acceptodds
Under review as a conference paper at ICLR 2027

Directing Safety Signals to the Appropriate Neurons: Multilingual Safety Alignment through Selective Neuron Masking

Abstract

The safety behavior of large language models varies considerably from one language to another, and prior work has traced such responses to a compact group of internal safety neurons. Current approaches to multilingual safety alignment usually formulate their optimization targets behaviorally, through loss terms or reward signals, yet seldom determine which of the model's safety-related internal representations actually absorb this training pressure. To address this gap, we introduce SNPO, a preference optimization scheme that operates at the level of individual neurons for aligning safety across languages. The method begins by locating the safety neurons tied to each language and then divides all languages into two sets according to how much training data each has. Optimization proceeds by alternating between these sets: on any given step, the safety-neuron activations of one set are masked and their weights held fixed, so that the alignment pressure is funneled onto the neurons belonging to the other set. The assignments are subsequently reversed, allowing the neurons of each set to receive focused optimization in turn. Evaluations spanning seven models and eight languages demonstrate that SNPO lowers how often unsafe outputs occur without sacrificing multilingual capability, and the improvements are most pronounced for languages with limited resources. Examining the neurons directly, we find that after SNPO those safety neurons that had responded to only one language begin firing across several. Taken together, our findings indicate that successful multilingual safety alignment depends not merely on carefully crafted behavioral preferences, but equally on directly governing which internal neurons end up receiving the resulting signal.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multilingual Safety Alignment, Safety Neuron, Preference optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_hyMpOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Safety-Neuron Preference Optimization (SNPO) for multilingual safety alignment. It identifies candidate safety-related neurons from activation differences between harmful and benign prompts, groups languages by resource level, and alternately masks one group’s safety neurons while updating only the other group’s exclusive neurons using multilingual preference data. All neurons are restored at inference. The main contribution is a neuron-level training strategy for directing safety updates across language groups. Evaluation covers seven models and eight languages using MULTI-STRONGREJECT for safety and MGSM/M-MMLU for utility, with comparisons against MPO, SWE, and additional training baselines. Results show consistent average safety improvements, while utility varies by model. Ablation and pruning studies support the strategy’s effectiveness and the functional importance of some neurons that become shared across languages, although safety specificity and over-refusal remain insufficiently validated.

Critical strengths

Strengths: 1.Consistent safety improvements across multiple models. SNPO achieves the largest average reduction in unsafe responses among the compared methods on all seven evaluated models across two model families and eight languages. 2.A distinctive training intervention supported by meaningful ablations. SNPO combines language-group-specific activation masking with restricted parameter updates to encourage cross-lingual safety learning.

Critical weaknesses

Weaknesses: 1.Insufficient validation that the detected neurons specifically mediate safety behavior. The detection procedure selects neurons whose activations distinguish harmful from benign prompts, which may also capture differences in topic or task rather than safety functionality. The post-training pruning experiment provides useful functional evidence for the Specific→Shared subset, but does not establish the safety specificity of the original detected sets; it also lacks matched functional controls and utility measurements after pruning. This matters because identifying language-specific safety carriers is central to both SNPO’s design rationale and its mechanistic interpretation. 2.Missing evaluation of over-refusal leaves the safety–helpfulness trade-off unresolved. The paper measures unsafe responses to harmful prompts but does not evaluate whether SNPO also rejects legitimate requests, particularly benign requests involving safety-sensitive topics across the target languages. MGSM and M-MMLU do not directly measure this behavior, and Table 1 already shows a substantial MGSM decline on Qwen2-1.5B, from 20.36 to 14.89.

Questions for authors

Did you validate the detected neuron sets in the original models through ablation, using controls matched for neuron count and layer distribution, while measuring both safety and utility? Such evidence would help distinguish safety-related functionality from general topic or task processing. Did you evaluate refusal rates on benign requests across the target languages, particularly safety-sensitive but legitimate requests? If so, how does SNPO compare with the original models and baselines? This would clarify whether the ASR improvements reflect better safety discrimination or broader refusal.

