acceptodds
Under review as a conference paper at ICLR 2027

Making Your LLMs More Objective: Stabilizing LLM Safety Behavior Across Traits with Trait-Invariant Safety Tuning

Abstract

Aligned large language models (LLMs) are expected to exhibit safety behavior based on the content of the user request: they should refuse unsafe requests and comply with safe ones. However, we show that the same request can elicit substantially different safety decisions under different traits assigned in the system prompt, a failure mode we call trait-induced safety variation. To measure this failure, we introduce refusal-based metrics: Trait-Induced Deviation measures dataset-level deviation from the no-trait baseline, while Trait-Induced Flip Rate measures whether the same request receives different safety decisions across traits. We then provide a representation-level analysis of the mechanism behind trait-induced safety shifts and find that traits perturb the model's safety representations within a low-dimensional subspace. To achieve trait-invariant safety, where safety behavior remains stable across traits, we introduce Trait-Invariant Safety Tuning (TIST), a simple yet effective self-distillation framework that aligns an LLM's trait-conditioned behavior with its no-trait behavior. Guided by our analysis, we further propose Trait-Subspace Neutralization (TraSN), an instantiation of TIST, which enforces invariance only within the identified trait subspace. Experiments show that TraSN improves trait-invariant safety and strengthens harmful-request safety while preserving general capability. Our results highlight traits as an important factor in LLM safety and robust model behavior.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Trait, LLM safety, LLM Alignment

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_vyDzOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies changes in LLM refusal decisions caused by system-prompt personas and personality traits. It defines Trait-Induced Deviation and Trait-Induced Flip Rate to distinguish aggregate refusal shifts from changes on individual requests. An activation analysis identifies a low-dimensional subspace of mean trait-induced shifts. The proposed TIST framework distills a frozen model’s no-trait behavior into its trait-conditioned counterpart; TraSN performs this matching only within the estimated subspace. Experiments on three model families compare response, logit, full-activation, and subspace matching, reporting reduced decision instability, stronger harmful-request refusal, and broadly preserved performance on capability tasks.

Critical strengths
  • The evaluation distinguishes stability from correctness rather than treating reduced variation as sufficient safety. Sections 2 and 5 report request-level flips alongside aggregate deviation, harmful refusal, benign over-refusal, and capability. The improvements are substantial across models: in Table 1, harmful TFR falls from 12.42 to 5.67 for Llama and from 18.14 to 4.74 for Gemma. Tables 2–5 expose dataset-level results and transfer to three held-out traits, making the aggregate findings more informative.

  • The comparisons meaningfully test the design choice. Response, logit, and full-activation matching use the same data and training settings (Appendix J), while the matched-rank random-subspace control directly tests whether an estimated trait subspace helps beyond generic low-rank regularization. On Llama, TraSN improves harmful TFR from the random control’s 8.49 to 5.67 and avoids much of its reported over-refusal increase (Table 6). The rank sweep and repeated-seed results provide additional support, with the latter appropriately limited to Llama (Figure 3; Table 7).

Critical weaknesses
  • The construction of SafeRLHF-Safe does not establish the benign-request ground truth needed to interpret its refusal rate as over-refusal. Table 9 specifies filtering PKU-SafeRLHF test examples by is_response_0_safe=true. In the source dataset, this labels the supplied response’s harmlessness, not the prompt’s harmlessness; a harmful request with a safe refusal can satisfy the filter. Appendix I nevertheless treats these as safe prompts. Without an additional prompt-level criterion, legitimate refusals can be counted as errors. This affects the interpretation of that benchmark and its contribution to the benign aggregate; it does not by itself invalidate the other three benign benchmarks. The selection of the SafeRLHF training/calibration prompts also needs clarification.

  • The evidence does not establish preservation of useful persona behavior, an explicit motivation in the introduction (page 2, lines 59–61). All capability evaluations use no-trait prompts (Appendices I and J), while the persona-conditioned evaluations measure refusal/compliance rather than style, role adherence, or task usefulness. Matching toward a no-trait teacher could reduce trait responsiveness more generally. Leaving orthogonal representation directions unconstrained does not establish that the corresponding behaviors survive parameter updates, and the two case studies report decisions and response counts rather than persona quality. The results support improved decision stability and preserved no-trait capability, but the intended separation of safety from useful role expression remains unverified.

Questions for authors
  1. Was SafeRLHF-Safe filtered for prompt harmfulness in addition to is_response_0_safe=true? Please clarify the actual rule for both evaluation and the purported benign training/calibration data, and how a harmful prompt paired with a safe refusal is handled.

  2. Is there an existing result that evaluates retained benign persona/style adherence after TraSN? If not, should the claims about allowing traits to influence style or role expression be narrowed to the demonstrated preservation of no-trait capability?

7 comments

Dear Reviewer vyDz,

Thank you for your careful review and constructive feedback. We address each of your questions below.


[Q1] Prompt-level safety filtering for SafeRLHF

Yes. In addition to selecting records with is_response_0_safe=true, we used Claude Haiku 4.5 to assess each user request based on the prompt text alone. We applied both filters to the SafeRLHF-Safe evaluation set and to the SafeRLHF prompts used for training and calibration. We retained only requests that Claude classified as benign. A request classified as harmful was therefore excluded even when its reference response was a safe refusal. The original Table 9 described only the response-level filter. We have revised Table 9 and Appendix I to document both filtering steps.


