acceptodds
Under review as a conference paper at ICLR 2027

Understanding and Improving Adversarial Answer Recovery via Activation Steering in Vision–Language Models

Abstract

Activation steering provides a controlled way to modify internal representations without updating backbone parameters. We study its use for *behavioral recovery*: restoring pre-attack responses after adversarial image perturbations while preserving responses to pre-attack inputs. Controlled analyses first examine where and how to intervene. We find that attack status is decodable at the answer-prediction state, yet detector and recovery directions are weakly aligned, and input-specific answer-margin directions recover substantially more responses under matched interventions. These findings motivate AM-GARD, which separates intervention selection from answer correction using a learned intervention gate and input-specific answer-margin updates. Across multiple LVLMs, tasks, and answer formats, AM-GARD improves recovery while maintaining high pre-attack response preservation. Further analyses show that broader selection trades preservation for recovery opportunity, while successful correction depends on identifying a useful answer boundary rather than merely changing the output. Together, these results show that behavioral recovery depends on separating *whether* to intervene from *how* to correct the answer, providing both a diagnostic account of recovery behavior and an effective inference-time intervention.

Primary area:
interpretability and explainable AI
Keywords:
Activation Steering, Vision–Language Models, Adversarial Robustness, Answer Recovery

Reviews

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

This paper proposes AM-GARD, an activation-steering-based defense method for recovering pre-attack responses in vision-language models (VLMs) against adversarial image perturbations. By investigating where interventions should be applied and which directions are effective for answer recovery, they observe that attack-detection directions are weakly aligned with answer-margin directions and that the latter achieve substantially better recovery under controlled interventions. Motivated by these findings, AM-GARD decouples intervention selection from answer correction using a learned intervention gate and input-specific answer-margin updates. The method employs local linearization to identify nearby answer decision boundaries and applies the resulting correction through activation steering without updating the backbone parameters. Experiments on Qwen2-VL-7B and LLaVA-1.5-7B across ten datasets demonstrate improvements in pre-attack answer recovery while largely preserving responses to clean inputs. Overall, the paper presents an interesting investigation of adversarial answer recovery from the perspective of internal representation geometry. The distinction between attack detection and answer correction is well motivated, and the controlled experiments provide useful evidence supporting this separation. However, several aspects of the theoretical motivation and experimental evaluation require further clarification, particularly regarding the relationship between local decision boundaries and the recovery of original answers.

Critical strengths
  • Insightful motivation for decoupling detection and correction.The paper identifies an important distinction between detecting adversarial perturbations and recovering original responses. This observation motivates separating intervention selection from answer correction, providing a meaningful perspective on adversarial defense in VLMs.
  • Effective answer recovery with minimal disruption to clean responses. The proposed method improves the recovery of pre-attack answers while largely preserving the model's original behavior on clean inputs. By combining selective intervention with input-specific answer-margin correction, it achieves a balance between adversarial recovery and clean-response preservation without modifying the backbone parameters.
Critical weaknesses
  • Insufficient justification for nearest-boundary selection. The theoretical analysis only establishes the minimum displacement required to cross a local decision boundary, without explaining its relationship to the pre-attack answer. Consequently, it remains unclear whether selecting the nearest boundary systematically facilitates answer recovery or merely produces coincidental corrections.
  • Limited generalizability of the proposed correction mechanism. The method primarily relies on first-token answer margins, which may not adequately capture the complexity of multi-token generation. Furthermore, its effectiveness under different adversarial objectives and attack distributions remains insufficiently established, raising concerns about its applicability to broader VLM defense scenarios.
Questions for authors
  • Question 1: Justification of nearest-boundary selection: The proposed method assumes that the nearest local answer boundary is a meaningful correction target, but the relationship between the selected competitor and the original pre-attack answer remains unclear. Could the authors provide further theoretical or empirical analysis to verify whether nearest-boundary selection consistently aligns with the original answer, rather than merely leading to alternative outputs that happen to improve recovery.
  • Question 2: Generalization to diverse attack mechanisms: The evaluation mainly focuses on margin-oriented adversarial perturbations, which may align with the proposed answer-margin correction strategy. It remains unclear whether AM-GARD can generalize to attacks that induce different representation shifts, such as patch-based attacks or universal perturbations. More evaluation under diverse attack settings would strengthen the robustness claims of the method.
9 comments
AuthorsOpenReview ↗

Response to Reviewer R3

We sincerely thank the reviewer for the thoughtful and constructive assessment of our work. We particularly appreciate the reviewer's recognition of our central design principle: combining selective intervention with input-specific answer-margin correction to balance recovery of adversarially altered responses and preservation of pre-attack behavior.