5 comments
AuthorsOpenReview ↗

Response to Reviewer hyMp(1/3): Weaknesses1 - Part 1

We thank the reviewer for recognizing the presentation and empirical gains of our method, and for the constructive comments. Below we respond to each weakness and question point by point, supported by additional experiments.

W1: Insufficient validation of the safety specificity and functional importance of detected neurons.

We thank the reviewer for this comment. We address it from both the detection design and causal evidence.

(i) Why the detected neurons reflect safety rather than topic.
The contrast is topic-overlapping by construction. A substantial part of our harmful set is HarmfulQA, whose 1,960 questions are organized under 10 academic topics (e.g., Science and Technology, Mathematics and Logic, Health and Medicine, Business and Economics), the same kinds of domains covered by our benign reasoning questions. Both sides therefore span diverse and overlapping topics, and the most consistent factor separating them is harmful intent rather than subject matter.

The selection criteria filter out topic-specific neurons. A neuron is selected only if its activation shift is consistent across the entire, topically diverse harmful set, as required by both Cohen's d against the pooled variance and the layer-wise z-scored mean shift. A topic-specific neuron (e.g., chemistry) fires on only a small fraction of harmful prompts and also on benign prompts of the same topic; it therefore has a small mean shift and a large variance, and is not selected. The surviving neurons respond to a factor shared across all harmful topics, most plausibly harmfulness. Consistently, Wang et al. [1] show that neurons detected separately on ten harm categories overlap substantially.

Language-specific sets are not driven by topic. Detection data for all languages are parallel translations with identical topic distributions, so differences among the language sets NℓN_\ell cannot come from topic. They reflect how each language's safety behavior is carried internally, which is exactly what SNPO exploits.

Residual differences in task format between the two sets cannot be fully excluded by design; we therefore verify the safety function of these neurons causally below.

(ii) Language-specific causal intervention.
Beyond these design-level arguments, our claims rest on causal evidence rather than on the assumption of pure safety functionality. Since identifying language-specific safety carriers is central to SNPO, we directly test whether each language's detected neurons causally support that language's safety. On the original (untrained) Qwen2.5-1.5B, for each language, we deactivate that language's detected safety neurons (setting their activations to zero) and measure the unsafe responses in that language; for utility, we deactivate the union of all languages' safety neurons and measure M-MMLU. As a control, we mask the same number of randomly selected neurons as the union, matched in layer distribution; since this exceeds the size of any single language's set, it is a conservative control. Unsafe responses out of 313 and M-MMLU accuracy (%) are reported:

SettingENZHVIJATHIDBNHEM-MMLU
With safety neurons60353663816021015942.85
Without safety neurons27027024524221923022917435.75
Layer-matched random6932611131136922914238.89

First, deactivating a language's safety neurons causes a drastic collapse of safety behavior in that language (e.g., EN: 60 → 270; ZH: 35 → 270), whereas masking even more randomly selected neurons causes much smaller changes (EN: 69; ZH: 32). This demonstrates that the detected language-specific neurons are strongly and causally associated with the corresponding language's safety capability, rather than the effect stemming from the number of masked neurons. Second, general capability also drops (M-MMLU: 42.85 → 35.75), but far less severely than safety, indicating that these neurons also carry part of the general capability rather than forming a purely safety-dedicated circuit. This is consistent with the reviewer's point on safety specificity. Notably, the increase is much smaller for the low-resource languages BN and HE, whose base unsafe rates are already high, reflecting their weak safety carriers, which is exactly the imbalance SNPO is designed to address.

AuthorsOpenReview ↗

Response to Reviewer hyMp(2/3): Weaknesses1 - Part 2

(Continued from our previous comment.)

(iii) Cross-lingual overlap.
We further compute the Jaccard similarity between the safety-neuron sets of each language pair. The overlap is far from complete (0.052–0.391) and particularly low for Bengali (0.052–0.123); moreover, it is markedly higher among high-resource languages (average 0.308) than among low-resource languages (0.179). This shows that multilingual safety relies on both shared and language-associated neurons, with safety carriers unevenly distributed across languages, which supports SNPO's design of routing alignment signals to different neuron groups rather than treating all safety neurons uniformly.