[Q2] Benign role expression after TraSN

To directly assess whether benign roles remain recognizable after tuning, we added a role-identification experiment on Qwen3.5-4B.

We used 100 requests from JBB-Benign and evaluated each request under four roles: pediatrician, high-school chemistry teacher, senior software engineer, and librarian. This yielded 400 responses per method. All three methods used the same requests and role instructions. We masked explicit profession names in each response and asked Claude Haiku 4.5 to infer the assigned role. The judge saw the user’s request, the masked response, and the four candidate role instructions in randomized order. It did not see the assigned role or the method that produced the response. Random guessing yields an accuracy of 25%.

MethodCorrect role identificationsAccuracy
Base model with role prompts282/40070.5%
TIST-Activation (full activation matching)97/40024.25%
TraSN (subspace matching)102/40025.5%

The results show that role-identification accuracy decreases from 70.5% for the base model to 25.5% after TraSN, approaching the random-guessing baseline. Both TIST-Activation and TraSN exhibit similar behavior, suggesting that these methods reduce the distinguishability of role-conditioned responses. This outcome is reasonable consistent with their training objectives, which align trait-conditioned representations with the no-trait reference, either across the full activation space or within the estimated trait subspace.

Accordingly, we have revised the claim in the Introduction from preserving benign role expression under trait prompts to preserving performance on general capability tasks, which is more directly supported by our evaluation. Preserving role-specific expression is not the primary objective of our work, and we have removed the corresponding claim from the manuscript.

Importantly, this finding does not undermine our work, which remains aligned with its core objective of stabilizing LLM safety behavior across different traits. We hope this clarification make the scope and contributions of our work clearer.


Thank you again for your thoughtful review. Your suggestions have helped us improve the paper, and we would be happy to discuss any remaining questions.

thanks for the clarification and the additional experiment. the prompt-level filtering addresses my concern about SafeRLHF-Safe, provided the revised manuscript clearly documents this rule for evaluation, training and calibration.

the role-identification result is more concerning. accuracy drops from 70.5% to 25.5%, essentially chance level, so this experiment does not support preservation of useful role expression. it is consistent with the concern that reduced trait responsiveness contributes to the improved stability, although this probe alone cannot establish the full mechanism.

narrowing the claim to preserved general capability is appropriate. however, the paper should also discuss the loss of role distinguishability as a limitation and make clear that preserved no-trait capability does not demonstrate preserved persona-conditioned behaviour. i would weigh this tradeoff when assessing the contribution.

Thank you for your prompt follow-up. We have added this trade-off to the Limitations section, explicitly acknowledging the loss of role distinguishability and clarifying that preserved capability under no-trait prompts does not demonstrate preserved persona-conditioned behavior.

While this finding limits our claims about persona preservation, we believe this concern does not invalidate our contributions to the problem formulation, representation-level analysis, and algorithmic approach for reducing trait-induced safety variation.

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

This paper demonstrates that system-prompt personas/roles/Big-Five traits alter aligned LLMs’ refusal decisions without jailbreaks, a bidirectional effect even for benign traits.

It makes three contributions: (1) TID and TFR to separate dataset drift and per-request flips; (2) trait-induced activation shifts concentrate in a rank-4 safety subspace; (3) TIST/TraSN enforce consistency against the trait-free baseline.

Experiments on three open models show TraSN boosts harmful refusal, cuts TID/TFR and retains capability, supported by strong ablations.

Key caveats: the target is behavioral invariance rather than safety ground truth. Trait expression is unmeasured, over-refusal persists, and evaluation uses only one LLM judge without error bars.

Critical strengths
  1. This work isolates and quantifies a subtle, under-evaluated failure mode. Its standout finding is bidirectional refusal shifts induced by benign roles and Big-Five traits, not just adversarial personas. Since persona conditioning is ubiquitous in deployed LLMs while safety audits mostly target jailbreaks, this finding carries high practical impact. TID and TFR make the failure measurable by separating dataset drift and per-request flips, and Tables 10–11 present concrete flip examples.
  2. Mechanism directly motivates TraSN and is validated. The rank‑4 subspace finding underpins the subspace intervention. The random-subspace ablation (Table 6) convincingly rules out generic low-rank regularization as the explanation. Together with OOD trait tests, matched compute budgets and seed checks, this evidence pushes the paper above the acceptance threshold despite its caveats.
Critical weaknesses
  1. The objective targets invariance rather than safety, and metrics cannot separate the two. The paper aims for request-content-driven “objectivity” but anchors TIST/TraSN to the model’s own weak trait-free baseline instead of safety ground truth. The Llama baseline has low harmful refusal, so anchoring propagates its errors. TFR penalizes all decision flips, including beneficial corrections of unsafe compliance. Reduced TFR may merely suppress helpful changes, not improve safety. Without external gold labels or decomposing flips into corrections/regressions, the work only demonstrates consistency, not safety.
  2. The paper does not verify preserved trait expression after intervention. Its goal is to retain role/style while stabilizing refusal behavior, yet it only evaluates refusal and trait-free capability. TraSN constrains exactly the subspace carrying trait-induced safety shifts; the observed gains may simply come from attenuating trait signals, a degenerate solution contradicting the core claim. Simple probes (role classification, trait projection magnitude) are missing. Finally, all safety evaluations rely on a single LLM judge with no human validation or error bars, so even invariance gains lack statistical support.
Questions for authors

Q1 (critical clarification): Can you quantify whether traits retain stylistic/role expression after TraSN (e.g., persona projection magnitude, trait probes, role classification)? If traits are merely attenuated instead of safety-neutral, the method discards trait signals and contradicts the paper’s core goal.