The reviewer's concerns address two fundamental questions underlying our approach: (1) why the nearest answer boundary should be relevant to recovering the pre-attack response, rather than merely inducing an alternative output, and (2) whether this correction principle extends beyond first-token margin attacks to different attack mechanisms and generation settings.

To address these questions, we substantially strengthened the manuscript along five directions:

  1. Theoretical justification: We derive an explicit condition under which the nearest answer boundary corresponds to the pre-attack leading token.
  2. Empirical mechanism validation: We examine the relationship between candidate accessibility, correct-boundary selection, genuine behavioral recovery, and unintended response changes.
  3. Attack-strength analysis: We investigate how attack budgets, margin deficits, and attack objectives affect the accessibility and selection of recovery-relevant boundaries.
  4. Attack-dependent recoverability: We evaluate the recovery opportunities remaining under different attacks and identify the stages that limit their realization.
  5. Beyond first-token recovery: We introduce non-oracle answer-position localization experiments for sentence-form responses.

Importantly, the additional attack analyses go beyond evaluating robustness under more attacks. By explicitly separating intervention selection from input-specific answer correction, AM-GARD provides a framework for diagnosing whether recovery is limited by intervention coverage, identification of a useful answer boundary, or the subsequent correction itself.

We address the two concerns below.

AuthorsOpenReview ↗

W1 / Q1: Nearest-Boundary Justification and Behavioral Recovery

We thank the reviewer for highlighting the missing connection between minimum-norm boundary crossing and recovery of the pre-attack response. We agree that reaching the nearest local decision boundary does not, by itself, establish that the boundary corresponds to the original answer.

We have therefore strengthened the analysis to distinguish geometric efficiency from recovery-relevant boundary identification and to evaluate their relationship to actual behavioral recovery.

1. Theoretical relationship between the nearest boundary and the pre-attack answer

Propositions 3.1 and 3.2 characterize the minimum-norm displacement required to cross a fixed linearized token-pair boundary and the effect of directional alignment on the required displacement.

However, these results alone do not establish whether the selected boundary corresponds to the pre-attack answer.

To address this gap, we introduce Proposition 3.3 (Correct-boundary selection condition) in Section 3.2, with its proof and a counterexample in Appendix C.4.

Let aa denote the attacked leading token, yy the pre-attack leading token, and cc another competitor. Define

m_c = z_y - z_c,$$ with margin gradients $$n_{ay} = \nabla_h(z_a-z_y),\qquad n_{ac} = \nabla_h(z_a-z_c).$$ Proposition 3.3 establishes that $$d_y < d_c \iff m_c > g_y\left( \frac{\lVert n_{ac}\rVert_2}{\lVert n_{ay}\rVert_2} - 1 \right).$$ When this condition holds for every other candidate and $y$ belongs to the candidate set, the nearest-boundary rule selects the boundary associated with the pre-attack token. This extends our theoretical analysis from how to reach a selected boundary efficiently to when that boundary is associated with the original answer. The accompanying counterexample also distinguishes this conditional result from an unconditional guarantee. ### 2. From candidate accessibility to genuine behavioral recovery We further examine whether the conditions relevant to correct-boundary selection arise in actual attacked representations. Our analysis follows three stages: **(A) Candidate accessibility.** Under the primary attack, the pre-attack leading token remains among the top four competitors for **96.6–100.0% of flipped inputs** across GQA, POPE, MMBench, and MathVista (Appendix H.1, Table 14). This establishes that the primary attack generally leaves a recovery-relevant candidate accessible across different answer formats. **(B) Correct-boundary selection.** In a separate five-seed diagnostic analysis, the nearest-boundary selector identifies the pre-attack token in **87.7% of available cases on GQA** and **86.3% on MMBench** (Appendix H.1, Table 15). **(C) Behavioral recovery.** More importantly, final response recovery differs substantially depending on whether the selected competitor matches the pre-attack token: | Selected competitor | GQA recovery | MMBench recovery | |---|---:|---:| | Pre-attack token | 89.9% | 93.4% | | Another competitor | 6.9% | 4.0% | These results demonstrate a strong empirical relationship between identifying the pre-attack boundary and recovering the original response, rather than merely producing an alternative answer. The conditional results are complemented by the controlled selector comparison in Section 6.3 (Table 3), which varies the competitor-selection rule while keeping the attacked inputs, intervention decisions, and correction settings fixed. Together, these analyses connect **boundary accessibility, target identification, and genuine response recovery** within the same correction framework.
AuthorsOpenReview ↗