LanguageENZHVIJATHIDBNHE
EN1.0000.3900.2480.2840.1910.2190.0520.133
ZH0.3901.0000.3330.3890.2730.2660.0720.194
VI0.2480.3331.0000.2790.3600.3910.0800.232
JA0.2840.3890.2791.0000.2560.2320.0800.187
TH0.1910.2730.3600.2561.0000.2930.1230.309
ID0.2190.2660.3910.2320.2931.0000.0710.196
BN0.0520.0720.0800.0800.1230.0711.0000.122
HE0.1330.1940.2320.1870.3090.1960.1221.000

(iv) Utility after pruning.
We supplemented the pruning analysis with M-MMLU (Table 4 in main text), where the random control is matched to Specific→Shared in the number of neurons:

Pruned neuronsAvg-HighAvg-LowM-MMLU
None0.724.3942.17
All-language Shared2.086.3942.63
Specific→Shared4.8713.7442.76
Random0.565.3542.66
Emerged-Safe1.445.7542.22

Pruning any of these safety-neuron sets leaves general capability nearly unchanged (M-MMLU changes by at most 0.59 points). Thus, the safety degradation reported in Table 4 is not a by-product of general capability loss, indicating that these neurons primarily contribute to safety, although we do not claim they are exclusively safety-dedicated.

We will incorporate these results and clarify the scope of our claims in the revised manuscript.

[1] Z. Wang, et al. SafeNeuron: Neuron-level safety alignment for large language models, 2026.

AuthorsOpenReview ↗

Response to Reviewer hyMp(3/3): Weaknesses2 and Questions For Authors

W2: Insufficient evaluation of over-refusal and the safety–helpfulness trade-off

We agree that a low ASR does not necessarily indicate better safety, as it may also result from excessive refusal. MGSM and M-MMLU consist of benign tasks and can indirectly reflect broad over-refusal behavior, since excessive refusal would prevent models from providing correct answers. Across seven models, SNPO changes MGSM by +0.54 and M-MMLU by −0.23 on average, suggesting that general utility is largely preserved. However, these benchmarks do not directly measure over-refusal on safety-sensitive benign requests, and we acknowledge the MGSM decline on Qwen2-1.5B (20.36 → 14.89).

To directly evaluate over-refusal, we additionally collected 20 benign requests involving safety-sensitive topics and used Claude to translate them into the other seven languages, covering all eight target languages. These requests include (1) harmless uses of potentially unsafe keywords (e.g., "kill" or "slash"), (2) definitions of sensitive concepts (violence, robbery), (3) questions involving nonsensical discriminatory premises, and (4) harmless privacy-related questions about public information. All methods were evaluated on the same set of requests.

The average refusal rates (%) across eight languages are reported below (lower is better):

ModelBaseSNPO (Ours)MPOSWEFull-DPO
Qwen2-1.5B20.630.032.531.287.5
Qwen2.5-1.5B21.236.237.539.460.6
Qwen2.5-7B11.916.218.117.5–

–: Full-DPO on Qwen2.5-7B was not trained due to time constraints.

Across all three models, SNPO achieves the lowest average over-refusal rate among the evaluated post-training methods. All post-training methods increase over-refusal relative to their base models, and Full-DPO does so most severely (87.5% on Qwen2-1.5B and 60.6% on Qwen2.5-1.5B). These results suggest that SNPO better preserves helpful responses to safety-sensitive benign requests than the evaluated post-training baselines on this evaluation set, although over-refusal remains a limitation. We will incorporate these results and discuss this limitation in the revised manuscript.

Q1: Validation of detected neurons through ablation with matched controls

Yes. On the original (untrained) Qwen2.5-1.5B, we deactivated the detected safety neurons and, as a control, the same number of random neurons matched in layer distribution, measuring both safety and utility. Deactivating a language's detected safety neurons causes a drastic collapse of safety in that language (e.g., EN: 60 → 270 unsafe responses out of 313), whereas random masking causes only minor changes (EN: 69). M-MMLU decreases from 42.85% to 35.75%, compared with 38.89% under random masking, indicating that these neurons are not exclusively safety-dedicated but contribute disproportionately to safety. Please refer to our response to W1 for details.

