acceptodds
Under review as a conference paper at ICLR 2027

How Narrative Wrapping Affects LLM Refusal: A Cross-Language Benchmark and Defense

Abstract

Safety-aligned language models often refuse a harmful request stated directly but answer the same request inside a role-play or narrative wrapper. We measure this vulnerability across languages and registers: attack success on Qwen3-1.7B is already 89.4% in English and 93.0% in modern Chinese, and reaches 95.7% in Classical Chinese. We build GUISE, a benchmark for systematically studying this vulnerability. It includes parallel requests in English, modern Chinese, and Classical Chinese, matched harmful and benign pairs, wrapper types held out for evaluation, and a stricter criterion that counts warn-then-answer responses as attack successes. Representation analysis shows that language and register move harmful-request representations only slightly away from the model’s refusal direction, whereas narrative wrappers move them much farther away. We propose AXIS, which combines preference optimisation with a rotation objective that aligns harmful-request representations with the refusal direction and a commitment objective that trains the model to refuse completely rather than produce a warn-then-answer response. Across Qwen3-1.7B, Qwen3-4B and GLM-4-9B, AXIS achieves the highest combined safety and usability score among the compared methods.

Primary area:
datasets and benchmarks
Keywords:
Cybersecurity, Large Language Model, Deep Learning

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_qAnkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how narrative wrapping affects LLM refusal across English, modern Chinese, and Classical Chinese. It first introduces GUISE, a benchmark containing harmful requests, benign controls, and held-out wrapper types, with an evaluator that counts warn-then-answer responses as attack successes. Its representation analysis associates wrapping with a shift away from a reference refusal direction. The proposed defense, AXIS, combines DPO, a representation-alignment objective, and a weighted response-supervision objective. Experiments on three models report improvements in the combined safety–usability score.

Critical strengths
  • The paper addresses a relevant safety problem and attempts to distinguish the effects of narrative wrapping from those of language and register.

  • The paper is generally easy to read.

  • The paper makes two contributions: a benchmark for studying this problem and a defence method.

Critical weaknesses
  • The presentation could be improved. Although the paper is generally understandable, the introduction contains many experimental details and numerical results that would fit better in the methods and results sections. These details can distract readers from the main motivation and ideas.

  • I am not convinced that the results sufficiently support the choice of the reference refusal direction. Section 4 describes the representation analysis as correlational. However, Table 3 shows that random and orthogonal anchors also achieve strong safety performance, particularly on the larger models. For example, on GLM-4-9B, the orthogonal anchor achieves 99.3% OOD safety versus 99.8% for the real anchor, while reducing XSTest over-refusal from 14.7% to 4.1%. It is therefore unclear how much of the improvement actually depends on aligning with the original refusal direction.

  • Another concern is that the setting is quite narrow. The method relies on aligning directions estimated from Classical and modern Chinese requests. Although the paper reports transfer to English, it is unclear how the approach would be applied more broadly to other languages and registers. A clearer discussion of how to use the method outside this particular setting would help establish its broader value.

Questions for authors
  • How do you interpret the strong performance of the random and orthogonal anchors in Table 3? What evidence shows that aligning with the original refusal direction is particularly beneficial?

  • How would AXIS be applied to other languages that do not have a clear classical–modern register distinction? How would you choose the reference direction in those settings?

1 comment

Thank you for highlighting the anchor controls and the applicability to other languages. We have added an intervention-based comparison of anchor choices, clarified the role of refusal-direction overlap, and made the cross-language application procedure explicit.

W1. Introduction contains too many numerical details

Clarification: The introduction highlights 20.4% attack success for bare Classical Chinese requests versus 95.7% for wrapped requests (Table 1). These aggregate rates use 334 and 1,067 intents, respectively; the caption now explicitly distinguishes their denominators. Table 1 retains the cross-register measurements, while the detailed defense comparisons are in Section 7.

W2 / Q1. Why use the original refusal direction instead of random or orthogonal anchors?

Key finding: Where the 1.7B model has room to improve, overlap with the refusal direction predicts held-out-wrapper generalization and a better safety–usability balance. The random anchor is a partial-overlap control: it preserves an initial overlap c₀=0.785 (Sec. 7.2) and therefore has an overlap of approximately c₀²≈0.62 with the reference refusal direction. The orthogonal anchor has zero overlap.