W1 / Q1: Recovery–Preservation Trade-offs and Attack-Strength Analysis (Continued)

3. Recovery and unintended harm: Why not simply select the highest-logit competitor?

We also investigate whether normalized boundary-distance selection offers benefits beyond increasing the frequency of answer changes.

Under the primary attack, Table 3 shows that normalized boundary-distance selection substantially improves recovery over random top-kk selection and achieves aggregate recovery comparable to highest-logit selection.

However, similar aggregate recovery does not imply identical intervention outcomes.

To examine this distinction, we added an analysis separating new recoveries from harmful changes to previously correct responses (Appendix H.1, Figure 11).

DatasetNew recovery: Highest-logit → NormalizedNew harm: Highest-logit → Normalized
GQA67% → 66%44% → 33%
POPE91% → 79%73% → 18%
MMBench86% → 84%93% → 80%
MathVista79% → 76%98% → 98%

On GQA, normalized selection maintains nearly identical recovery while reducing harmful changes. On POPE, it substantially reduces harm, accompanied by a reduction in recovery.

These measurements use separate conditioning populations for recovery and harm and are therefore interpreted as complementary outcomes rather than a single net-effect score.

These findings provide an empirical reason to use normalized boundary-distance selection beyond its minimum-displacement interpretation. Unlike raw logit ranking, it considers both the answer margin and local gradient geometry when selecting an input-specific correction target. Proposition 3.3 characterizes when this criterion selects the pre-attack token's boundary, while Table 15 connects correct-target selection to actual recovery. The controlled comparison (Table 3, Figure 11) further shows competitive recovery with fewer harmful changes in several settings: on GQA, new recovery is nearly unchanged (66% versus 67%) while harm falls from 44% to 33%; on POPE, harm falls from 73% to 18%, accompanied by lower recovery (79% versus 91%). Thus, normalized boundary-distance selection offers a geometrically grounded recovery--preservation trade-off, rather than merely inducing alternative outputs or claiming uniform superiority over simpler selectors.

4. Attack-strength analysis: When does correct-boundary selection become difficult?

We further investigate how the relationship between the nearest boundary and the pre-attack answer changes with attack strength and objective (Section 6.3, Figure 6).

The analysis distinguishes three factors.

Attack-budget sweep (S1–S4). Increasing the optimization budget does not immediately eliminate the pre-attack token from the candidate set. Candidate accessibility remains relatively stable across the evaluated primary-attack budgets.

Attack-depth analysis. In contrast, candidate accessibility generally decreases as the pre-attack token's margin deficit gyg_y increases. This indicates that the depth of the attacked answer relative to the original answer is more informative than optimization budget alone.

Matched-depth comparison. Even at comparable margin deficits, correct-boundary selection depends on the attack objective.

For gy∈[5,10)g_y \in [5,10), the observed correct-boundary selection rates are:

DatasetFirst-token margin attackSequence-NLL attack
GQA28%2%
POPE76%37%

Thus, the accessibility and selection of a recovery-relevant boundary depend not only on attack depth but also on the answer-boundary structure induced by the attack objective.