Q2: Refusal rates on benign safety-sensitive requests

Yes. We collected 20 benign but safety-sensitive requests and translated them with Claude into all eight target languages. Average refusal rates (%) across languages (lower is better):

ModelBaseSNPOMPOSWEFull-DPO
Qwen2-1.5B20.630.032.531.287.5
Qwen2.5-1.5B21.236.237.539.460.6
Qwen2.5-7B11.916.218.117.5–

–: Full-DPO on Qwen2.5-7B was not trained due to time constraints.

Across all three models, SNPO has the lowest refusal rate among the post-training methods, although it remains higher than that of the original models. Since SNPO achieves the largest ASR reduction while incurring the smallest increase in refusals, its gains reflect better safety discrimination rather than broader refusal. Notably, on Qwen2-1.5B, where MGSM declines, SNPO also shows the lowest refusal rate, suggesting that the decline is not caused by increased refusals. Please refer to our response to W2 for details.

Reviewer_De13OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The core of the paper is SNPO, a DPO variant. It first identifies language-specific safety neurons and roughly splits languages into a high-resource and a low-resource group. During training it masks the safety neurons of one group, forcing the exclusive neurons of the other group to absorb the preference signal, and then swaps the roles and repeats. Experiments on Llama and Qwen show lower ASR than MPO and SWE, more clearly in low-resource languages, while average MGSM and M-MMLU scores change little, with variation across models.

Critical strengths

(1) The idea is not complicated, but it fits well: it is built directly on DPO and adds nothing at inference. The motivation that safety carriers are unevenly distributed across languages also makes sense.

(2) The gains in the main table are indeed large and fairly consistent across the two model families. Low-resource languages see the largest drops, and on average MGSM and M-MMLU do not drop much.

Critical weaknesses

(1) Table 1 relies on a single automatic judge, and I could not find a benign over-refusal evaluation or variability across training runs. Unsafe counts fall sharply on the smaller Qwen models, but MGSM also drops in some cases, particularly Qwen2-1.5B. This leaves uncertainty about whether the safety gains preserve helpful responses beyond the reported utility benchmarks.

(2) The evidence for neuron routing would be stronger across more backbones: the Full-DPO comparison and ablations in Section 4.3 use only Qwen2.5-1.5B. Section 3.3 allows the two steps to be trained independently and merged, while the abstract describes alternating updates; the actual training schedule needs clarification. I also could not locate the promised comparison of restoring all neurons against keeping either group suppressed at inference; the pruning results test different neuron sets.

(3) Several implementation details remain unclear in Sections 3.2 and 3.3 and Appendix C. These include the separability and strength threshold values, exclusive-set sizes, trainable parameter counts, and training duration. The paper explains which activations are masked, but the precise masking implementation would help reproduction.

Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer De13(1/2): Weaknesses (1)

We thank the reviewer for recognizing our method's motivation, empirical gains, and analysis. Below we address the three weaknesses point by point with additional experiments.

W1: Insufficient evaluation of safety–utility trade-offs, including over-refusal, judge reliability, and training variability.

(i) Judge. We additionally evaluated all seven backbones with PolyGuard, a non-Qwen safety judge. The average ASR (%) reduction relative to None, averaged over the seven backbones and the four languages supported by PolyGuard (EN, ZH, JA, TH), is summarized below:

Qwen3GuardPolyGuard
MPOSWESNPOMPOSWESNPO
Avg. Δ vs. None−13.44−10.01−19.66−5.73−2.74−11.54

PolyGuard yields higher absolute ASRs for all methods because it sometimes misclassifies explanatory refusals as compliance. Nevertheless, SNPO achieves the largest ASR reduction under both judges, and the ranking SNPO > MPO > SWE is preserved. Please refer to our response to Q1 of Reviewer XSV6 for more results.

(ii) Over-refusal. We agree that a low ASR does not necessarily indicate better safety, as it may also result from over-refusal. This is one reason why we report MGSM and M-MMLU alongside safety. Both benchmarks consist of benign tasks and can indirectly reflect over-refusal behavior, since over-refusal of legitimate requests would prevent models from providing correct answers. Across the seven models, SNPO changes MGSM by +0.54 and M-MMLU by −0.23 on average, suggesting that general utility is largely preserved. However, we acknowledge that these benchmarks cannot directly measure over-refusal on safety-sensitive but legitimate requests.