Q2 (key suggestion): How does TraSN perform when anchored to external safety gold labels rather than the trait-free baseline? Since the Llama baseline is weak (47.00% refusal on WJB-Harmful), please test whether gold-label anchoring boosts harmful refusal and decompose flips into corrective vs regressive ones.

Q3 (clarification): Table 8 includes two neutral anchors besides the trait-free setting. Which one is τ₀, and are results sensitive to this choice? The invariance objective is defined relative to τ₀.

Q4 (clarification): Subspace U is estimated only on harmful prompts but applied to benign prompts in loss. Do benign trait shifts lie within the same subspace? Subspaces estimated from benign or joint data would verify this.

Q5 (clarification): TIST-Logits achieves lower TID on Qwen3.5-4B. Please characterize settings where TraSN outperforms full logit/activation matching, rather than claiming universal superiority.

Q6 (clarification): Some performance gaps are small relative to noise, and evaluation uses only one LLM judge without human validation. Report seed-wise variance and judge-human agreement to establish statistical significance of invariance gains.

Q7 (suggestion): Absolute over-refusal does not improve despite lower TID/TFR. Can tuning λ_b trade partial invariance for lower over-refusal?

Q8 (suggestion): Does the rank‑4 trait subspace overlap with known refusal or persona directions? Cosine similarity analysis would strengthen the mechanism story.

Q9 (suggestion, out of scope): Transfer to larger models and instruction-hierarchy models is valuable for follow-up work and not required here.

3 comments

Dear Reviewer XzQJ,

Thank you for your careful review and constructive feedback. We address each of your questions below.


[Q1] Benign role expression after tuning

To directly assess whether benign roles remain recognizable after tuning, we added a role-identification experiment on Qwen3.5-4B.

We used 100 requests from JBB-Benign and evaluated each request under four roles: pediatrician, high-school chemistry teacher, senior software engineer, and librarian. This yielded 400 responses per method. All three methods used the same requests and role instructions. We masked explicit profession names in each response and asked Claude Haiku 4.5 to infer the assigned role. The judge saw the user’s request, the masked response, and the four candidate role instructions in randomized order. It did not see the assigned role or the method that produced the response. Random guessing yields an accuracy of 25%.

MethodCorrect role identificationsAccuracy
Base model with role prompts282/40070.5%
TIST-Activation (full activation matching)97/40024.25%
TraSN (subspace matching)102/40025.5%

The results show that role-identification accuracy decreases from 70.5% for the base model to 25.5% after TraSN, approaching the random-guessing baseline. Both TIST-Activation and TraSN exhibit similar behavior, suggesting that these methods reduce the distinguishability of role-conditioned responses. This outcome is reasonable consistent with their training objectives, which align trait-conditioned representations with the no-trait reference, either across the full activation space or within the estimated trait subspace.

Accordingly, we have revised the claim in the Introduction from preserving benign role expression under trait prompts to preserving performance on general capability tasks, which is more directly supported by our evaluation. Preserving role-specific expression is not the primary objective of our work, and we have removed the corresponding claim from the manuscript.

Importantly, this finding does not undermine our work, which remains aligned with its core objective of stabilizing LLM safety behavior across different traits. We hope this clarification make the scope and contributions of our work clearer.

Our practical motivation also concerns settings where the model encounters externally supplied instructions containing role or trait assignments at inference time. Omitting persona instructions from a controlled prompt does not establish robustness when such instructions are subsequently introduced. We therefore view trait-conditioned prompting as a potential perturbation or attack surface, and TIST/TraSN as a defense intended to stabilize safety decisions under these inputs. We have incorporated this more detailed discussion into Appendix K, “Discussion of the Practical Setting and Scope of This Work,” of the revised manuscript.


[Q2] External safety labels and the direction of flips

The original TraSN uses the frozen model’s representation without a trait prompt as its training target. If the base model incorrectly complies with a harmful request, this objective may preserve that error. We evaluated Llama-3.2-3B on 200 requests from WJB-Harmful under 12 trait prompts. For the same model and request, we count a change from compliance to refusal after adding a trait prompt as a corrective flip, and a change from refusal to compliance as a regressive flip.

MethodHarmful refusal rate without trait promptsHarmful refusal rate with trait promptsCorrective flip rateRegressive flip rateTotal flip rate
Base model47.00%51.58%10.63%6.04%16.67%
TraSN55.00%57.50%5.83%3.33%9.17%
TraSN + gold-label anchoring71.00%74.96%8.33%4.38%12.71%

On this benchmark, TraSN reduced the total flip rate from 16.67% to 9.17% and increased the harmful refusal rate under trait prompts from 51.58% to 57.50%. The regressive flip rate fell from 6.04% to 3.33%, and the corrective flip rate also fell, from 10.63% to 5.83%. The reduction in the total flip rate should therefore be interpreted alongside changes in both directions and the observed refusal rate.