Existing anchor ablation, Qwen3-1.7B (ood834 refusal; Table 3, mean of 3 seeds):

AnchorRealRandom (partial overlap)Orthogonal
Refusal (%) ↑81.071.155.9
Overlap with reference direction1.00~0.620

Source: Table 3, ood834 column of the 1.7B block (rows "AXIS (real anchor)", "AXIS, random anchor", "AXIS, orthogonal anchor"); overlaps from the definitions in Sec. 7.2.

The ordering holds across seeds, and all nine paired contrasts are significant (Sec. 7.2).

New intervention: rotation-only training, with no refusal demonstrations (new; Sec. 7.2, paragraph "Rotation without refusal demonstrations").

Qwen3-1.7B anchorharm334 refusal ↑ood834 refusal ↑Plain benign over-refusal ↓XSTest over-refusal ↓
Real57.587.98.724.9
Random / partial overlap42.584.17.323.7
Orthogonal64.496.046.751.4

Source (new): Rotation-only anchor comparison reported in this response; the real-anchor result is also described in Sec. 7.2. Bold: best per column.

Compared with the partial-overlap anchor, the real anchor improves harm334 refusal by 15.0 points (Holm-adjusted p=3e-6) and ood834 refusal by 3.8 points (p=0.02), with 1.4 and 1.2 points more over-refusal on plain benign requests and XSTest, respectively (the rotation-only results in this response). Although the orthogonal anchor yields higher attack refusal, it also refuses many more benign requests: 46.7% versus 8.7% on plain benign requests and 51.4% versus 24.9% on XSTest.

At larger scales, anchor variants already achieve high ood834 refusal (≥97%; Table 3, 4B and 9B blocks), making safety rates less informative for distinguishing directions. The revised claim therefore focuses on the setting where direction choice produces measurable differences in generalization. We report the complete larger-model anchor results in Table 3.

W3 / Q2. Applying AXIS to languages without a classical–modern distinction

Key finding: AXIS requires matched harmful/benign contrasts across two input conditions—not a classical–modern language pair. The objective in Section 5 can be applied using the following recipe:

  1. Reference condition: Choose bare requests in a language where the model reliably refuses harmful content (e.g., English).
  2. Target condition: Use the same intents expressed in the target language, dialect, register, or encoding.
  3. Estimate and align: Compute the harmful-minus-benign reference direction and optimize the target-condition contrast toward it using the AXIS rotation objective.

Existing and new evidence supports the feasibility of this setup:

EvidenceResultWhere
English ↔ modern Chinese refusal-direction cosine0.865 (1.7B); 0.899 (4B)new; reported here
Held-out English jailbreak displacementWithin 6° of the angle measured against the Chinese anchor at 1.7Bnew; the geometry results in this response
Zero-shot English jailbreak strict refusal, 1.7BBase 8.8% → AXIS 43.8% (λ=0.1) / 57.6% (λ=1)new; Sec. 7.3, the English-transfer results in this response

Thus, the method generalizes conceptually to any setting with a reliable reference condition and a matched target condition. The procedure above explains how to adapt the objective to other languages; those adaptations still require evaluation.

Reviewer_KVvbOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper uses feature vector analysis to understand how narrative wrapping make jailbreaking success, under different languages, then propose a defense method, including an additional rotation loss for feature level regularization. It conducts experiments on Qwen and GLM and report improved safety-utility tradeoff.

Critical strengths

Some intresting interpretations of narrative wrapping in feature space.

Critical weaknesses
  1. The training process directly involves narrative wrapping, which means it is hard to justify if: (a) the model's stronger defense on narrative wrapping is gained from the feature rotation (where no narrative wrapping is in training set) or, (b) the defense is gained from supervised training with some potential semantic correlations between narrative wrapping from training to test set. It would be convincing if the method is purely doing representational defense, without making model see the narrative mapping during training phase.

  2. The design of the defense method seems only targeting on narrative mapping, however the defender should be unaware of what kind of jailbreak methods adversary can use, so the actual usefulness of this defense is questionable. A broader defense, or a better abstraction for other jailbreak methods may be better contribute to the current jailbreak robustness research.