To evaluate overly conservative refusal behavior, we additionally collected 20 benign requests involving safety-sensitive topics and used Claude to translate them into the other seven languages, covering all eight languages in our evaluation. These requests span four categories: (1) harmless requests containing potentially unsafe keywords (e.g., "kill" or "slash"); (2) definitions of sensitive concepts (violence, robbery); (3) questions involving nonsensical discriminatory premises; and (4) harmless privacy-related questions involving public information. The refusal rates (%), averaged across eight languages (lower is better), are:

ModelBaseSNPO (Ours)MPOSWEFull-DPO
Qwen2-1.5B20.630.032.531.287.5
Qwen2.5-1.5B21.236.237.539.460.6
Qwen2.5-7B11.916.218.117.5–

–: Full-DPO on Qwen2.5-7B was not trained due to time constraints.

All post-training methods increase over-refusal relative to the base model, while SNPO achieves the lowest refusal rate among them. These results indicate that over-refusal remains a common challenge in multilingual safety alignment rather than an issue specific to SNPO. Notably, Full-DPO, trained on the same preference data, exhibits a higher refusal rate than SNPO, suggesting that routing alignment updates to safety neurons also limits collateral refusal. We will discuss this limitation in the revised manuscript.

(iii) Variability across training runs. We re-trained SNPO with an additional random seed on three backbones. Avg-High / Avg-Low ASR (%) are reported as mean ± std over two runs (MPO and SWE from Table 1; lower is better):

BackboneSNPO (mean ± std)MPOSWE
Llama-3.2-3B1.76±0.23 / 10.42±1.522.56 / 15.730.80 / 13.74
Qwen2-1.5B0.04±0.06 / 1.16±1.070.16 / 3.193.27 / 21.65
Qwen2.5-1.5B0.76±0.05 / 4.90±0.710.72 / 8.715.03 / 25.48

SNPO maintains the lowest mean Avg-Low ASR across these runs, providing preliminary evidence of training robustness.

We further evaluated the inference stability of SNPO on Qwen2.5-1.5B with three sampling seeds:

SeedAvg-HighAvg-Low
10.724.39
20.964.31
30.724.55
Mean ± std0.80±0.144.42±0.12

The results are highly consistent across sampling seeds, indicating that the reported improvements are not artifacts of decoding randomness.

AuthorsOpenReview ↗

Response to Reviewer De13(2/2): Weaknesses (2) (3)

W2: More backbones for ablation, unclear training schedule, and missing inference-time masking ablations.

1. Ablations on an additional backbone.
We repeated the Section 4.3 ablations on Qwen2-1.5B under the same settings. Representative results are reported below (ASR %, lower is better); the full per-language results will be included in the Appendix of the revised manuscript.

VariantAvg-High ↓\downarrowAvg-Low ↓\downarrow
None13.1036.58
SNPO0.000.40
Full-DPO10.9423.48
SS-Shared0.084.47
All-SN-DPO1.046.15
Random-Matched1.449.90
w/o NshN_{sh}-mask1.928.15
Step-1 Only3.1917.09
Step-2 Only1.369.35
Mixed Grouping0.324.63
Corpus-Neuron-Matched0.166.63
Corpus-Neuron-Crossed0.323.04

The results are consistent with those on Qwen2.5-1.5B: SNPO achieves the best Avg-High and Avg-Low among all variants. Full-DPO, which updates all parameters with the same data, remains far behind (10.94% / 23.48%), and restricting updates to all safety neurons (All-SN-DPO) or to layer-matched random neurons (Random-Matched) is also clearly inferior, showing that where the alignment signal is routed matters more than the signal itself. Removing shared-neuron masking or altering the grouping strategy consistently degrades Avg-Low, and the two steps again show asymmetric but complementary contributions (Step-2 Only: 9.35%; Step-1 Only: 17.09%; SNPO: 0.40%). The advantage is most pronounced on low-resource languages, where SNPO reduces Avg-Low from 36.58% to 0.40% while all variants remain at 3.04%–23.48%.