This result provides an empirical characterization of when nearest-boundary selection remains useful and when its ability to identify the pre-attack answer becomes limited.

Revisions: We added Proposition 3.3 and its proof in Section 3.2 and Appendix C.4; strengthened the controlled selector and attack-strength analyses in Section 6.3 (Table 3 and Figure 6); and added candidate-accessibility, conditional-recovery, and recovery–harm diagnostics in Appendix H.1 (Tables 14–15 and Figure 11).

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

This paper works on utilizing activation steering to restoring responses after adversarial image perturbations. Activation steering is a controlled way to modify internal representations without updating backbone parameters. The authors use it for behavioral recovery: restoring pre-attack responses after adversarial image perturbations while preserving responses to pre-attack inputs.

The authors first conduct experiments to examine where and how to intervene. They find that attack status is decodable at the answer-prediction state, yet detector and recovery directions are weakly aligned, and input-specific answer-margin directions recover substantially more responses under matched interventions. Then, they propose AM-GARD, which separates intervention selection from answer correction using a learned intervention gate and input-specific answer-margin updates.

In experiments, across multiple LVLMs, tasks, and answer formats, AM-GARD improves recovery while maintaining high pre-attack response preservation. Further analyses show that broader selection trades preservation for recovery opportunity, while successful correction depends on identifying a useful answer boundary rather than merely changing the output.

Critical strengths
  • The figures and tables are well-designed.

  • The proposed method AM-GARD shows improved results across multiple LVLMs, tasks, and answer formats.

  • The authors conducted comprehensive analyses, supporting their arguments and designs.

Critical weaknesses
  • AM-GARD needs additional data for pre-attack and attacked images and use the data for calibration. I'm wondering if this will limit the practical value and robustness of the method. For example, if we have no data for the attacking or if there is a new attack, the method may fail to recover responses.

  • The experiments are mainly based on Qwen2-VL-7B and LLaVA-1.5-7B, which are kind of old. I'm not sure if the findings and the proposed method still works on latest VLMs.

Questions for authors
  • I'm sorry I am not familiar with either activation steering or adversarial image perturbation on VLMs, so I'm afraid that I might miss some important issues. I need to read other reviewers' comments and then make a decision.

  • I'm wondering how to evaluate it on open-ended tasks. Maybe the model gives different but same-meaning responses. And when we use some sampling strategies, the model will give slightly different answer each time. In these cases, behavior recovery for the exact response seems kind of too strict.

7 comments
AuthorsOpenReview ↗

Response to Reviewer R2

We sincerely thank the reviewer for the thoughtful and constructive assessment of our work. We particularly appreciate the reviewer's recognition that AM-GARD improves recovery across multiple LVLMs, tasks, and answer formats, and that our comprehensive analyses support the proposed design and its underlying arguments.

The central objective of our work is not merely to improve prediction accuracy under adversarial attacks, but to understand and improve how an LVLM can recover its own pre-attack responses through interventions in its internal representations.

Our controlled experiments first reveal that the representation directions useful for detecting adversarial inputs differ from those effective for recovering the original response. Under identical intervention masks and matched update norms, the detector direction achieves only 1.85% recovery, whereas the input-specific answer-margin direction achieves 83.39% recovery (Figure 3). Their weak directional alignment further supports this distinction.

Motivated by this finding, AM-GARD explicitly separates intervention selection from input-specific answer-margin correction. This separation not only enables behavioral recovery but also provides a structured way to investigate which stages govern its success and failure.

The reviewer's concerns regarding attack-data dependence, generalization to newer LVLMs, and evaluation of open-ended responses are important for examining whether these recovery principles extend beyond the original experimental conditions.

Accordingly, we have strengthened the revised manuscript through three additional lines of investigation:

  1. Attack-dependent recoverability: We analyze the recovery opportunities remaining under different attacks and identify the stages that limit their realization.
  2. Additional backbone evaluation: We extend the experiments to InternVL3.5-8B, evaluating recovery and pre-attack response preservation across ten datasets.
  3. Semantic behavioral recovery: We evaluate recovery using semantic agreement in addition to exact match, accounting for responses that differ in wording but convey equivalent meanings.

