acceptodds
Under review as a conference paper at ICLR 2027

SafetyTransfer: Distilling Harmful and Harmless Discrimination for Selective Refusal in Small Language Models

10 views

Abstract

Safety alignment requires language models to refuse harmful requests while preserving useful responses to harmless ones. Transferring safety capabilities from an aligned teacher can improve compact models, but output-level distillation does not explicitly transfer the internal discrimination that determines when refusal is appropriate. We introduce SafetyTransfer, a representation-level framework that transfers harmful–harmless discriminative information from an aligned teacher to a smaller student. SafetyTransfer first selects teacher layers where this distinction is both geometrically separated and linearly accessible, then extracts a low-rank contrastive subspace rather than matching complete hidden states. Each prompt is projected onto this subspace to preserve an input-specific discriminative component. To bridge heterogeneous hidden dimensions, the component is expressed with sparse token-semantic coefficients derived from the teacher LM head and reconstructed with the corresponding student token-semantic dictionary. The student is then trained with representation alignment to these reconstructed targets together with standard response supervision. Across three teacher–student pairs and four safety datasets, SafetyTransfer achieves the highest harmful refusal rate minus harmless over-refusal rate (Gap) among the main baselines and the highest harmonic mean of harmful refusal and harmless acceptance (SR-H) in 10, while targeted removal of the transferred subspace produces larger behavioral changes than norm-matched orthogonal perturbations. The code is provided in the supplementary material.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Knowledge Distillation, Safety Alignment, Harmful and Harmless Discrimination

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_H3Z9OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SafetyTransfer, a knowledge transfer framework for safety alignment in small language models, aiming to improve their ability to distinguish between harmful and harmless requests while mitigating over-refusal. The method first extracts low-rank safety contrastive subspaces from the teacher model using SVD. It then constructs a token-semantic interface based on the LM Head to transfer safety-related representations to the student model, which is trained through response supervision and representation alignment. Experiments across multiple model families and safety evaluation datasets demonstrate that the proposed method effectively reduces over-refusal and improves selective-refusal performance.

Critical strengths
  1. The paper addresses the balance between refusal capability and helpful responses in the safety alignment of small language models. The research motivation is clear, and the problem has practical relevance.

  2. The proposed method extracts safety-related representations using low-rank contrastive subspaces and employs the LM Head for cross-dimensional transfer, avoiding direct alignment of complete hidden states or the need to train additional cross-dimensional projection networks. Its effectiveness is validated across three model families and multiple datasets, demonstrating consistent improvements in Gap and SR-H. Ablation studies further support the effectiveness of several key design choices.

Critical weaknesses
  1. All three teacher–student model pairs evaluated in the paper belong to the same model family. Since the proposed method relies on shared token identities to establish semantic correspondence across models, its effectiveness may be affected by differences in vocabularies, tokenizers, and hidden representation structures. Therefore, its applicability to heterogeneous models remains to be explored.

  2. The extracted subspaces may be influenced by the training data distribution, and the dominant deviation directions may not necessarily capture generalizable safety-discriminative information. It remains unclear whether the current results demonstrate robust cross-dataset transferability.

  3. The experiments are limited to student models with 0.6B–1B parameters, without evaluation on larger student models. In addition, the existing metrics mainly focus on harmful refusal rates and over-refusal rates on harmless requests. However, a lower over-refusal rate does not necessarily indicate that the model can provide accurate and helpful responses to benign requests. The paper lacks an evaluation of the student's general task capabilities after safety transfer.

  4. The proposed method requires scoring and selecting hidden layers from the teacher model. However, it remains unclear whether different datasets require different layer selections or whether a fixed set of layers can be consistently used across multiple datasets.

  5. The paper constructs low-rank contrastive subspaces using SVD, where the subspace dimensionality directly determines the amount of representation information retained. However, a systematic analysis of different subspace dimensions is missing. I suggest conducting a rank sensitivity analysis to investigate how different values of (r) affect HR, OR, Gap, and SR-H.

Questions for authors

None

5 comments
AuthorsOpenReview ↗

Response for Reviewer H3Z9 [1/2]

[Weakness 1] The applicability of SafetyTransfer to heterogeneous models remains unclear.