2. Training schedule. Thank you for catching this inconsistency. The term "alternating" was inaccurate and has been removed from the revised manuscript. The two steps are fully independent training runs, each initialized from the same base model; they update disjoint group-exclusive safety neurons, and their parameter updates are merged afterward.

3. Inference-time masking. We thank the reviewer for pointing out that the inference-time comparison promised in Section 3.3 was missing. Unlike the pruning analysis, which tests neuron sets that change after training, this experiment directly compares the SNPO model (all neurons restored) with variants that keep a given safety-neuron group suppressed at inference, using the same number of random neurons as controls. Results on Qwen2.5-1.5B are reported below (ASR % and M-MMLU %).

Condition#MaskedAvg-High ↓\downarrowAvg-Low ↓\downarrowM-MMLU ↑\uparrow
SNPO00.724.3942.17
Mask NAN_A8,30172.1263.1838.17
Random (same size as NAN_A)8,3010.722.8839.95
Mask NBN_B4,98633.1524.8440.51
Random (same size as NBN_B)4,9860.966.1541.73
Mask NshN_{sh}2,96819.5716.7741.03
Mask NAexcN_A^{exc}5,33339.4643.3741.02
Mask NBexcN_B^{exc}2,0180.727.5142.29

Restoring all neurons improves safety compared with keeping either group suppressed, while same-size random masking barely affects safety. The two updated neuron groups jointly support multilingual safety, consistent with the claim in Section 3.3.

W3: Insufficient implementation details for reproducibility.

We thank the reviewer for pointing this out. All details below, together with the full per-model statistics, will be added to Appendix C of the revised manuscript.

1. Detection thresholds. Safety neurons are detected in the gate_proj and up_proj layers. For each language, a neuron is selected if its steering-vector z-score (separability) exceeds 2.0 (with a positive mean activation difference) and its Cohen's d (strength) exceeds 1.0.

2. Exclusive sets, trainable parameters, and training time. Each neuron corresponds to one output channel (one weight row) in gate_proj or up_proj. Across the seven backbones, the high- and low-resource exclusive sets contain 2,194–11,694 and 719–4,500 neurons, respectively, so SNPO updates only 0.53%–0.76% of the model parameters; one training epoch (both steps) takes 140–1,128 minutes. For example, on Qwen2.5-1.5B, ∣NAexc∣|N_A^{exc}| = 5,333 and ∣NBexc∣|N_B^{exc}| = 2,018 (11.3M trainable parameters, 0.73%), and one epoch takes 144 minutes.

3. Masking implementation. During training, we register forward hooks on gate_proj and up_proj to zero out the output channels of the masked safety neurons. All other parameters are frozen, and gradient row masks ensure that only the selected group-exclusive safety-neuron rows are updated. After the two independent steps, only the updated rows are merged into the base model. Masking is applied only during training; at inference, no hooks or masks are used.

Thank you for the detailed response. The PolyGuard results address my concern about relying on a single judge. The new ablations on Qwen2-1.5B and the inference-time masking comparison also help. I will keep my score.

AuthorsOpenReview ↗

Response to Reviewer De13

Thank you for your reply.

We do apologize for posting our response only a few days ago, as the additional experiments took more time to complete. We are glad that the additional results addressed your concerns.

If any concerns remain, we would be happy to address them during the discussion period.
If you feel your concerns have been resolved, we would be grateful if you would consider raising your score.

Many thanks again for your time and effort you invested in reviewing our paper.

Reviewer_XSV6OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SNPO, targeting uneven LLM safety across languages: models refuse harmful requests in English but comply in low-resource ones. It detects language-specific safety neurons, splits languages into high-/low-resource groups, and alternately masks one group’s safety-neuron activations during DPO, forcing alignment gradients onto the other group’s exclusive neurons. Across seven LLMs and eight languages, it cuts unsafe responses more than MPO and SWE baselines—especially on Bengali and Hebrew—while preserving utility. Neuron analysis shows language-specific neurons become shared carriers, confirming they carry the gained multilingual safety.