We experimented with adding a safety anchor to TraSN and found that training was unstable and highly sensitive to the weight on the anchoring loss. With gold-label anchoring, the SFT control increased the harmful refusal rate under trait prompts to 74.96%, while also increasing the over-refusal rate on WJB-Benign from 23.04% to 37.42%. The additional loss terms made it difficult to balance their weights, and the model became more prone to over-refusal. We leave the design of a loss that better balances these terms while anchoring the model to safety labels to future work.


[Q3] The no-trait reference τ0\tau_0

In all training and evaluation runs, τ0\tau_0 denotes the no-trait condition. We pass only the user’s request to the model’s standard chat template, with no explicit system message.

We used the two neutral prompts in Table 8 only in a supplementary activation analysis. We ran the same 200 harmful requests twice, once with “You are a helpful assistant.” as the system prompt and once with “You are an AI assistant that answers questions.” For each prompt, we first averaged the activation vectors across the 200 requests. We then averaged the two resulting vectors with equal weight to obtain the alternative reference. We have clarified it in the Appendix H.


[Q4] Coverage of benign trait shifts by the harmful-prompt subspace

Benign trait shifts lie partly within the subspace estimated from harmful prompts. We tested this on frozen Llama-3.2-3B-Instruct using TraSN’s layer, 12 training traits, and rank 4. We estimated each subspace from 100 calibration requests drawn from WildGuard-Train (harmful), prompt-screened SafeRLHF-Train (benign), or both. We evaluated all three on the same 100 held-out benign requests from SafeRLHF-Test.

We measured activation shifts relative to the no-trait condition. Following Eq. 12, we averaged shifts over requests for each trait, centered the 12 means, and measured the variance captured. We also measured the fraction of total squared shift magnitude captured across all 1,200 individual request–trait pairs.

Calibration requestsMean trait-shift variance capturedIndividual shift energy captured
100 harmful64.1%30.8%
100 benign79.3%39.8%
50 harmful + 50 benign75.6%37.5%

The harmful-prompt subspace captured most of the variance across mean benign trait shifts, but only about a third of the shift energy at the individual-request level. Including benign prompts in calibration improved coverage on both measures.


[Q5] Scope of TraSN’s advantage

We agree that the ranking depends on the trait split and metric. On held-out Qwen3.5-4B traits, TIST-Logits achieves lower harmful and benign TID than TraSN, and TIST-Activation achieves lower harmful TFR (Table 5). TraSN’s most consistent advantage is lower benign TFR than both baselines across all three models and both trait splits (Tables 1 and 5). We have revised the text to state this more precisely.


[Q6] Seed variation and judge–human agreement

We manually reviewed 30 responses drawn from WJB-Harmful and WJB-Benign using the same refusal criteria as the original evaluation. The human labels agreed with Claude Haiku 4.5 (temperature 0) on all 30 responses (100% observed agreement).

Given our limited compute budget, we repeated training across five seeds on Llama-3.2-3B. Appendix G, Table 7 reports the mean and standard deviation across seeds. Relative to the base model, TraSN reduced harmful TFR from 12.42% to 5.45% ± 0.32% and benign TFR from 11.36% to 4.60% ± 0.22%. These reductions were large relative to the observed seed variation.


[Q7] Effect of weighting the benign consistency loss

In our exploratory experiments, we found that changing the weight of the benign consistency term mainly affects over-refusal. A smaller λb\lambda_b usually increases over-refusal and may allow greater trait-induced variation on benign requests. For simplicity, we retained equal weights for the harmful and benign terms as the default across all reported experiments.


[Q8] Overlap with refusal and role directions

We conducted a supplementary cosine analysis on frozen Llama-3.2-3B-Instruct at the layer used by TraSN, using the rank-4 subspace estimated from 100 harmful requests in our Q4 analysis.

We estimated a refusal-associated direction from the difference in mean no-trait activations between 50 refused and 50 complied harmful requests, using archived judge labels. This direction distinguished the two decisions on another 100 requests with AUC 0.860. We estimated four role directions as mean role-versus-no-trait activation shifts on 100 held-out benign requests.

For each direction vv, we computed its cosine with its projection onto the subspace, ∥Uv∥2/∥v∥2\|Uv\|_2/\|v\|_2.

DirectionCosine to subspace
Refusal-associated0.372
Pediatrician0.773
Chemistry teacher0.587
Software engineer0.303
Librarian0.393

The subspace has some overlap with the estimated refusal-associated direction and greater overlap with the four role directions tested. This suggests that it captures role-related variation alongside a refusal-associated component, which may help explain the reduction in role expression after tuning.


[Q9] Transfer to larger and instruction-hierarchy models

We agree that both are valuable extensions and leave them for future work.