  3. The logic of rotation loss is confusing. Rotation loss is applied on bare Classical Chinese, this is a language transform, rather than narrative wraping, so the logical linkage from aligning refusal direction for bare classical chinese to better defense on narrative wraping is not persuasive enough, explain it from theoretical perspective, or making it intuitively understandable would be better.

Questions for authors
  1. What is the logical linkage between rotation loss on bare classical chinese to better defense on narrative wraping? This is to make the story easier to understand.
  2. Do other popular defenses follow similar feature deviations as narrative wrapping? Since this is a possible way for generalizable defense of your method.
  3. If the training set does not include narrative wrapping, how the defense could be like? Because there have been plenty of defense methods that are zero-shot already, and achieve good jailbreak robustness on multiple jailbreak attacks. (CircuitBreaker, LAT, etc. ), this is to assess if this paper has solid contribution on jailbreak defense research.
6 comments

Thank you for the constructive questions. We have added direct wrapper-free controls, evaluations on 11 external English jailbreak families, geometric analyses, and comparisons with Circuit Breakers and LAT. Together, these experiments clarify what rotation learns and where the robustness gains originate.

W1. Does robustness come from rotation or from seeing wrappers in supervised training?

Key finding: Robustness transfers even when training contains no wrappers; rotation is particularly effective for generalizing wrapped preference supervision without severe over-refusal. We tested two complementary wrapper-free designs.

Experiment A: Rotation only, no wrappers and no refusal demonstrations (new; Sec. 7.2, paragraph "Rotation without refusal demonstrations"). Only the rotation loss is optimised on bare Classical Chinese requests (λ_pref=λ_cmt=0).

Qwen3-1.7BBaseRotation only
harm334 strict refusal ↑9.657.5
ood834 strict refusal ↑33.287.9

Source (new): Sec. 7.2, paragraph "Rotation without refusal demonstrations"; base and trained outputs were scored in the same batch.

The result directly demonstrates that a representation-level intervention trained on bare prompts can improve refusal on held-out wrapped attacks. Rotation-only over-refusal is 24.9% on XSTest and 8.7% on plain benign requests (Sec. 7.2).

Experiment B: Wrapper-free versus wrapper-based preference training (new; Table 4, the wrapper-supervision results in this response). For the wrapper-free arms, we replaced each wrapped prompt with the corresponding bare Classical Chinese request and tested two seeds.

Training armharm334 refusal ↑ood834 refusal ↑XSTest over-refusal ↓
Base2.122.90.4
Wrapped supervision, no rotation53.039.86.9
Wrapped supervision + rotation74.381.013.1
Wrapper-free, no rotation (seeds 0 / 1)72.2 / 71.098.2 / 98.854.3 / 61.6
Wrapper-free + rotation (seeds 0 / 1)71.3 / 66.296.4 / 95.853.5 / 55.9

Source (new): Table 4, Sec. 7.2; the wrapper-free seeds are listed separately. Rotated arms use λ_rot=1. Bold: best value in each column, base included; no single arm wins every column.

What this resolves:

  • Seeing wrappers is not required for transfer: even wrapper-free NLL-based training reaches 96–99% held-out refusal, though with substantial over-refusal (Table 4).
  • Wrapped supervision provides discrimination: combined with rotation, it reduces XSTest over-refusal to 13.1% while improving in-distribution refusal (Table 4).
  • Rotation enables transfer of wrapped supervision: adding rotation improves harm334 from 53.0% to 74.3% and ood834 from 39.8% to 81.0% (Table 4).

The wrapper-free preference arms retain the rejected responses from the original pairs; the relative contributions of preference loss and chosen-response NLL are not isolated by these arms. The prompt substitution is stated explicitly in Sec. 7.2.

W2. Generalization beyond narrative wrappers

Key finding: AXIS targets a shared representational displacement, not a specific narrative wrapper format. We extended evaluation to 11 external English jailbreak families, including non-narrative methods such as refusal suppression, prefix injection, JSON-style injection, distractors, Wikipedia-with-title, and GCG suffixes. Training uses Classical Chinese wrappers only, while evaluation uses English attacks on 200 harmful intents.