We clarify that SafetyTransfer is originally designed for teacher–student pairs with compatible vocabularies, where shared token identities establish a common semantic coordinate system through their LM heads. However, we further evaluate cross-family transfer from Llama-3.1-8B-Instruct to Qwen3-4B on WildGuardMix. To connect their different vocabularies, we construct an explicit mapping between tokens with matching token strings and verify that each mapped pair decodes to the same text. We then map the teacher’s sparse token-semantic coefficients to the corresponding student token indices and reconstruct the supervision targets using the student’s semantic dictionary. This preserves the transfer procedure while replacing the original shared-vocabulary correspondence with an explicit mapping.

Qwen3-4BHR ↑OR ↓Gap ↑SR-H ↑
Original checkpoint68.308.0460.2678.38
SafetyTransfer89.3915.0374.3687.13
These results support the feasibility of cross-family transfer with an explicit vocabulary mapping. We add this experiment and the mapping procedure to the revised manuscript.

[Weakness 2] Are the extracted subspaces sensitive to training-data distributions, limiting cross-dataset transferability?

To examine cross-dataset transfer, we trained Qwen3-0.6B models on WJB and evaluated them on WGM without target-domain retraining or model selection. The results are as follows:

MethodHR (%) ↑OR (%) ↓Gap ↑
SFT79.3122.5456.77
SafetyTransfer79.9719.6860.29

Compared with SFT, SafetyTransfer increases HR by 0.66 percentage points while reducing OR by 2.86 points on the unseen WGM dataset. This provides direct evidence that the benefits of SafetyTransfer can extend beyond its training distribution. We will incorporate this result into the revised manuscript and provide broader cross-dataset generalization for further investigation.

[Weakness 3] The experiments are limited to small student models and lack an evaluation of general task capabilities after safety transfer.

(1) Student model scale. We clarify that the goal of SafetyTransfer is to transfer harmful–harmless discrimination from larger aligned teachers to smaller students, enabling compact models to reject harmful requests while remaining responsive to harmless ones. We initially focused on compact students (0.6B–1B parameters), for which safety alignment is particularly challenging: stronger safety tuning can improve harmful refusal but also increase over-refusal on benign inputs. To further examine scalability beyond this range, we additionally trained on the Qwen3-4B student. The results on the WGM are shown in the table in the Response for Weakness 1. This extends the empirical evaluation to a 4B student and demonstrates the feasibility of the SafetyTransfer.

(2) General task capabilities. Our main experiments focus on safety datasets and assess harmful refusal and harmless acceptance through HR and OR. To further examine how safety-transfer training affects capabilities beyond selective refusal, we conducted additional PIQA and IFEval evaluations. PIQA evaluates everyday physical commonsense reasoning on 1,838 questions, while IFEval evaluates adherence to 834 verifiable instructions across 541 prompts, including formatting and length requirements.

The table reports SafetyTransfer accuracy minus SFT accuracy in percentage points, with positive values indicating improvements. For IFEval, prompt-level accuracy requires satisfying every instruction within a prompt, whereas instruction-level accuracy evaluates each requirement separately. Both use strict scoring, which checks the original response directly against the requirements.

Student modelTraining dataΔ PIQA accuracy (pp)Δ IFEval strict prompt accuracy (pp)Δ IFEval strict instruction accuracy (pp)
Qwen3-0.6BWGM+0.43−2.03−0.36
Qwen3-0.6BSTL+0.27−1.11−0.84
Llama-3.2-1BWGM+0.11−3.33−4.44
Llama-3.2-1BSTL−0.60+0.55−0.84

The results indicate little adverse impact on physical commonsense reasoning measured by PIQA, with changes ranging from −0.60 to +0.43 percentage points. IFEval results are more mixed: instruction-level accuracy decreases by less than one percentage point in three settings, while larger declines occur for Llama/WGM. These findings should be considered alongside our primary objective of safety-capability transfer. The instruction-following declines do not negate the selective-refusal improvements demonstrated in our safety evaluations, but reveal a capability trade-off accompanying this specialized training. We will include these evaluations, protocol details, and discussion of this trade-off in the revised appendix.

AuthorsOpenReview ↗

Response for Reviewer H3Z9 [2/2]

[Weakness 4] Does SafetyTransfer require dataset-specific teacher-layer selection, or can fixed layers be reused across datasets?