We address each concern below.

AuthorsOpenReview ↗

W1. Dependence on Attack Data and Generalization to Unseen Attacks

We thank the reviewer for raising the important concern that reliance on pre-attack and attacked data for calibration may limit the practical applicability and robustness of AM-GARD, particularly when attack data are unavailable or new attack mechanisms emerge.

Our approach is motivated by recovering the model's original response through input-specific correction at the observed answer-prediction state, rather than learning a separate correction rule for each attack type. To investigate the implications of this design, we distinguish the information requirements of intervention selection and answer correction and evaluate their behavior under different attacks.

1. Information requirements during calibration and inference

AM-GARD explicitly separates the intervention gate, which determines whether correction should be applied, from input-specific answer-margin correction, which determines how to modify the representation.

The Probe-Gate is trained using pre-attack and attacked representations, while the gate threshold and correction settings are selected using calibration data. However, at inference time, AM-GARD does not require the paired pre-attack image, pre-attack response, ground-truth answer, or attack identity.

Importantly, answer-margin correction does not rely on a correction direction learned for a particular attack category. Instead, it computes an input-specific direction from the token-pair margins and gradients at the observed answer-prediction representation.

Thus, attack-data dependence primarily arises from learning and calibrating intervention selection, rather than from learning attack-specific correction directions.

We also evaluate Outlier-Gate, which is fitted using only pre-attack representations and shares the same answer-margin corrector as Probe-Gate (Section 5, Table 2). On GQA, it improves recovery over no-op from 42.4% to 45.7% on Qwen2-VL-7B and from 17.4% to 24.2% on LLaVA-1.5-7B, with pre-attack preservation of 91.0% and 91.4%, respectively. These results demonstrate that attacked examples are not inherently required to fit an intervention gate that enables nonzero recovery gains, although Probe-Gate achieves higher recovery in these settings when attacked data are available. This clean-only alternative still requires pre-attack data and appropriate calibration; it does not establish that the deployed pipeline is invariant to unseen attacks.

2. Recovery opportunities and bottlenecks under different attacks

To investigate recovery beyond the original first-token margin attack, we extended our analyses to sequence-NLL attacks, continued margin optimization, adversarial patches, universal perturbations, and gate-aware attacks (Section 6.3, Appendix I).

These experiments examine not only recovery under different attacks but also how attack-dependent changes in answer representations affect the possibility of restoring pre-attack behavior. By separating intervention selection from answer correction, AM-GARD allows us to identify which stages constrain recovery.

Specifically, we compare three conditions while retaining the same answer-margin corrector: Deployed, using the learned Probe-Gate and competitor selector; Oracle Gate, providing ideal intervention decisions; and Oracle Gate + Target, additionally providing the pre-attack target (Appendix I.2, Table 19).

On GQA, the primary margin attack achieves 70.5% deployed recovery, which increases to 88.5% with Oracle Gate and 92.9% with Oracle Gate + Target. Under sequence-NLL, deployed recovery is substantially lower at 4.3%, while Oracle Gate + Target increases recovery to 36.6%. The larger recoverable gap therefore shifts from intervention selection under the primary attack to target identification under sequence-NLL.

These findings demonstrate that the effectiveness of behavioral recovery depends on the answer-boundary structure remaining after an attack. Importantly, the oracle-assisted results distinguish recovery opportunities accessible to the correction mechanism from those realized by the deployed system.

Thus, the additional experiments extend our understanding of when input-specific answer-margin correction can restore pre-attack behavior and how different attack objectives constrain that recovery, rather than merely reporting robustness across additional attacks.

AuthorsOpenReview ↗

W1: Correction Reach and Generalization Across Attack Families (Continued)

3. Correction reach under patch and universal perturbations

We further investigated whether answer-margin correction remains effective under attacks that do not directly optimize the first-token margin, including adversarial patches and universal perturbations (Appendix I.4, Figure 14).