Qwen3-1.7B methodEnglish jailbreak strict refusal ↑English XSTest-safe over-refusal ↓
Base8.86.8
DPO19.3—
DPO + commitment31.2—
AXIS (λ=0.1)43.815.6
AXIS (λ=1)57.622.4
LAT95.362.4

Source (new): Sec. 7.3, Table 6 (pooled refusal, Base to AXIS). English XSTest results are reported here; LAT is from a separate comparison batch, whose base over-refusal is 7.6% (6.8% in the DPO/AXIS batch). Bold: best among trained methods; base shown for reference. The cross-batch LAT comparison is descriptive.

Across the 11 families, AXIS improves pooled refusal by 35.0 percentage points (λ=0.1) or 48.9 percentage points (λ=1) over the base model, with Holm-adjusted p < 1e-34 (the English-transfer results in this response; differences are computed from unrounded rates). The LAT comparison illustrates why joint assessment of refusal and benign usability is necessary. At 4B, base/DPO + commitment/AXIS achieve 48.0% / 89.1% / 91.5% refusal (the English-transfer results in this response), respectively, confirming transfer while showing a smaller incremental AXIS gain at this scale.

W3 / Q1. Why does rotation on bare Classical Chinese help with held-out wrappers?

Key finding: Both register changes and jailbreak transformations displace the harmful–benign contrast away from the refusal direction; rotation learns to preserve that content signal. The contrast remains separable within each transformed input family, which motivates alignment rather than suppressing the representation.

We now provide two direct tests of this explanation:

  1. Training intervention: Rotation-only training uses no wrapped prompts and no refusal demonstrations, yet improves harm334 refusal from 9.6% to 57.5% and ood834 from 33.2% to 87.9% (W1; Sec. 7.2).
  2. Geometric measurement on held-out English attacks: AXIS draws the displaced contrast back toward the reference refusal direction.
Qwen3-1.7B methodMedian angle to refusal direction ↓Median projection onto refusal direction ↑
Base77°21
DPO75°27
DPO + commitment67°121
AXIS (λ=0.1)62°242
AXIS (λ=1)50°476

Source (new): the W3 / Q1 table in this response (medians over the 11 English families); angles also quoted in Sec. 7.3.

This geometric recovery extends to non-narrative attacks: with λ=1, the angle reduction is 43° for Wikipedia-with-title and 41° for GCG, compared with 31–32° for ReNeLLM and DeepInception (the geometry results in this response).

Q2. Do other jailbreak attacks exhibit the same representation shift?

Key finding: Yes. The same displaced-but-separable harmful contrast appears in all 11 tested English attack families and in two encoding attacks.

ModelAngle across 11 familiesHarmful/benign AUCGeometric effect of AXIS
Qwen3-1.7B65–87° (median 77°)0.72–0.99Median moves to 50° (λ=1)
Qwen3-4B54–65° (median 60°)0.69–0.995Median moves to 47°

Source (new): Sec. 4 (added sentences) and the geometry results in this response; 1.7B at layer 24, 4B at layer 19, 450 harmful and 450 benign intents each.

At 1.7B, base64 is partially realigned (82° base; 69° with AXIS λ=1; the geometry results in this response). rot13 remains displaced because the base model does not decode it. These measurements clarify the scope of the representational phenomenon and distinguish a successfully decoded attack from an undecoded encoding.

Q3. Comparison with existing zero-shot representation defenses

Key finding: AXIS has higher refusal rates on the evaluated wrapped attacks than the reproduced Circuit Breakers and LAT baselines; LAT also exhibits substantial over-refusal. We added comparisons on Qwen3-1.7B, with judging-batch boundaries noted below.

Methodharm334 refusal ↑ood834 refusal ↑English jailbreak refusal ↑English XSTest-safe over-refusal ↓
Base2.122.98.87.6
Circuit Breakers15.946.439.2—
LAT15.672.295.362.4
AXIS (λ=0.1)68.687.943.815.6
AXIS (λ=1)74.381.057.622.4

Source (new): Defense comparisons reported here; English refusal for Base through AXIS also appears in Table 6. The English XSTest column uses the language-matched results in W2: the base reference is from the LAT batch and AXIS rates are from the DPO/AXIS batch. Those cross-batch usability comparisons are descriptive. The 0.4%, 8.2% and 13.4% rates previously shown here belong to the Classical Chinese XSTest evaluation and are excluded from this English column. Bold: best among trained methods; base shown for reference.