Although teacher-layer selection is data-driven, our additional experiments show that dataset-specific reselection is not always necessary. For Qwen3-8B→Qwen3-0.6B, the top seven layers selected from PAN ({19, 20, 21, 22, 23, 24, 25}) overlap with six of the seven layers selected from WGM ({19, 20, 22, 23, 24, 25, 26}). To further test layer reusability, we fixed the PAN-selected layers and trained SafetyTransfer on WGM without reselecting teacher layers, while keeping the training configuration and number of supervised layers identical. The test results are:

Layer selectionHR (%) ↑OR (%) ↓Gap ↑
WGM-specific selection89.1226.8862.25
Fixed PAN-selected layers88.4624.8763.59

The fixed-layer setting achieves comparable HR while reducing OR by 2.01 percentage points and improving Gap by 1.35 points. These results demonstrate that teacher-layer selection can be reused across datasets without sacrificing selective-refusal performance in this setting. We will incorporate this comparison into the revised manuscript to further demonstrate the flexibility of SafetyTransfer in teacher-layer selection and its ability to reuse selected layers across datasets.

[Weakness 5] The sensitivity of SafetyTransfer to the contrastive-subspace rank has not been systematically evaluated.

Thank you for highlighting the importance of subspace dimensionality. We conducted an additional rank sensitivity analysis on Qwen3-0.6B trained on WildGuardMix, evaluating four fixed SVD ranks (r∈{8,16,24,32}r\in\{8,16,24,32\}). These values represent the actual contrastive-subspace dimensionality, rather than the LoRA rank. We kept the teacher-layer selection, training data, optimization settings, random seed, and training duration fixed, changing only the SVD rank. All configurations were evaluated on the same full test set containing 754 harmful and 945 harmless prompts.

Rank strategyHR (%) ↑OR (%) ↓Gap ↑SR-H ↑
Fixed r=8r=888.8625.1963.6781.23
Fixed r=16r=1688.7327.6261.1179.72
Fixed r=24r=2488.9927.4161.5879.96
Fixed r=32r=3289.2628.2561.0079.55
Adaptive (80% energy, max rank 32)89.9224.4465.4882.11

HR remains highly stable across the four fixed ranks, varying by only 0.53 percentage points despite a fourfold increase in subspace dimensionality. However, increasing the rank does not consistently improve selective-refusal performance: from r=8r=8 to r=32r=32, HR increases by only 0.40 percentage points, while OR increases by 3.06 points, resulting in lower Gap and SR-H. More importantly, our default adaptive strategy, which selects the rank independently for each teacher layer using an 80% cumulative squared-singular-value energy criterion with an upper bound of 32, outperforms all four fixed-rank configurations across all four metrics. These results support the use of layer-wise adaptive rank selection: retaining more dimensions does not necessarily improve selective refusal, whereas adapting the retained dimensionality to each layer's singular-value spectrum provides a more effective representation of the contrastive information. We will incorporate this analysis into the revised manuscript.

AuthorsOpenReview ↗

Follow-up to Reviewer H3Z9

Thank you for your careful review and constructive comments. We would greatly appreciate a brief reply on whether our rebuttal addresses your main concerns.

We tested transfer from Llama to a Qwen3-4B student and across datasets, evaluated commonsense reasoning and instruction following, and examined the effects of reusing teacher layers and varying the subspace rank.

Thank you again for your time and consideration.

Reviewer_sSUiOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary
  • This paper proposes SafetyTransfer, a representation-level distillation framework for transferring harmful–harmless discrimination from a larger aligned teacher to a smaller student.
  • The method extracts low-rank contrastive subspaces from teacher representations, maps the resulting components through an LM-head-derived token-semantic interface, and aligns the student representations with the reconstructed targets.
  • Experiments across three teacher–student pairs and four safety datasets show improved Gap and SR-H, mainly through substantially reduced over-refusal.
Critical strengths
  • The paper addresses an important practical problem. It targets selective refusal in compact models and explicitly considers both harmful refusal and harmless acceptance rather than only maximizing refusal.

  • The representation-transfer design is technically interesting. The combination of contrastive subspace extraction and token-semantic reconstruction provides a concrete way to transfer selected teacher representations across models with different hidden dimensions.

  • The empirical improvements are substantial on over-refusal. Across the main experiments, SafetyTransfer consistently improves Gap and usually SR-H, with particularly large reductions in harmless over-refusal compared with SFT and vanilla KD.