Critical strengths
  • The paper reframes multilingual safety alignment from “what signal to optimize” to “which neurons receive it,” identifying a genuine, well-motivated gap: safety-neuron carriers are themselves imbalanced across languages.
  • The evaluation is thorough: seven models across two families and multiple scales, eight languages, consistent improvements over MPO and SWE, and targeted ablations isolating each design choice.
Critical weaknesses

The paper’s core claim, i.e., alternately masking language groups drives the gains, is weakened by its own ablations: Step-2 Only already captures a large portion of SNPO’s benefit while Step-1 Only fails badly, suggesting asymmetric reliance on high-resource training; the claimed symmetric routing mechanism thus lacks strong support, and the neuron-reorganization explanation remains purely correlational.

Questions for authors
  • Q1: Since the judge and five of the seven tested backbones are all Qwen-family, could you show that the main results hold with a non-Qwen judge? Human evaluation is better.
  • Q2: How sensitive are the results to the neuron-detection method (Wang et al., 2026) and threshold choices? If detection is noisy, the entire routing premise is fragile.
2 comments
AuthorsOpenReview ↗

Response to Reviewer XSV6 (1/2): Weaknesses

We deeply appreciate the time and effort you invested in reviewing our paper and thank the reviewer for recognizing that SNPO reframes multilingual safety alignment from "what signal to optimize" to "which neurons receive it". Below, we provide point-by-point responses to address each of your concerns.

W1: Asymmetric contributions of the two steps and the role of neuron reorganization

(i) Training procedure. We apologize for the ambiguity. The term "alternating" was an inaccurate choice of wording and has been removed from the revised manuscript. The two steps are fully independent training runs: each is initialized separately from the same base model, and their disjoint parameter updates are merged afterward. The second step does not inherit any parameter updates from the first.

(ii) "Symmetric" refers to the routing design, not to equal contributions.
The observed asymmetric contributions are consistent with our motivation for SNPO and provide further evidence of the differences in safety-related capacity between high- and low-resource language neurons. Given that low-resource languages have fewer identified safety neurons than high-resource languages, such asymmetric gains are expected. Importantly, the two steps remain complementary. Step-2 Only already achieves substantial gains (Avg-Low: 7.35%), while incorporating Step 1 further reduces Avg-Low from 7.35% to 4.39% and Avg-High from 1.12% to 0.72%. The same asymmetric-but-complementary pattern also holds on Qwen2-1.5B (see our response to W2 of Reviewer De13).
Our goal is not to achieve equal gains from the two steps, but that directing alignment signals to different safety-neuron groups and merging their updates improves multilingual safety. The unequal contributions therefore do not contradict the motivation or design of SNPO.

(iii) The neuron reorganization analysis is not purely correlational.
Section 4.3 aims to investigate how SNPO reshapes the safety-related roles of neurons and to validate the functional importance of the identified changes through targeted interventions, rather than to establish neuron reorganization as the direct cause of performance gains. Specifically, pruning the Specific→Shared neurons raises Avg-Low ASR from 4.39% to 13.74%, whereas pruning the same number of random neurons only raises it to 5.35%. This provides intervention-based evidence that these neurons play an important role in maintaining multilingual safety after SNPO.
Moreover, the performance gains are not simply determined by the extent of neuron reorganization. Full-DPO produces more emerged safety neurons (8,686 vs. 2,014) and more Specific→Shared transitions (812 vs. 728) than SNPO, yet performs worse. We have revised and highlighted the corresponding discussion to clarify the purpose of this analysis and avoid overstating its causal implications.

Due to the character limit, our responses to the "Questions For Authors" are provided in a separate comment.

AuthorsOpenReview ↗

Response to Reviewer XSV6 (2/2): Questions for Authors

Q1: Do the main results hold with a non-Qwen judge? -- Answer: Yes

We adopt an automatic judge pipeline because expert human evaluation is highly costly and time-consuming for multilingual research; automatic judging is also common practice in existing multilingual LLM safety work [1], [2], [3]. Qwen3Guard-Gen-8B is one of the few open-source guard models with broad multilingual coverage (119 languages and dialects).