For LAT, English XSTest over-refusal reaches 62.4%; the English-language comparison in W2 reports 15.6% and 22.4% for the two AXIS doses. The reproduced Circuit Breakers model primarily increases incoherent or off-topic outputs rather than firm refusals: incoherent benign outputs rise from 11% to 30%. Its implementation follows the authors' published configuration and is reported here as a best-effort port. Because SRR includes off-topic outputs, these additional comparisons require joint assessment of refusal quality and preserved benign behavior.

Thank you for the detailed rebuttal and the substantial additional experiments. I appreciate the authors' efforts to address my concerns. The new results clarify several aspects of the method, although some questions remain.

Concerns that have been addressed

1. Separating the effect of rotation from wrapper-based supervision (W1).

The new rotation-only experiment is informative. It demonstrates that optimizing the representation-level rotation objective on bare Classical Chinese inputs, without wrapper exposure or refusal demonstrations, can substantially improve refusal on wrapped attacks. This addresses my initial concern that the improvement might be entirely attributable to supervised exposure to narrative wrappers.

The additional wrapper-free versus wrapper-based preference training experiments also help clarify the respective contributions of representation alignment and supervised training.

2. Generalization beyond narrative wrapping (W2/Q2).

The evaluation on 11 additional English jailbreak families, including non-narrative attacks, provides meaningful evidence of cross-attack generalization. The comparisons with Circuit Breakers and LAT further strengthen the empirical evaluation. I consider this concern largely addressed, although the incremental benefit of rotation appears to depend on model scale.

Remaining concerns

1. The necessity of wrapper-based supervision remains unclear (W1/Q3).

The new experiments show that wrapper-free training can achieve very high refusal rates on held-out wrappers, although with substantial over-refusal. The authors argue that wrapper-based supervision helps preserve the safety–utility balance.

However, this does not establish that narrative wrappers are specifically necessary for learning such discrimination. Could similar results be achieved by combining rotation with ordinary harmful/benign preference data, particularly benign examples near the safety decision boundary, without any wrapper-based supervision?

If so, wrapped input is only for avoiding over-refusal, I am suspecting wrapped input still plays an important role for safety, otherwise why not using safety boundary data for this purpose, which is the standard way for other defense methods.

2. The mechanism connecting bare-input rotation to wrapper robustness remains insufficiently explained (W3/Q1).

The new geometric analyses show that AXIS brings harmful–benign contrasts from held-out attack families closer to the reference refusal direction. Together with the rotation-only experiment, these results provide stronger empirical support for the method.

However, they do not fully explain why aligning the population-level harmful–benign contrast on bare Classical Chinese inputs should improve robustness to substantially different narrative or adversarial transformations.

In particular, the observed geometric recovery remains correlational. It is not yet clear whether the alignment itself drives the improved refusal behavior or whether it accompanies other changes induced by training, such as a broader shift in refusal propensity.

A more explicit mechanistic explanation, or an intervention that distinguishes these possibilities, would strengthen the central claim.

Overall assessment

The rebuttal has meaningfully improved the paper, particularly through the wrapper-free controls and cross-attack evaluations. I acknowledge that several of my original experimental concerns have been addressed.

My remaining reservation primarily concerns the interpretation and technical novelty of the proposed defense. The results demonstrate that AXIS can improve robustness, but the unique necessity of wrapper-based supervision and the mechanism underlying transfer from bare-input rotation to unseen attacks remain insufficiently established.

I would appreciate the authors' clarification on these two points.

Reviewer_7M6cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the vulnerability of safety-aligned Large Language Models (LLMs) to narrative wrapping and register-shift attacks across different languages and registers (English, Modern Chinese, and Classical Chinese).

The authors introduce GUISE, a multi-lingual (English, Modern Chinese, and Classical Chinese) benchmark including prompts, harmful and benign responses.

Through representation analysis, the paper reveals that narrative wrapping shifts the representation of harmful requests away from the model's refusal axis. To address this issue, the authors propose AXIS, a defense mechanism with rotation and commitment objectives working with DPO.