Critical weaknesses
  • It is unclear whether the extracted subspace really captures harmful–harmless discrimination: The subspace may also contain other differences between harmful and harmless prompts. In addition, harmful refusal often decreases together with over-refusal. This may indicate a general shift toward lower refusal, rather than better safety discrimination. The intervention results show the same pattern.
  • Gap assumes an unjustified linear and symmetric trade-off between safety and over-refusal: Gap = HR − OR implicitly treats a one-point loss in harmful refusal as equivalent to a one-point improvement in harmless acceptance. This assumption may be inappropriate in safety-critical settings, where the cost of missing harmful requests may be much higher. A Pareto analysis or helpfulness comparison under fixed safety constraints would be more informative.
  • The necessity of the proposed representation transfer over output-level distillation is not fully established: The token-semantic bridge itself is derived from LM-head geometry, making it unclear how much information is transferred beyond safety-related output/logit information. Although the paper includes vanilla KD and limited comparisons with GKD and DistiLLM-2, stronger distillation baselines are evaluated only on a subset of model–dataset settings.
Questions for authors
  • Can you evaluate the safety–helpfulness trade-off without relying primarily on Gap? For example, it would be useful to report Pareto curves or harmless acceptance at matched harmful-refusal levels.

  • How much additional information does the proposed representation transfer provide beyond safety-specific logit distillation? A useful control would be to directly distill the vocabulary-space signal induced by the extracted safety component, e.g. (W_T c_T), and compare it with the proposed token-semantic reconstruction.

  • How robust is the token-semantic bridge across model families? Since the current experiments use within-family teacher–student pairs, a cross-family setting such as Qwen→Llama would better test whether shared token identities genuinely provide a general semantic interface.

5 comments
AuthorsOpenReview ↗

Response for Reviewer sSUi [1/2]

[Weakness 1]: Does the extracted subspace capture harmful–harmless discrimination or merely a general refusal tendency?

SafetyTransfer is designed to extract class-discriminative information rather than a generic refusal direction. We select teacher layers based on both harmful–harmless geometric separation and held-out linear-probe performance, and construct a low-rank subspace from harmful deviations relative to the harmless reference. More importantly, our controlled ablation directly supports the importance of this contrastive structure: replacing the proposed subspace with a single harmful-minus-harmless mean direction, while keeping the remaining transfer procedure unchanged, reduces Gap and SR-H across all six PAN and WGM settings. For example, on Qwen3.5/WGM, the mean-direction variant achieves even higher HR (89.26% vs. 84.35%), but substantially higher OR (44.55% vs. 19.58%), resulting in a much lower Gap (44.71% vs. 64.77%). This demonstrates that stronger refusal alone does not explain our gains; preserving multiple contrastive directions is important for selective refusal.

Furthermore, although targeted subspace removal decreases both HR and OR, the effects are asymmetric: across all four intervention settings, HR decreases by 12–15 percentage points, whereas OR decreases by 24–35 points, resulting in a 12–21-point increase in Gap. Norm-matched orthogonal perturbations produce substantially smaller changes. This indicates that the transferred subspace affects harmful and harmless inputs differently, rather than simply shifting the model's overall refusal tendency. Together with SafetyTransfer achieving the highest Gap in all 12 model–dataset settings, these results support the interpretation that the transferred components carry behaviorally relevant harmful–harmless discriminative information.

[Weakness 2 + Question 1]: Does Gap impose an unjustified symmetric trade-off between safety and helpfulness, and can this trade-off be evaluated independently of Gap?

Our primary objective is to improve selective refusal by maintaining or improving harmful refusal (HR) while reducing harmless over-refusal (OR), rather than trading off one against the other. We acknowledge that Gap (HR − OR) assigns equal numerical weights to the two metrics and is therefore a descriptive summary rather than a cost-sensitive utility function. Importantly, we report HR and OR separately, and several comparisons show simultaneous improvements in both metrics, without requiring any weighting assumption. Representative results are shown below.

SettingComparisonHR (%) ↑OR (%) ↓
Qwen3/PANSFT94.7918.13
SafetyTransfer97.7110.83
Qwen3.5/STLSFT87.104.84
SafetyTransfer89.923.63
Llama/WGMVanilla KD85.6816.63
SafetyTransfer87.0014.71

In these settings, SafetyTransfer improves both objectives simultaneously, demonstrating that its advantages are not merely consequences of the equal-weight Gap formulation. These comparisons represent Pareto improvements at the evaluated operating points, rather than a complete Pareto-frontier or matched-HR analysis. We therefore do not claim that Gap captures the asymmetric costs of safety errors or establishes optimal performance under arbitrary safety constraints.