To address this concern, we re-evaluated the main results with PolyGuard-Ministral [4], a non-Qwen safety judge. Since PolyGuard supports only four of our eight target languages (EN, ZH, JA, TH), results are reported on these four languages:

BackboneQwen3GuardPolyGuard
NoneMPOSWESNPONoneMPOSWESNPO
Llama-3.2-1B13.749.278.956.7135.5432.9136.4231.07
Llama-3.2-3B9.116.234.073.9928.4326.9223.4025.08
Qwen2-0.5B65.6531.1552.321.3681.9563.0275.1644.73
Qwen2-1.5B15.730.563.990.1648.4043.2947.6036.42
Qwen2.5-1.5B19.091.445.270.9649.4440.9754.3138.10
Qwen2.5-3B21.097.194.711.4448.1644.8937.5438.18
Qwen2.5-7B7.912.402.960.0836.0235.8634.3533.55
Avg. Δ vs. None–−13.44−10.01−19.66–−5.73−2.74−11.54

PolyGuard yields higher absolute ASRs for all methods because it sometimes misclassifies explanatory refusals as compliance. Nevertheless, SNPO achieves the largest average ASR reduction under both judges, supporting the robustness of our main findings.

[1] W. Zhao, et al. MPO: Multilingual safety alignment via reward gap optimization. ACL 2025.
[2] S, Dan, et al. Multilingual Safety Alignment via Representation-Space Separability. ICML 2026.
[3] J. Liang, et al. Multilingual safety alignment via sparse weight editing. ICML 2026.
[4] Kumar, P., et al. PolyGuard: A Multilingual Safety Moderation Tool for 17 Languages. COLM 2025.

Q2: Sensitivity to the neuron-detection method and thresholds

SNPO does not require the detected neurons to be only safety-related; it only requires that they are strongly associated with safety behavior. We first verify this with a causal intervention experiment on the original (untrained) Qwen2.5-1.5B. For each language, we deactivate that language's detected safety neurons (setting their activations to zero) and measure the unsafe responses in that language; for utility, we deactivate the union of all languages' safety neurons and measure M-MMLU. As a control, we mask the same number of randomly selected neurons as the union, matched in layer distribution; since this exceeds the size of any single language's set, it is a conservative control.

Unsafe responses out of 313 and M-MMLU accuracy (%) are reported:

SettingENZHVIJATHIDBNHEM-MMLU
With safety neurons60353663816021015942.85
Without safety neurons27027024524221923022917435.75
Layer-matched random6932611131136922914238.89

First, deactivating a language's safety neurons causes a drastic collapse of safety behavior in that language (e.g., EN: 60 → 270; ZH: 35 → 270), demonstrating that the detected language-specific neurons are strongly and causally associated with the corresponding language's safety capability. Second, general capability also drops (M-MMLU: 42.85 → 35.75), but far less severely than safety, indicating that these neurons also carry part of the general capability rather than forming a purely safety-dedicated circuit.

In addition, we conducted experiments with different detection thresholds to analyze the sensitivity. We vary the z-score threshold (z) and the Cohen's d threshold (d) around the default setting, and mask each resulting neuron set in the original Qwen2.5-1.5B. ASR and M-MMLU (%) are reported.

The first row is the original model, the bold row is the default detected set.

zdAvg-HighAvg-LowM-MMLU
––15.2640.9742.85
3179.0774.6839.44
2.5180.8375.6438.17
21.581.3977.5637.48
21.2580.9177.7236.60
2184.1978.4335.75
20.7585.0677.5631.04
20.581.5578.2725.10
1.5186.0280.3526.34
1190.5081.6323.55

The results show that the stricter the threshold, the more pronounced the functional specificity of the selected neurons. Since looser sets contain the stricter ones, masking any of them causes a severe safety collapse, but stricter thresholds better preserve M-MMLU (39.44% for z = 3 vs. 35.75% for the default), whereas looser thresholds include more general-capability neurons and M-MMLU drops sharply (down to 23.55%). The default thresholds thus balance the coverage and specificity of safety neurons, providing SNPO with a reliable routing target.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.