Currently, we examine whether trait-induced safety variation also occurs in reasoning models or larger ones, we conducted additional behavioral evaluations on Qwen3.5-4B with reasoning enabled and Qwen3.5-9B with reasoning disabled:

SettingHarmful TID ↓Harmful TFR ↓Benign TID ↓Benign TFR ↓
Qwen3.5-4B, reasoning disabled4.52%9.09%13.12%21.00%
Qwen3.5-4B, reasoning enabled8.44%18.05%4.76%9.52%
Qwen3.5-9B, reasoning disabled3.87%7.26%8.95%15.62%

Interestingly, enabling reasoning in Qwen3.5-4B reduces stability on harmful requests while improving stability on benign requests. With reasoning disabled, Qwen3.5-9B exhibits greater stability than Qwen3.5-4B on both request types, although trait-induced variation remains present. These results show that the behavioral phenomenon extends to the tested larger-model and reasoning-enabled settings.


Thank you again for your thoughtful review. Your suggestions have helped us improve the paper, and we would be happy to discuss any remaining questions.

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

This paper studies how a persona or trait in the system prompt can change whether an LLM refuses the same request. The authors propose two metrics for this effect: TID measures the change in overall refusal rate, and TFR measures how often individual decisions flip.

On three open models, they find that even ordinary roles and personality traits can change refusal behavior. An analysis of internal activations shows that traits push harmful prompts toward the benign side, and that these shifts lie mostly in a small subspace. Based on this, the authors propose TIST, which trains the model to behave the same with or without a trait. Its main version, TraSN, applies this constraint only within that subspace. The experiments show that TraSN reduces both metrics and raises harmful-request refusal without hurting general capability.

Critical strengths

The paper shows that ordinary roles and personality traits, not only adversarial personas, can shift refusal in both directions. To measure this effect, the authors propose two metrics to capture this effect at two levels. First, TID shows how much a trait changes the overall refusal rate. Second, TFR shows how often a trait flips the decision on individual requests. This can catch changes that overall rates can hide.

The method design also follows from the paper’s own analysis. The authors first find that trait-induced shifts lie mostly in a small subspace, and then build TraSN to constrain only that subspace. Basing the design on an observed property of the model is a sensible approach, even though its benefit is not yet clearly shown.

Critical weaknesses
  1. The paper does not show that restricting the constraint to the subspace actually helps.

Compared with the TIST-Activation baseline, TraSN mainly adds one step: it restricts the activation constraint to the subspace U. Therefore, the value of TraSN depends on whether this restriction helps. The paper supports it in two ways, but neither is convincing.

First, the analysis in Section 3 does not directly justify the subspace that TraSN uses. Its main finding is that traits share a common push toward the benign side. However, U is estimated after this average shift is removed (Eq. 12). As a result, U captures how traits differ from each other, and it may not include the shared push at all. The authors may need to show how much of this push lies in U.

Second, the gains over TIST-Activation are small, and the baselines are run with only one seed. On Gemma, for example, harmful TID is 2.22 for TraSN and 2.41 for TIST-Activation, a gap of only 0.19. In comparison, the seed study on Llama reports a standard deviation of 0.48 for the same metric. This will affect how I interpret the results on Gemma.

  1. The evaluation does not fully support the paper’s two main claims.

The paper claims that TraSN makes the model safer and keeps safety decisions stable across traits. However, the evaluation leaves both claims open.

For the safety claim, the reported refusal rates seem to come from a single condition, most likely the no-trait prompt. If so, the gain in harmful refusal is shown only without traits, and refusal under traits is not reported for any trained model. In addition, the gain comes with more over-refusal. On Llama, harmful refusal rises from 71.38 to 77.75, while over-refusal also rises from 14.00 to 15.50. Part of the gain may therefore come from the model refusing more in general. The authors may need to clarify which condition these rates are computed under, and report refusal rates under traits.

For the stability claim, lower TID and TFR may partly mean that the model pays less attention to the system prompt. To rule this out, the authors may need to check whether the trained model still follows the persona.

Questions for authors
  1. Under which prompt condition did the authors compute the refusal rates in Tables 1–4? Could you also report harmful refusal and over-refusal under traits for each method, separately for the training traits and the held-out traits?

  2. Does the trained model still follow the assigned persona? A persona-adherence score on the existing benign responses would help show that lower TID and TFR do not come from ignoring the system prompt.

  3. How much of the average trait shift, which is removed in Eq. 12, lies in the subspace U? It would also help to compare TraSN with a variant that skips this centering step.

3 comments

Dear Reviewer ZfRg,

Thank you for your careful review and constructive feedback. We address each of your questions below.


[Q1] Refusal rates under trait conditioning

The harmful refusal and benign over-refusal rates in Tables 1–3 are computed under the no-trait condition, τ0\tau_0, using the standard chat template with no explicit system message. Table 4 reports capability results under the same condition.

Below, we report additional results on Gemma-4-E2B for all five methods, separating the 12 training traits (ID) from the three held-out traits (OOD). We average refusal rates across traits within each dataset, then equally across datasets within each group. All values are percentages.

MethodID harmful refusal ↑ID over-refusal ↓OOD harmful refusal ↑OOD over-refusal ↓
Base72.0736.3872.2130.71
TIST-Response84.6526.6084.2526.17
TIST-Logits88.9427.2187.4225.71
TIST-Activation90.9625.8990.4625.58
TraSN92.1925.7191.4625.50

TraSN increases harmful refusal and reduces over-refusal relative to the base model under both trait splits.


[Q2] Retention of benign role expression

To directly assess whether benign roles remain recognizable after tuning, we added a role-identification experiment on Qwen3.5-4B.

