 | 0.0 | 15.6 | 11.7 | 91.0 | 91.0 |
- Qwen3.5-4B
| Method | Fixed SR↓ | Fixed AB↓ | Fixed HB↓ | Deep SR↓ | Deep AB↓ | Deep HB↓ | ORR XS↓ | ORR OK↓ | GSM8K | MATH500 |
|---|---|---|---|---|---|---|---|---|---|---|
| Base | 81.8 | 88.0 | 67.9 | 98.7 | 98.0 | 96.9 | 4.4 | 1.0 | 94.0 | 92.0 |
| Refusal Vector | 83.4 | 92.0 | 69.2 | 98.7 | 98.0 | 95.0 | 24.4 | 21.0 | 90.0 | 85.0 |
| CAST | 77.0 | 80.0 | 56.6 | 99.4 | 99.0 | 94.3 | 4.0 | 4.0 | 94.0 | 86.0 |
| AlphaSteer† | 59.7 | 57.0 | 47.8 | 95.5 | 95.0 | 95.0 | 44.0 | 26.0 | 95.0 | 92.0 |
| ADA-RK | 0.0 | 0.0 | 0.6 | 0.0 | 0.0 | 0.0 | 20.4 | 14.0 | 94.0 | 92.0 |
| ADA-LP† | 0.0 | 0.0 | 0.0 | 1.9 | 2.0 | 8.8 | 4.8 | 1.7 | 86.0 | 80.0 |
| J-Anchor (Ours) | 0.6 | 1.0 | 0.6 | 1.0 | 0.0 | 1.3 | 6.8 | 3.3 | 92.0 | 90.0 |
The results are consistent with our main findings on both fixed and deep prefill attacks while largely preserving mathematical reasoning performance. Although the ORR values vary across judges, the relative changes between J-Anchor and the undefended base model remain generally modest. These additional results provide further evidence of J-Anchor’s effectiveness across different safety judges.
References
[1] Zou et al. Improving Alignment and Robustness with Circuit Breakers. NeurIPS, 2024.
[2] Qi et al. Safety Alignment Should Be Made More Than Just a Few Tokens Deep. ICLR, 2025.
[3] Xu et al. DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence. arXiv preprint arXiv:2606.19348, 2026.
We thank the reviewer for the positive assessment and constructive suggestions. We address the three concerns below.
W1: Component-Level Ablations
Thank you for the suggestions that have helped us further improve and strengthen our manuscript. In the original submission, we focused on ablations of intervention directions and representation spaces (Table 3). Following your suggestions, we further conducted five component-level ablation variants on Llama-3.1-8B-Instruct and Qwen3.5-4B under the deep prefill attacks (k=500). We have incorporated these results into the revised manuscript (Section 5 and Appendix D).
Without gate: Removing the gate maintains low ASR and similar ORR but reduces MATH500 accuracy from 87% to 43% on Qwen and from 55% to 23% on Llama. Llama’s GSM8K accuracy also decreases from 81% to 53%, highlighting the gate’s role in preserving reasoning utility.
Fixed global target: Replacing the request-specific restoration target with the median value from a separate AdvBench development split, shared across all requests while keeping the gate and decoding trigger unchanged, maintains low ASR but increases XSTest ORR from 6.4% to 42.0% on Qwen and from 7.6% to 19.2% on Llama, highlighting the benefit of request-specific restoration strength.
Prefilling only and decoding only: Prefilling-only and decoding-only interventions yield higher overall deep-prefill ASR than their combination on both models, with decoding-only generally more effective than prefilling-only.
W/o harmfulness in target: At the same restoration multiplier, using only refusal to construct the restoration target increases deep-prefill ASR from 0.0–1.0% to 9.0–23.9% on Qwen and from 0.0–2.5% to 31.0–44.4% on Llama, highlighting the contribution of harmfulness-aware restoration.
These comparisons support the contributions of gating, request-specific anchoring, intervention stages, and harmfulness-aware restoration across both models.