We evaluate recovery under oracle intervention and target information to measure the recovery opportunities accessible to the existing corrector.

Attack mechanismGQAPOPEMMBenchMathVista
Continued margin51.2%54.1%96.2%43.8%
Adversarial patch80.3%85.7%97.6%56.2%
Universal perturbation90.9%100.0%100.0%83.3%

These results provide empirical evidence that the recovery capability of answer-margin correction is not inherently restricted to first-token margin attacks.

In particular, substantial behavioral recovery remains possible under patch and universal perturbations when appropriate intervention and target information are available.

Rather than establishing identical deployed robustness across attack families, these experiments demonstrate that the proposed correction mechanism can exploit recovery opportunities remaining in the internal representations under different attack mechanisms.

4. Transfer of intervention selection to unseen attacks

We additionally examined whether the intervention gate transfers across attack families (Appendix I.4, Table 21).

Within a separately reconstructed evaluation environment, a gate trained on minimum-budget margin attacks was evaluated on other attack mechanisms without attack-specific refitting.

Gate detection remained relatively effective under continued margin optimization and the evaluated universal perturbations, whereas patch attacks exhibited reduced intervention coverage.

Nevertheless, replacing the transferred gate with oracle intervention decisions substantially improved recovery on several patch-attack settings.

Patch attackTransferred-gate recoveryOracle-gate recovery
GQA18.0%44.3%
MMBench48.8%73.2%

These findings indicate that recovery opportunities may remain available to the answer-margin corrector even when intervention selection does not successfully identify the attacked inputs.

Together, the clean-only Outlier-Gate evaluation and cross-attack diagnostics demonstrate two complementary properties of AM-GARD: intervention selection can be constructed under different data requirements, while answer-margin correction remains input-specific without learning attack-specific correction directions. Moreover, the explicit separation of selection and correction makes it possible to examine how intervention coverage, target identification, and correction reach limit behavioral recovery under different attack objectives and perturbation mechanisms. Thus, these experiments strengthen both the evaluation of attack transfer and our mechanistic understanding of when recovery remains possible.

Revisions: We clarified the information-access assumptions during calibration and inference in Sections 2 and 4. We also expanded the attack-strength, sequence-NLL, patch, universal-perturbation, and gate-transfer analyses in Section 6.3 and Appendix I.

Reviewer_2ugxOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the recovery of responses in LVLMs under adversarial attacks by adjusting activation functions. It demonstrates that attack detection and response recovery require distinct intervention strategies. They propose AM-GARD, a method that combines a learned intervention gate with input-specific response error correction. Experiments conducted across two LVLMs and ten datasets show that this approach improves recovery performance while largely preserving the original, pre-attack responses.

Critical strengths
  • It provides insightful distinction between attack detection and response recovery.
  • Extensive experimental evaluation covers proposed method improves recovery performance while largely preserving the original, pre-attack responses.
Critical weaknesses
  • Performance is inconsistent across different models, and the advantages over existing defense mechanisms are limited. As shown in Table 2, while AM-GARD may perform better in terms of preserving pre-attack performance on certain datasets—such as LLaVA GQA—its recovery performance is significantly lower than that of F3 (45.3% v.s. 90.2%). Furthermore, on the LLaVA MathVision and MathVista datasets, AM-GARD performs even worse than the "no-op" (no-action) baseline. These results raise questions regarding the effectiveness of this methodology across a diverse range of models and tasks; it also remains unclear under what conditions it actually offers a tangible benefit compared to existing defense techniques.
  • Experiments involving sentence-form responses rely on oracle answer-position localization, a method that uses the model’s pre-attack response to pinpoint the exact location for the correct answer, which is information unavailable in real-world deployment scenarios. Furthermore, the computational cost associated with iterative gradient-based correction has not been quantified, making it difficult to assess its practical efficiency.
Questions for authors
  • As shown in Table 2, F3 achieves a significantly higher recovery performance (90.2% v.s. 45.3%) compared to AM-GARD on LLaVA GQA dataset, while maintaining over 90% preservation. Explain this performance gap and clarify the practical advantages of AM-GARD.
  • As shown in Table 2, AM-GARD performance on LLaVA MathVision and MathVista datasets is lower than No-op baseline. What is the cause of this performance degradations, and how does it relate to the proposed correction mechanism?
  • Table 3 indicates that the use of normalized boundary-distance selection lead to a slight improvement in recovery performance(60.7% v.s. 57.6%). Could you confirm whether this improvement is statistically significant and whether it is consistently observed across other models or datasets.
8 comments
AuthorsOpenReview ↗