We used 100 requests from JBB-Benign and evaluated each request under four roles: pediatrician, high-school chemistry teacher, senior software engineer, and librarian. This yielded 400 responses per method. All three methods used the same requests and role instructions. We masked explicit profession names in each response and asked Claude Haiku 4.5 to infer the assigned role. The judge saw the user’s request, the masked response, and the four candidate role instructions in randomized order. It did not see the assigned role or the method that produced the response. Random guessing yields an accuracy of 25%.

MethodCorrect role identificationsAccuracy
Base model with role prompts282/40070.5%
TIST-Activation (full activation matching)97/40024.25%
TraSN (subspace matching)102/40025.5%

The results show that role-identification accuracy decreases from 70.5% for the base model to 25.5% after TraSN, approaching the random-guessing baseline. Both TIST-Activation and TraSN exhibit similar behavior, suggesting that these methods reduce the distinguishability of role-conditioned responses. This outcome is reasonable consistent with their training objectives, which align trait-conditioned representations with the no-trait reference, either across the full activation space or within the estimated trait subspace.

Accordingly, we have revised the claim in the Introduction from preserving benign role expression under trait prompts to preserving performance on general capability tasks, which is more directly supported by our evaluation. Preserving role-specific expression is not the primary objective of our work, and we have removed the corresponding claim from the manuscript.

Importantly, this finding does not undermine our work, which remains aligned with its core objective of stabilizing LLM safety behavior across different traits. We hope this clarification make the scope and contributions of our work clearer.

Our practical motivation also concerns settings where the model encounters externally supplied instructions containing role or trait assignments at inference time. Omitting persona instructions from a controlled prompt does not establish robustness when such instructions are subsequently introduced. We therefore view trait-conditioned prompting as a potential perturbation or attack surface, and TIST/TraSN as a defense intended to stabilize safety decisions under these inputs. We have incorporated this more detailed discussion into Appendix K Discussion of the Practical Setting and Scope.


[Q3] Coverage of the average trait shift

We apply centering to focus the limited rank-4 subspace on differences across traits, preventing a large shared shift from dominating the principal directions. The resulting subspace can still overlap with the shared mean shift.

In our supplementary analysis on Llama-3.2-3B-Instruct, the centered rank-4 subspace captured 47.2% of the shared mean-shift energy, measured by ∥UΔˉ∥22/∥Δˉ∥22\|U\bar{\Delta}\|_2^2/\|\bar{\Delta}\|_2^2. We estimated this subspace using 100 harmful calibration requests and the 12 training traits at the same layer used by TraSN. This shows that centering retains partial coverage of the common shift identified in Section 3.


Thank you again for your thoughtful review and constructive feedback. If our response has adequately addressed your concerns, we would greatly appreciate your consideration of a higher score. We would also be happy to provide further clarification or discuss any remaining questions.

Thank you for the response and the new experiments. The clarification on the no-trait condition and the Gemma results under traits are helpful. However, my main concerns remain, so I will keep my score.

The new role-identification experiment confirms my main concern. After TraSN, role identification drops to 25.5%, which is the level of random guessing, so the trained model no longer reflects the assigned role. The paper originally defined trait-invariant safety as letting traits shape role and style while keeping safety decisions fixed. Without this part, the goal can be reached by simply ignoring the system prompt. It is therefore unclear what TraSN offers beyond removing the system prompt at inference time.

In addition, TraSN and full activation matching behave almost the same, both in role identification (25.5% vs. 24.25%) and in refusal under traits (92.19 vs. 90.96). The evidence that the subspace step helps is therefore still missing.

Thank you for taking the time to review our additional results and explain your remaining concerns. We appreciate the opportunity to clarify these questions:

  1. We have narrowed the goal and preservation claim. The revised manuscript no longer claims to preserve role and stylistic expression while stabilizing safety decisions. Instead, our goal is to reduce trait-induced safety variation while maintaining performance on the evaluated capability tasks under no-trait prompts. We explicitly acknowledge reduced role distinguishability as a limitation.

  2. Our practical motivation concerns externally supplied trait instructions at inference time. Omitting persona instructions from a controlled prompt does not establish robustness when downstream applications subsequently introduce them. Moreover, prompts always contain complex instructions in which role assignments are intertwined with task requirements, making them difficult to remove without altering the intended task. We therefore view trait-conditioned prompting as a potential perturbation or attack surface, and TIST/TraSN as a defense for stabilizing safety decisions under these inputs. Trait instructions enter through the system message; the defender can modify model parameters before deployment but may not fully control downstream role instructions. We have incorporated this discussion into Appendix K, “Discussion of the Practical Setting and Scope of This Work,” of the revised manuscript.

  3. TIST-Activation and TraSN are both instantiations of TIST, the overarching training framework proposed in our paper. Both align trait-conditioned representations with the no-trait reference, so their comparable reductions in role distinguishability are consistent with their shared objective. We do not interpret this result as evidence of selective safety–style decoupling. The distinction lies in the scope of alignment: TIST-Activation matches the full hidden representation, whereas TraSN constrains only its projection onto the estimated trait subspace. The purpose of this restriction is to improve safety consistency while limiting degradation in general capability. The reported results in the original paper favor TraSN on these measures. We have distinguished the contribution of the shared TIST framework from the incremental benefits of its subspace-based instantiation.