AuthorsOpenReview ↗

Response for Reviewer sSUi [2/2]

[Weakness 3 + Question 2]: Does representation-level transfer provide additional benefits beyond safety-specific output-level distillation?

(1) Direct vocabulary-space supervision versus representation-level transfer.

The token-semantic bridge uses LM-head geometry as a shared coordinate system, but the transferred targets are reconstructed in the student hidden space rather than directly matched as output logits. To examine whether the corresponding vocabulary-space signal alone is sufficient, we additionally compare direct supervision using WTcTW_Tc_T with representation-level transfer on Qwen3/WGM. The results are:

MethodHR (%) ↑OR (%) ↓
Direct vocabulary-space supervision46.0227.94
SafetyTransfer89.1226.88

SafetyTransfer achieves substantially higher HR while also slightly reducing OR, suggesting that directly supervising the vocabulary-space signal does not reproduce the benefit of representation-level transfer.

(2) Extended comparisons with stronger distillation baselines.

We further extended the comparisons with GKD and DistiLLM-2 under shared response supervision from PAN and STL to all four datasets on Qwen3. The results are reported separately for Gap and SR-H.

DatasetMetricGKDDistiLLM-2SafetyTransfer
PANGap ↑83.3382.5086.88
SR-H ↑91.3290.8293.24
STLGap ↑81.4581.0587.90
SR-H ↑90.3290.0993.89
WGMGap ↑61.2464.6465.48
SR-H ↑79.9882.3282.11
WJBGap ↑12.8112.5426.64
SR-H ↑48.1053.7258.16

SafetyTransfer achieves the highest Gap across all four datasets and the highest SR-H on three. These additional results provide broader evidence of the practical benefits of representation-level transfer over the evaluated distillation baselines, although the expanded comparison remains limited to the Qwen3 model pair.

[Question 3]: How robust is the token-semantic bridge across different model families?

We clarify that SafetyTransfer is originally designed for teacher–student pairs with compatible vocabularies, where shared token identities establish a common semantic coordinate system through their LM heads. However, we further evaluate cross-family transfer from Llama-3.1-8B-Instruct to Qwen3-4B on WildGuardMix. To connect their different vocabularies, we construct an explicit mapping between tokens with matching token strings and verify that each mapped pair decodes to the same text. We then map the teacher’s sparse token-semantic coefficients to the corresponding student token indices and reconstruct the supervision targets using the student’s semantic dictionary. This preserves the transfer procedure while replacing the original shared-vocabulary correspondence with an explicit mapping.

Qwen3-4BHR ↑OR ↓Gap ↑SR-H ↑
Original checkpoint68.308.0460.2678.38
SafetyTransfer89.3915.0374.3687.13

Relative to the original checkpoint before our fine-tuning, SafetyTransfer improves Gap by 14.10 points and SR-H by 8.75 points, with increased harmful refusal and some increase in over-refusal. These results support the feasibility of cross-family transfer with an explicit vocabulary mapping. We will add this experiment and the mapping procedure to the revised manuscript.

AuthorsOpenReview ↗

Follow-up to Reviewer sSUi

Thank you for your careful review and constructive questions. We would greatly appreciate a brief reply on whether our rebuttal addresses your main concerns.

We provided evidence that the subspace supports selective refusal, identified settings where harmful refusal improves while harmless over-refusal declines, and reported additional comparisons with output-level distillation and a cross-family transfer experiment.

Thank you again for your time and consideration.

Reviewer_1ZZhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to transfer safety alignment from a larger teacher model to a smaller student model. It first selects informative teacher layers based on the distributions of hidden states elicited by harmful and harmless prompts, as well as the linear accessibility of safety-related information. For the selected layers, the method computes harmful deviations and their associated low-rank subspaces. Features are projected onto these subspaces and then mapped into the student’s hidden space through an LM-head-derived token-semantic decomposition and reconstruction process. The student is trained using a combination of language modeling loss and representation alignment loss. Experiments show that the proposed method substantially improves harmful refusal while keeping over-refusal close to the level of the original model (NoSFT), whereas baselines such as SFT and vanilla KD tend to increase over-refusal.