Critical strengths

Strengths

  1. The paper contributes GUISE, a benchmark encompassing parallel and matched prompts in English, Modern Chinese, and Classical Chinese.
  2. The paper designs AXIS, which effectively rotates representation vectors back toward the refusal direction while reinforcing complete refusal via a commitment term.
Critical weaknesses

Weaknesses

  1. While the benchmark spans three languages/registers, a major portion of the novelty and focus lies in Classical Chinese, which may limit its application domain.
  2. The technical contributions of the proposed components are somewhat limited. Vector alignment and activation steering are already well-explored in related research. The Commitment objective is somewhat somewhat ad-hoc because it uses hyper-parameters such as weighting ‘w’ and the first “k” tokens.
  3. Some of the experimental results in this paper are mixed (e.g., Table 2).
  4. The paper only evaluates its method on Qwen3-1.7B, Qwen3-4B, and GLM-4-9B. Experiments on larger models or other model families are necessary.
Questions for authors

The title of the paper mentioned "CROSS-LANGUAGE", but the core experiments and methods appear to process each language independently. Is there any explicit cross-lingual transfer mechanism to investigate? If the evaluation simply tests models on multiple individual languages independently, would framing the work as "Multi-Lingual" be more accurate than "Cross-Language"?

2 comments

Thank you for your thoughtful feedback. We have added cross-language and cross-attack evaluations, a sensitivity analysis, and a clearer comparison with existing methods. We summarize the new evidence below and have incorporated the supervision controls, rotation-only result, English-transfer summary and geometric summaries into the revised manuscript; the other new results are reported in this response.

W1. Generality beyond Classical Chinese

Key finding: The representation displacement studied in this paper is not specific to Classical Chinese; AXIS also transfers to English jailbreak attacks that were absent from training. Classical Chinese provides a controlled setting in which the underlying intent remains fixed while the register and wrapper vary.

Attack success under the same wrapper (Qwen3-1.7B; Table 1, strict judge, wrapped prompts):

Language/registerEnglishModern ChineseClassical Chinese
Attack success (%)89.493.095.7

Source: Table 1, row "Wrapped success" (strict judge).

We additionally evaluated 11 English jailbreak families (eight StrongREJECT templates, GCG transfer suffixes from HarmBench, ReNeLLM, and DeepInception), plus two encoding attacks. These attacks show a consistent geometric signature: the harmful–benign contrast remains distinguishable but shifts away from the refusal direction.

Geometric measurementQwen3-1.7BQwen3-4B
Angle to refusal direction across 11 families65–87° (median 77°)54–65° (median 60°)
Split-half noise floor7.4°7.3°
Within-family harmful/benign separability (AUC)0.72–0.990.69–0.995

Source (new): Sec. 4 and the aggregate geometry results in this response. Estimator as in Sec. 4: last prompt token, layer 24 at 1.7B and layer 19 at 4B, 450 harmful and 450 benign intents.

At 1.7B, projections onto the refusal direction fall to 7–178 for the English attacks (reported here), compared with 761 for bare modern Chinese requests (Sec. 4). The encoding attacks show comparable displacement (base64: 82°; rot13: 79–87°; the geometry results in this response). Thus, the phenomenon extends beyond Chinese registers and narrative wrappers.

Zero-shot English transfer (new). AXIS was trained only on Classical Chinese wrappers and evaluated on 11 English jailbreak families constructed from 200 harmful intents (template-based attacks, without LLM rewriting).

Qwen3-1.7B methodBaseDPODPO + commitmentAXIS (λ=0.1)AXIS (λ=1)
Pooled strict refusal (%) ↑8.819.331.243.857.6

Source (new): Sec. 7.3, Table 6; pooled results are reproduced here.

These results demonstrate transfer beyond the training language and wrapper distribution. The same experiment at 4B shows that commitment training already transfers strongly (base 48.0%, DPO + commitment 89.1%, AXIS 91.5%; new, the English-transfer results in this response).

W2. Technical novelty and the choices of ω and k

Key finding: AXIS learns content-dependent alignment during training, rather than applying a fixed inference-time steering vector. Our intervention study in Section 4 ("Can the change be corrected without retraining?") shows that rotating the wrapped harmful contrast onto the refusal direction at constant norm does not restore refusal. A fixed additive shift can restore refusal, but also produces nonsensical outputs for approximately one-third of benign requests. This motivates learning a content-sensitive correction instead of a constant shift.