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

The paper identifies trait-induced safety variation: the same user request receives different refusal decisions under different system-prompt traits (roles, personas, personality descriptions), even for benign traits. It proposes refusal-based metrics (TID for dataset-level deviation from the no-trait baseline, TFR for per-request decision flips), and shows representation-level evidence that traits shift harmful-prompt activations along the harmful-benign axis within a low-dimensional subspace (~77-79% variance at rank 4). It then proposes TIST, a self-distillation framework anchoring trait-conditioned behavior to the model's own no-trait behavior, instantiated as TraSN, a LoRA objective penalizing activation discrepancies only within the SVD-estimated trait subspace. On three <=4B open-weight LLMs, TraSN reduces TID/TFR and raises harmful-request refusal while preserving benign compliance and general capability.

Critical strengths
  1. Well-motivated problem formulation. Trait-induced safety variation is real, under-measured, and broader than prior persona-jailbreak work suggests: the paper shows benign roles and personality traits (not only adversarial personas) flip safety decisions, and frames "trait-invariant safety" as a clean invariance requirement. TID and TFR are simple, reusable audit metrics, and TFR's request-level view catches flips that aggregate refusal rates hide.
  2. Coherent mechanism-to-method pipeline. The low-dimensionality result (rank-4 captures ~77-79% of trait-shift variance at the safety layer, aligned with the harmful-benign axis) is consistent across all three models and directly motivates subspace-localized tuning rather than serving as decoration.
  3. The random-subspace control (Table 6) is a genuinely informative ablation: a random rank-4 subspace under the identical objective underperforms TraSN and degrades capability, ruling out generic low-rank regularization as the explanation.
  4. TraSN avoids the degenerate "refuse more" solution: benign TID/TFR improve while over-refusal stays near baseline and capability is preserved or slightly improved. Case studies concretely show both flip directions (trait->compliance on harmful, trait->refusal on benign).
  5. Reasonable experimental hygiene where present: disjoint alignment and evaluation data, held-out traits not used for subspace estimation, an early-stopping dev split, and a multi-seed stability check on Llama.
Critical weaknesses
  1. The key margin over the paper's own ablation is modest. TIST-Activation — literally TraSN minus the subspace projection — reaches harmful TID 3.40 vs 2.57 (Llama) and benign TFR 6.24 vs 4.50; on Qwen, benign TFR is 8.89 vs 7.47. Much of the gain comes from self-distillation, an established idea, rather than from the subspace constraint that is the paper's distinctive claim. Also missing is the strongest zero-cost baseline: inference-time steering/subtraction along the same subspace (the persona-vector recipe).
  2. The experiments do not yet establish whether the claimed mechanism generalizes beyond the three evaluated <=4B models with reasoning disabled. This limits the scope of the generalization claim rather than establishing a minimum model-size requirement. Trait coverage is 15 traits with only 3 held out, providing limited evidence for generalization to unseen traits. Each safety dataset has <=200 prompts (n=100 for two). Without confidence intervals or paired statistical comparisons that account for shared prompts and trait conditions, the reliability of the smaller 1-3-point differences remains unclear. The multi-seed results cover TraSN on Llama only, leaving uncertainty about its margin over TIST-Activation across seeds and model families.
  3. The trait-invariance goal admits a degenerate solution never ruled out: a model that simply attenuates trait conditioning scores TID=TFR=0. No persona-fidelity evaluation is reported, so it is unknown whether TraSN preserves the trait's legitimate influence on style/role or erases it. Robustness under adaptive attacks (PAIR/TAP/PAP, multi-turn) is likewise untested — "objectivity" is shown only for static one-line traits.
  4. The mechanism-to-method link is correlational: the subspace comes from trait shifts on harmful prompts at one layer, and no ablation shows the orthogonal complement is ineffective, nor that the subspace is causally tied to refusal rather than generic style variation. The Claude judge is not validated against human labels, and a 256-token budget can misread delayed compliance.
Questions for authors
  1. Persona fidelity / degenerate solution: TID=TFR=0 is also achieved by a model that ignores trait conditioning entirely. Please report trait-following metrics before vs after tuning (persona consistency, or at minimum whether trait-conditioned outputs still reflect the assigned role on non-safety tasks) to show TraSN preserves the trait's legitimate influence on style rather than attenuating it.
  2. Training-free baseline: please compare TraSN against inference-time steering/subtraction along the same estimated subspace U (the persona-vector recipe), which requires no fine-tuning. If steering achieves comparable TID/TFR, the LoRA training pipeline needs stronger justification; please also report TraSN vs TIST-Activation across seeds with significance tests.
  3. Scope across models: please clarify whether the mechanism and generalization claims are intended to apply beyond the evaluated <=4B, reasoning-disabled setting, and qualify those claims where evidence is unavailable. As an optional extension, an audit on a larger or reasoning-enabled model would help test this scope; additional large-model training is not a prerequisite for answering this question.
  4. Trait coverage: the OOD claim rests on 3 held-out traits. Please clarify the diversity of those traits and the scope of generalization supported by this split. As optional extensions, a broader held-out persona sample (e.g., drawn from the cited PersonaHub resource) or an additional refusal benchmark such as HarmBench or XSTest would strengthen the evidence. These are suggestions for extending the study, not a requirement to complete every listed experiment.
2 comments

Dear Reviewer ysQZ,

Thank you for your careful review and constructive feedback. We address each of your questions below.