Critical strengths
  1. The problem setting is well motivated. Larger models may have greater capacity and stronger safety alignment, making the transfer of safety-related representations to smaller models a reasonable and promising approach.
  2. The experiments demonstrate strong selective refusal performance across multiple datasets and models. In particular, Table 1 shows that SafetyTransfer achieves a better balance between harmful refusal and harmless acceptance than SFT and vanilla KD.
Critical weaknesses
  1. The intervention results are difficult to reconcile with the main results. The paper states that the intervention experiments use “independently trained single-seed mechanism checkpoints.” However, the harmful refusal and harmless acceptance rates of the checkpoints reported in Figures 2 and 3 differ substantially from those in Table 1. These discrepancies raise concerns about training variability or differences in the experimental setup and require a clearer explanation. The procedure for constructing the student-side target subspace is also insufficiently clear.
  2. The contribution beyond SemAlign is not sufficiently isolated. As the paper acknowledges, the proposed algorithm builds on SemAlign, but the extent of the methodological differences and their contribution to performance remain unclear. In particular, the original SemAlign method is not included in Table 1. Moreover, Table 5 in the appendix shows only a small performance gap between NoProj and SafetyTransfer. This raises the possibility that projection onto the harmful-related subspace—a central novel component of the method—provides only a limited incremental benefit. A direct comparison with SemAlign and a clearer analysis of this component’s contribution would strengthen the paper.
Questions for authors
  1. How does SafetyTransfer affect general capabilities and instruction-following ability? Evaluations beyond harmful refusal (HR) and over-refusal (OR) would be helpful, since OR alone does not capture changes in these abilities.
  2. Do tokens with large absolute token-semantic coefficients exhibit recognizable semantic patterns when decoded into text? Representative examples could help clarify what the transferred components encode.
5 comments
AuthorsOpenReview ↗

Response for Reviewer 1ZZh [1/2]

[Weakness 1] Discrepancies between the main and intervention results, and construction of the student-side target subspace

Thank you for raising this concern. The main and intervention experiments use different training and evaluation settings. Our description of “independently trained” models did not sufficiently explain these differences, which we summarize below. In particular, WGM/WJB in Figures 2–3 denotes the source of evaluation requests, whereas Table 1 evaluates models trained on the corresponding domains.

AspectTable 1: overall performanceFigures 2–3: mechanism interventions
Training dataCorresponding training domain, such as WGM or WJBPAN
Training configurationDomain-specific main-study configuration, including Qwen3 auxiliary-loss weights of 0.05 for WGM and 0.5 for WJBDefaulted mechanism-study configuration with auxiliary-loss weight 0.25
Evaluation samplesFull test sets, including 754 harmful and 945 harmless requests for WGMFixed 100 harmful and 100 harmless requests per displayed domain
PurposeCompare overall performance across methodsCompare behavioral changes within the same model under different interventions

Figures 2 and 3 present complementary views of the same mechanism experiments. Differences in training data, configuration, and evaluation samples mean that their absolute refusal rates are not directly comparable with Table 1. Within the mechanism study, however, the no-intervention, norm-matched orthogonal-control, and target-removal conditions use the same model, requests, and generation settings. The mechanism conclusions therefore follow from changes between these paired conditions. We will make this distinction explicit in the revised description.

The student-side target subspace is constructed from the reconstructed target vectors used to supervise harmful training examples. We group these vectors by student layer, normalize and stack them, and combine targets from multiple teacher layers when they are assigned to the same student layer. We then apply uncentered SVD and retain directions accounting for 80% of the squared-singular-value energy, capped at 32 directions. These directions define the student-side target subspace. It is constructed in student coordinates from training targets, rather than copied directly from the teacher subspace or fitted using test examples. We will add these construction details to the revision.

[Weakness 2] Contribution beyond SemAlign and the role of subspace projection

SafetyTransfer adopts SemAlign’s cross-model semantic interface, while our additional design addresses which safety information to transfer. We select teacher layers that distinguish harmful from harmless requests, construct contrastive subspaces from harmful representations relative to a harmless reference, and transfer input-specific components while retaining response supervision. The resulting targets focus on harmful–harmless discrimination rather than complete layer representations. To supplement this methodological distinction, we implemented a baseline following the SemAlign paper’s formulas and evaluated Qwen3-8B→Qwen3-0.6B on the same WGM test set of 754 harmful and 945 harmless requests. All values below are percentages.

