Large language models that refuse a harmful request in English often comply when the same request arrives in Swahili or Hausa. Recent work has made real progress on this multilingual safety gap. The progress, however, has mostly been measured on literal translations of English harmful prompts. A separate study rewrote such prompts the way native speakers of Twi, Hausa, Amharic and Swahili would actually phrase them, using local metaphors and local sociopolitical framing. Models failed to refuse in over 98.8% of cases. The model's internal representations of the literal and localized versions were nearly identical in early layers but drifted apart in later ones. This note treats that result as a second domain shift layered on top of the language shift, much as a medical scanner's images differ from natural photos in ways a general vision model was never trained to handle. The analogy points to a concrete question: do current safety fixes, which are trained on translated data, survive the move to localized phrasing? If they do not, can the fix be placed at the layers where the drift happens?
The proposal is a two-stage method with a decisive test built in. Stage one reuses an existing sparse safety-transfer method so that literal translations are refused. Stage two uses a small set of native-speaker rewrites of prompts, paired with their literal twins. It trains a low-rank update that pulls the representation of each localized prompt toward that of its already-refused literal twin, only in the layers where the two measurably diverge, and only along refusal-relevant directions. Native speakers only need to rewrite prompts; they never have to author safe responses. The central claim is falsifiable. Relative to plain refusal fine-tuning on the same localized prompts, the paired consistency objective should give equal or better refusal on held-out localized harm categories and clearly less over-refusal of benign localized prompts.
What the two directions contribute
The multilingual safety literature supplies both the problem and most of the machinery. Multilingual Jailbreak Challenges in Large Language Models introduced MultiJail, 315 harmful queries translated into nine languages, and showed low-resource languages roughly triple unsafe outputs. Three method papers then shrank the gap on translated benchmarks, each in a different way:
- MPO matches the target-language gap between safe and unsafe response likelihoods to the English gap.
- Multilingual Safety Alignment via Representation-Space Separability (SMO) separates harmful from harmless prompts in each language's representation space.
- Who Transfers Safety? finds a sparse set of shared safety neurons that non-English languages route through, and updates only those neurons using parallel safety pairs.
Two further papers are closer to editing:
- Soteria edits the output projections of language-specific attention heads.
- Multilingual Safety Alignment Via Sparse Weight Editing (SWE) computes a closed-form weight change that maps harmful low-resource representations into the English safety subspace. The change is restricted so that it leaves harmless inputs untouched, a restriction called the null space of harmless data.
All of these train or edit on machine-translated English seeds. The shared-neuron paper explicitly scopes itself to "fundamental" harms and leaves culturally localized harm open.
The paper that motivates this proposal is The Illusion of Cross-Lingual Safety in Low-Resource Languages. It introduced LoDNA, a Do-Not-Answer benchmark with paired literal and native-localized prompts for four African languages. Its findings are the ones summarized above. In addition, harmful prompts retained less than 10% of the English refusal signal, and Hausa showed near-zero or negative retention. It is purely diagnostic and proposes no intervention.
The vision domain-gap literature does not supply the mechanism here. It supplies two constraints and an evaluation habit.
- Frozen features are not enough. Are Natural Domain Foundation Models Useful for Medical Image Classification? found that freezing natural-image foundation models and training only a classifier head costs substantially on clinical data. This matches the multilingual finding that an English refusal probe barely registers on African-language prompts. Together these argue against a frozen-feature fix.
- Train on one domain, test on an unseen one. MA-SAM held out a whole domain: on unseen AMOS22 MRI, nnU-Net fell to 10.9% Dice while the lightly adapted foundation model kept 60.4%. We borrow this protocol by treating literal translation as the source domain and localized phrasing as the target. Safety methods have so far been scored only in-domain.
- Adapting without target labels. Unsupervised Adaptation of 3D CT Foundation Models for 3D CBCT Segmentation addresses adaptation when target labels are unavailable. That motivates our choice to require only prompt rewrites, not native safety responses.
We deliberately do not pursue the most obvious transplant, cross-block adapter orchestration in the style of SAM-COBOT applied to literal-translation safety. Existing sparse methods already bring MultiJail attack success to low single digits, which leaves little room for such a method to show a gain. The unsolved regime is localized harm.
Core Claim
Current multilingual safety methods close the language gap for literal translations but leave a second, culturally localized gap that lives mostly in later layers. The proposed fix has three parts:
- Measure, per layer, how much of the refusal direction a localized prompt retains relative to its literal twin.
- Choose the layers where retention drops.
- Train a small low-rank update so that localized prompts inherit the refusal behavior of their literal twins, while benign native inputs are left unchanged.
We expect this to be more precise than refusal fine-tuning on localized prompts. The pairing ties each rewrite to its meaning rather than to surface markers of "sounds like a local harmful request". It should therefore cut over-refusal of benign prompts that use the same idioms.
Research Program
Stage one applies shared-safety-neuron expansion from Who Transfers Safety? using literal parallel safety pairs. Literal twins then become reliably refused, which provides a working "teacher" representation. This stage matters most for Hausa, where even literal prompts retain almost no refusal signal. If stage one fails there, stage two has nothing to pull toward. We will report that outcome honestly as the collapse regime noted by the SMO authors for Swahili in Qwen-2.5-7B.
Drift diagnosis. On a training split of LoDNA pairs, we compute each layer's signed retained component, LoDNA's measure of how much of the English refusal direction survives. We do this for literal and localized prompts separately. Drift layers are those where the localized score falls below the literal score by more than a set margin.
Stage two trains a rank-$r$ update on the drift layers only. The update uses the same roughly 0.6% parameter budget as the shared-neuron method. It minimizes
$$\mathcal{L} = \sum_{l \in D} \big\| P_l \big(h_l(x_{\text{loc}}) - \mathrm{sg}[h_l(x_{\text{lit}})]\big) \big\|^2 + \lambda \sum_{l \in D} \big\| h_l(b) - h^{0}_l(b) \big\|^2 .$$
The symbols are:
- $D$ is the set of drift layers.