[Q1 & W3] Retention of benign role expression

To directly assess whether benign roles remain recognizable after tuning, we added a role-identification experiment on Qwen3.5-4B.

We used 100 requests from JBB-Benign and evaluated each request under four roles: pediatrician, high-school chemistry teacher, senior software engineer, and librarian. This yielded 400 responses per method. All three methods used the same requests and role instructions. We masked explicit profession names in each response and asked Claude Haiku 4.5 to infer the assigned role. The judge saw the user’s request, the masked response, and the four candidate role instructions in randomized order. It did not see the assigned role or the method that produced the response. Random guessing yields an accuracy of 25%.

MethodCorrect role identificationsAccuracy
Base model with role prompts282/40070.5%
TIST-Activation (full activation matching)97/40024.25%
TraSN (subspace matching)102/40025.5%

The results show that role-identification accuracy decreases from 70.5% for the base model to 25.5% after TraSN, approaching the random-guessing baseline. Both TIST-Activation and TraSN exhibit similar behavior, suggesting that these methods reduce the distinguishability of role-conditioned responses. This outcome is reasonable consistent with their training objectives, which align trait-conditioned representations with the no-trait reference, either across the full activation space or within the estimated trait subspace.

Accordingly, we have revised the claim in the Introduction from preserving benign role expression under trait prompts to preserving performance on general capability tasks, which is more directly supported by our evaluation. Preserving role-specific expression is not the primary objective of our work, and we have removed the corresponding claim from the manuscript.

Importantly, this finding does not undermine our work, which remains aligned with its core objective of stabilizing LLM safety behavior across different traits. We hope this clarification make the scope and contributions of our work clearer.

Our practical motivation also concerns settings where the model encounters externally supplied instructions containing role or trait assignments at inference time. Omitting persona instructions from a controlled prompt does not establish robustness when such instructions are subsequently introduced. We therefore view trait-conditioned prompting as a potential perturbation or attack surface, and TIST/TraSN as a defense intended to stabilize safety decisions under these inputs. We have incorporated this more detailed discussion into Appendix K, “Discussion of the Practical Setting and Scope of This Work,” of the revised manuscript.


[Q2 & W1] Training-free steering baselines

We explored training-free inference-time steering along the estimated trait subspace during our initial method development. In these exploratory experiments, performance was sensitive to steering strength: settings tended toward either over-refusal or reduced harmful-request refusal, alongside substantial degradation in general capability. The results were poor, and we will include them as negative results documenting unsuccessful approaches. These observations also motivated the training-based formulation of TraSN.

For significance testing, we followed the original paper’s setup on Llama-3.2-3B across five random seeds.

MethodHarmful Refusal Rate ↑Harmful TID ↓Harmful TFR ↓Benign Over-Refusal Rate ↓Benign TID ↓Benign TFR ↓
None71.386.3812.4214.006.8511.36
TIST-Activation73.08 ± 0.973.37 ± 0.417.52 ± 0.2417.26 ± 0.302.55 ± 0.386.01 ± 0.43
TraSN77.25 ± 1.222.60 ± 0.485.45 ± 0.3216.00 ± 0.650.98 ± 0.244.60 ± 0.22

The results show that TraSN outperforms TIST-Activation stably.


[Q3 & W2] Scope across models and reasoning settings

To examine whether trait-induced safety variation also occurs beyond this setting, we conducted additional behavioral evaluations on Qwen3.5-4B with reasoning enabled and Qwen3.5-9B with reasoning disabled:

SettingHarmful TID ↓Harmful TFR ↓Benign TID ↓Benign TFR ↓
Qwen3.5-4B, reasoning disabled4.52%9.09%13.12%21.00%
Qwen3.5-4B, reasoning enabled8.44%18.05%4.76%9.52%
Qwen3.5-9B, reasoning disabled3.87%7.26%8.95%15.62%

Interestingly, enabling reasoning in Qwen3.5-4B reduces stability on harmful requests while improving stability on benign requests. With reasoning disabled, Qwen3.5-9B exhibits greater stability than Qwen3.5-4B on both request types, although trait-induced variation remains present. These results show that the behavioral phenomenon extends to the tested larger-model and reasoning-enabled settings.


[Q4] Trait coverage and scope of generalization

The three held-out traits represent the three families in our library: survivalist prepper (adversarial role), museum guide (benign role), and curiosity/openness (personality trait). They are excluded from both subspace estimation and training.

This split evaluates generalization to unseen trait prompts within the represented families. We agree that three held-out traits provide limited coverage and will qualify our OOD claims accordingly. Broader persona samples and additional refusal benchmarks would help assess generalization beyond this setting.


[W4] Judge–human agreement

We manually reviewed 30 responses drawn from WJB-Harmful and WJB-Benign using the same refusal criteria as the original evaluation. The human labels agreed with Claude Haiku 4.5 (temperature 0) on all 30 responses (100% observed agreement).


Thank you again for your thoughtful review and constructive feedback. If our response has adequately addressed your concerns, we would greatly appreciate your consideration of a higher score. We would also be happy to provide further clarification or discuss any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.