MethodHR ↑OR ↓Gap ↑
SemAlign (paper-formula implementation)46.1510.1636.00
SafetyTransfer89.1226.8862.25

SafetyTransfer achieves 42.97 percentage points higher harmful refusal and 26.25 points higher Gap, alongside higher harmless over-refusal. This comparison uses their respective objectives and data configurations. The SafetyTransfer results come from the additional GPU experiment and will be reported separately from Table 1. Together, the methodological distinction and supplementary results support the value of explicitly targeting harmful–harmless discrimination beyond adopting the semantic interface, with the complete SafetyTransfer configuration achieving a higher Gap under the evaluated training settings.

Projection contributes by selecting components associated with harmful–harmless discrimination to form the student’s learning targets. Table 5 isolates its incremental contribution by removing only projection while retaining the remaining design. Projection improves average Gap by 0.82 percentage points and SR-H by 0.61 points across four datasets. On WGM, it reduces harmless over-refusal from 17.88% to 14.71% and increases Gap from 70.18% to 72.29%. Gap improves on three datasets, with a small decrease on PAN. The ablation therefore supports projection as a useful refinement of the transfer targets, improving aggregate selective-refusal performance without being the sole source of the framework’s effectiveness. We will clarify the distinction between this overall effectiveness and the incremental contribution of projection.

AuthorsOpenReview ↗

Response for Reviewer 1ZZh [2/2]

[Question 1] How does SafetyTransfer affect general capabilities and instruction following?

The goal of SafetyTransfer is to transfer harmful–harmless discrimination from a larger teacher to a smaller student, enabling the student to reject harmful requests while remaining responsive to harmless ones. Our main experiments therefore focus on safety datasets and assess these complementary objectives through HR and OR. To further examine how this safety-transfer training affects capabilities beyond selective refusal, we conducted additional PIQA and IFEval evaluations for the Qwen3 and Llama settings below. We compare SafetyTransfer with SFT using response supervision only. PIQA evaluates everyday physical commonsense reasoning on 1,838 questions, while IFEval evaluates adherence to 834 verifiable instructions across 541 prompts, including formatting and length requirements.

The table reports SafetyTransfer accuracy minus SFT accuracy in percentage points, with positive values indicating improvements. For IFEval, prompt-level accuracy requires satisfying every instruction within a prompt, whereas instruction-level accuracy evaluates each requirement separately. For example, satisfying two of three requirements yields a failed prompt but two correct individual instructions. Both columns use strict scoring, which checks the original response directly against the requirements.

StudentTraining dataΔ PIQA accuracyΔ IFEval strict prompt accuracyΔ IFEval strict instruction accuracy
Qwen3-0.6BWGM+0.43−2.03−0.36
Qwen3-0.6BSTL+0.27−1.11−0.84
Llama-3.2-1BWGM+0.11−3.33−4.44
Llama-3.2-1BSTL−0.60+0.55−0.84

The results indicate little adverse impact on the commonsense reasoning performance measured by PIQA, with small improvements in three settings and a decline of 0.60 percentage points in the fourth. On IFEval, instruction-level accuracy decreases by less than one percentage point in three settings, with a larger decline of 4.44 points for Llama/WGM and a corresponding prompt-level decline of 3.33 points. These findings should be considered alongside our primary objective of safety-capability transfer. The instruction-following declines do not negate the selective-refusal improvements demonstrated in the safety evaluations, but identify a capability trade-off accompanying this specialized training. We will provide the supplementary evaluations, protocol details, and discussion of this trade-off in the revised appendix.

[Question 2] Do large token-semantic coefficients reveal recognizable semantic patterns?

We analyzed teacher-side token-semantic coefficients computed on STL and WJB training examples for transfer from Qwen3-8B to Qwen3-0.6B. To understand the information being transferred, we decoded tokens associated with large absolute coefficients. These coefficients combine token-semantic directions to reconstruct learning targets in the student’s representation space. The table presents representative examples, with layer indices referring to the Qwen3-8B teacher and coefficient signs indicating their direction in the reconstruction.