Unlike methods that reroute harmful activations (Circuit Breakers), push them away (RepBend), or anchor a fixed refusal direction during fine-tuning (ProCon), AXIS aligns the harmful-minus-benign contrast under distribution shift. On plain benign requests at 1.7B, over-refusal is 0.3% for the base model and 5.7% / 6.3% for AXIS (λ=0.1 / 1) (new; defense-comparison results reported in this response).

New sensitivity analysis (reported in this response). We varied k ∈ {2, 4, 8} with ω=2 and ω ∈ {1, 2, 4} with k=4, using a separate training pool (rotation enabled, seed 0).

MeasurementRange across tested settings
harm334 strict refusal ↑94.0–97.3%
ood834 strict refusal ↑93.8–95.9%
XSTest over-refusal ↓14.7–18.8%

Source (new): the sensitivity results in this response.

Only one of 16 paired comparisons with the default (k=4, ω=2) remains significant after Holm correction (the sensitivity results in this response). In particular, ω=1—equivalent to a length-normalized chosen-response NLL, as in RPO / DPO+NLL—performs within 1–2 points of the default. Accordingly, we now characterize the commitment term as a stable NLL anchor and identify representation alignment as the method's distinguishing component.

The rotation ablation (new; Table 4) directly quantifies its contribution when combined with wrapped preference training:

Qwen3-1.7B trainingharm334 strict refusal ↑ood834 strict refusal ↑
DPO + commitment, without rotation53.039.8
DPO + commitment + rotation (λ=1)74.381.0

Source (new): Table 4 (rows "Wrapped, Off" and "Wrapped, On (AXIS)"), Sec. 7.2; Qwen3-1.7B, λ_rot=1, same batch.

This establishes the specific role of rotation in generalizing wrapped preference supervision to held-out wrappers.

W3. Mixed results and the safety–usability trade-off

Key finding: AXIS delivers a strong balance between attack refusal and avoiding over-refusal. The apparent differences between methods in Table 2 are clarified by the two-dimensional safety–usability view (Figure 2). At 1.7B, alpha-DPO and Booster achieve higher refusal at the cost of 67.3 and 42.0 additional points of over-refusal relative to AXIS (Table 2, xs column; Sec. 7.1), while SimPER and DOOR preserve usability largely by declining to refuse harmful attacks.

Model sizeAXIS composite-score lead over strongest external baselineConfidence interval
1.7B+14.5 points[11.1, 18.1]
4B+3.0 points[1.0, 5.2]
9B+1.4 points[-0.2, 3.0]

Source: Table 2, Comp column (AXIS against Booster at 1.7B, SimPER at 4B and 9B); paired-bootstrap intervals in App. M.

AXIS ranks first on all three models for usability weights between 0.4 and 0.7 (App. M). This comparison emphasizes the joint safety–usability objective rather than refusal alone.

W4. Model scale and family coverage

Key finding: The empirical evaluation already covers two model families and three model sizes. Qwen3-1.7B, Qwen3-4B, and GLM-4-9B (Tables 2 and 3) establish the main findings across architectures and scales. The revised manuscript reports model-wise gains and refrains from claiming a scale-invariant advantage. The stronger, statistically resolved improvements are observed at 1.7B and 4B; the 9B result is reported with its confidence interval above (Table 2, Comp column; App. M). Further scaling and additional families are natural follow-up evaluations.

Q. “Cross-Language” versus “Multi-Lingual”

Key finding: The method explicitly trains on one language/register and transfers to another; “Cross-Language” describes the evaluated setting. The study evaluates three transfer paths:

  1. The refusal anchor is estimated on modern Chinese and applied when training on Classical Chinese.
  2. Adapters trained on Classical Chinese transfer to modern Chinese and English (Table 5).
  3. Those adapters also transfer zero-shot to 11 English jailbreak families (new evaluation above; Sec. 7.3).

The English/modern-Chinese refusal directions are also closely aligned (cosine 0.865 at 1.7B; 0.899 at 4B; new measurement reported here). Cross-register results are reported in Table 5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.