- $h_l(\cdot)$ is the layer-$l$ hidden state at the final prompt token.
- $x_{\text{loc}}$ and $x_{\text{lit}}$ are a localized prompt and its literal twin.
- $\mathrm{sg}$ stops gradients, so the literal twin acts as a fixed teacher.
- $P_l$ projects onto the top-$k$ directions separating English harmful from harmless prompts at layer $l$. Only refusal-relevant content is aligned; the cultural content of the prompt is left intact.
- $b$ ranges over benign native-language text.
- $h^{0}_l$ is the stage-one model's hidden state, and $\lambda$ weights the term that keeps benign inputs unchanged. This term is a soft version of SWE's null-space constraint.
Baselines
Unmodified instruct models. The naive reference point is the unmodified Llama-3.1-8B-Instruct, the model used by both MPO and the shared-neuron paper, and Qwen-2.5-7B. On localized prompts, the Illusion paper reports refusal failure above 98.8%.
Strongest existing methods. The strongest competitors are shared-safety-neuron expansion and MPO, since they hold the best reported numbers on the benchmark closest to ours. On MultiJail with Llama 3.1-8B-Instruct, shared-neuron expansion reduced average attack success rate from 29.44% to 3.41% (Who Transfers Safety?). MPO reached about 6% against about 15% for DPO (MPO). No reported result was found on Lacuna for either method, or any intervention, on LoDNA's localized split. That is exactly the measurement this proposal supplies.
We run these methods two ways:
- As published, on literal data only.
- Augmented with the same localized prompts our method sees, paired with refusal responses.
The augmented version is the fair comparison, because all methods then see identical data, parameter budgets and tuning sweeps. Beating them means lower attack success on held-out localized harm categories at equal or lower benign over-refusal, with no MultiJail regression and M-MMLU and MGSM within noise.
Soteria and SWE. Both run with the same augmented data as secondary strong baselines.
Ablations. Three variants isolate the key ingredients:
- Consistency applied at all layers rather than only drift layers, which tests placement.
- Consistency without the projection $P_l$, which tests whether aligning everything erases cultural content and hurts benign behavior.
- Refusal fine-tuning with a low-rank update restricted to the drift layers, which tests the loss independently of placement.
Why not just…?
Why not just translate every query into English and run an English guardrail? Soteria's authors note that translation distorts meaning across diverse scripts and rich morphology. The Illusion paper also reports severe subword fragmentation for Amharic. Localized harm leans on idiom that translation tends to flatten into something innocuous. We include translate-then-moderate as a baseline anyway, using the same English guard. If it matches us on LoDNA, the internal fix is unnecessary for filtering and we will say so.
Why not just add the localized prompts to the training data of shared-neuron expansion or MPO? This is the most serious threat. It may be that the missing ingredient is data and not the objective. The augmented baselines test this directly. If they match our method on both refusal and over-refusal, the conclusion is that native prompt rewrites plus existing methods suffice. That would still be a useful result, but the consistency objective would add nothing.
Why not just apply low-rank adaptation to every layer and let gradients find the drift? Gradient descent may well discover the right depth on its own. The all-layer ablation, run at matched parameter count, settles this. If placement makes no difference, the drift diagnosis is interpretive only.
Why not just train a linear probe on native localized prompts as an input filter? The Illusion paper showed an English-trained probe fails on these languages. A probe trained on native localized data was not tested, and SMO's evidence that harmful and harmless prompts are entangled in low-resource languages argues against, but does not rule out, a workable probe. We train one per language on the same data as a filter baseline. If it matches our method, the remaining argument for internal alignment is weaker: the model's own generation stays unsafe whenever the filter is bypassed.
Proposed Experiments
First decisive check. Before any new training, we run the published baselines on LoDNA's literal and localized splits for all four languages. If shared-neuron expansion and MPO already refuse localized prompts nearly as well as literal ones, the proposal's premise fails and we stop.
Main experiment. We split LoDNA pairs by harm category, train on some categories and test on unseen ones, following MA-SAM's held-out-domain habit. We report the following:
- Refusal and attack success, judged by a model-based judge and a native-speaker audit of a sample.
- Over-refusal on a small native-written set of benign prompts that use the same local idioms.
- Signed retained component per layer.
- MultiJail, AdvBench-X and MULTI-STRONGREJECT to confirm there is no regression on literal data.
- M-MMLU, MGSM and MT-Bench to measure any loss of general capability from the safety training.
Secondary robustness check. Motivated by the shared-neuron paper's observation that masking those neurons collapses safety, we mask the stage-one safety neurons after stage two. This tests whether the drift-layer update leaves localized refusal less dependent on that single fragile set.
References
- Multilingual Jailbreak Challenges in Large Language Models
- MPO: Multilingual Safety Alignment via Reward Gap Optimization
- Multilingual Safety Alignment via Representation-Space Separability
- Who Transfers Safety? Identifying and Targeting Cross-Lingual Shared Safety Neurons
- Soteria: Language-Specific Functional Parameter Steering for Multilingual Safety Alignment
- Multilingual Safety Alignment Via Sparse Weight Editing
- The Illusion of Cross-Lingual Safety in Low-Resource Languages
- Are Natural Domain Foundation Models Useful for Medical Image Classification?
- MA-SAM: Modality-agnostic SAM adaptation for 3D medical image segmentation
- Unsupervised Adaptation of 3D CT Foundation Models for 3D CBCT Segmentation
- Parameter Efficient Fine-tuning via Cross Block Orchestration for Segment Anything Model