Data / request type / teacher layerPositive-coefficient examplesNegative-coefficient examples
STL / harmful / 19disclaimer (+0.0803), cautious (+0.0769), ethical (+0.0768)هو (−0.0641), FString (−0.0633), 小事 (−0.0628)
WJB / harmful / 23我不想 (+0.0945), 我不 (+0.0915), cannot (+0.0902)легко (−0.0820), ToolTip (−0.0642), פיל (−0.0638)
STL / harmless / 23下面就 (+0.0804), certains (+0.0712), İşte (+0.0704)无法 (−0.0941), Cannot (−0.0851), cannot (−0.0848)

In the harmful examples, prominent positive coefficients correspond to caution and ethical language, such as cautious and ethical, and refusal expressions, such as 我不 (“I do not”), 我不想 (“I do not want to”), and cannot. In the harmless example, refusal-related tokens instead contribute with large negative weights, including 无法, which expresses inability. These examples provide interpretable semantic cues associated with safety discrimination. Not every high-magnitude token has an obvious safety-related meaning, as the table also contains code fragments and tokens from other languages. This is a valuable suggestion, and we will add this discussion and representative examples to the revised manuscript to make the transferred information more intuitive.

AuthorsOpenReview ↗

Follow-up to Reviewer 1ZZh

Thank you for your careful review and constructive comments. We would greatly appreciate a brief response on whether our rebuttal addresses your main concerns.

We clarified the intervention results, compared SafetyTransfer with SemAlign, and provided additional capability evaluations and examples of the tokens associated with refusal.

Thank you again for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary Response to the Area Chair

Dear Area Chair,

We thank you and all reviewers for the careful evaluation and constructive feedback. We have revised the manuscript to address the main concerns through additional experiments, clearer methodological explanations, and expanded analysis. The corresponding changes are highlighted in blue. Our detailed responses and experimental results are provided in the replies to individual reviewers. We summarize the principal revisions below.

Shared concerns across reviewers:

  • Larger students and cross-family transfer. We added an experiment transferring from Llama-3.1-8B-Instruct to Qwen3-4B through an explicit vocabulary mapping. Compared with the original student checkpoint, SafetyTransfer improves Gap from 60.26 to 74.36 and SR-H from 78.38 to 87.13. This extends the evaluation beyond compact students and provides evidence for the feasibility of cross-family transfer with aligned token correspondences.
  • General capabilities beyond harmless acceptance. We added PIQA and IFEval evaluations across Qwen3 and Llama students trained on WGM and STL, together with an additional Qwen3.5 comparison. These evaluations distinguish answer correctness and instruction following from refusal behavior. We report the small PIQA changes and the observed instruction-following degradation, providing a more complete account of the effects of safety-transfer training.

Additional concerns raised by Reviewer 1ZZh:

  • Interpretation of the mechanism experiments. We clarified the different checkpoints and evaluation panels used in the intervention study and the main table. We also added the construction of the student-side intervention subspace and explained how the controlled interventions relate transferred directions to refusal behavior.
  • Contribution beyond SemAlign. We clarified how informative-layer selection and input-specific contrastive targets complement the adopted semantic interface. We added a SemAlign comparison and distinguished the contribution of the complete framework from the incremental effect of projection.
  • Interpretation of token-semantic coefficients. We added examples of positive and negative coefficients and explained their role as signed representation coordinates.

Additional concerns raised by Reviewer sSUi:

  • Discrimination and the safety–utility trade-off. We clarified the meaning of harmful–harmless discrimination and the scope of the intervention evidence. We also explained Gap’s symmetric weighting and retained separate HR and OR reporting to make the underlying trade-off explicit.
  • Alternative supervision and stronger distillation baselines. We added direct supervision using the safety-component vocabulary-space signal WTcTW_Tc_T and expanded the GKD and DistiLLM-2 comparisons to all four Qwen3 datasets. SafetyTransfer achieves the highest Gap in all four datasets and the highest SR-H in three, strengthening the empirical comparison with alternative transfer objectives.

Additional concerns raised by Reviewer H3Z9:

  • Subspace dimensionality. We added a controlled sensitivity study with actual SVD ranks 8, 16, 24, and 32, reporting HR, OR, Gap, and SR-H. HR varies by only 0.53 percentage points, while OR shows greater sensitivity. This directly characterizes how retained dimensionality affects selective refusal.
  • Layer reuse and cross-dataset transfer. We added a fixed-layer comparison and evaluated WJB-trained models on WGM and PAN without target-domain retraining. These results provide direct evidence about layer reuse and cross-dataset transfer, including the conditions where improvements do not carry over.

We thank the Area Chair and reviewers again for their time and consideration.

AuthorsOpenReview ↗

A Brief Follow-up on Our Rebuttal

Dear Reviewers,

Thank you for the time you have devoted to our submission. As the discussion period draws to a close, we would be grateful if the reviewers could briefly let us know whether our responses address their main concerns.

If any questions remain, we would be happy to clarify them before the deadline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.