Response to Reviewer R1

We sincerely thank the reviewer for the thoughtful and constructive assessment of our work. We particularly appreciate the recognition of the distinction between adversarial attack detection and behavioral recovery, as well as the extensive evaluation across LVLMs, tasks, and answer formats.

Our study aims to both understand and improve adversarial behavioral recovery through activation-level interventions. While existing defenses primarily intervene through image purification, generation guidance, or general-purpose representation steering, our work investigates a complementary question: how can an attacked LVLM recover its own pre-attack response through its internal representations, without access to the original input or response at inference time?

Our controlled experiments reveal that detecting an attacked state and restoring its original response require different intervention directions. In particular, under identical intervention masks and matched update norms, the detector direction achieves only 1.85% recovery, compared with 83.39% for the input-specific answer-margin direction (Figure 3). Their weak directional alignment further supports this distinction.

These findings motivate AM-GARD, which explicitly separates whether to intervene from how to recover, operating directly on the answer-prediction representation while keeping the observed image and backbone parameters unchanged.

To address the reviewer's concerns and further substantiate these contributions, we substantially strengthened the manuscript through:

  1. Recovery and failure analyses: Additional paired evaluations identifying the sources of model-dependent performance differences, including the F3 comparison and mathematical-task regressions.
  2. Practical applicability: Expanded recovery–preservation comparisons with input-space defenses and computational profiling across three LVLMs.
  3. Generalization beyond the primary setting: Non-oracle sentence-form recovery, alternative attack objectives, and attack-family transfer evaluations.
  4. Theoretical and empirical justification: A new correct-boundary selection condition and experiments connecting candidate accessibility, boundary selection, and behavioral recovery.

We address each concern individually below.

AuthorsOpenReview ↗

W1, Q1, Q2: Model-Dependent Performance and F3 Recovery Gap

We appreciate the reviewer's concern regarding inconsistent performance across models and the practical advantages of AM-GARD relative to existing defenses. We agree that these questions are important for establishing the applicability of the proposed mechanism.

Our additional analyses clarify where AM-GARD improves recovery, why certain settings remain challenging, and which recovery–preservation–efficiency trade-offs it provides.

1. Overall effectiveness across models and tasks

Before examining the individual exceptions, we would like to clarify the overall recovery results.

Across the 20 model–task configurations in Table 2, AM-GARD with Probe-Gate improves behavioral recovery over no intervention in 18 configurations: all 10 tasks on Qwen2-VL-7B and 8 of 10 on LLaVA-1.5-7B.

This includes open-ended, binary, multiple-choice, and numeric answer formats. On LLaVA GQA, for example, recovery increases from 17.4% with no-op to 45.3% with AM-GARD, while preserving 96.6% of pre-attack responses.

These results demonstrate that input-specific representation-level intervention can restore pre-attack responses across diverse tasks without modifying the visual input or backbone parameters. Importantly, they establish improvement over unmodified inference, rather than implying superiority over every existing defense.

We further investigated the model-dependent differences highlighted by the reviewer.

2. Understanding the recovery gap with F3 on LLaVA GQA

We acknowledge that F3 achieves substantially higher recovery on LLaVA GQA at the default operating point (90.2% vs. 45.3%), while AM-GARD achieves higher pre-attack preservation (96.6% vs. 91.2%; Table 2).

F3 performs cross-modal-attention-based image purification before answer generation, intervening directly on the visual input. This mechanism is well suited to counteracting image-level perturbations in the evaluated LPIPS-constrained PGD setting, where small visual changes can alter the model's response. AM-GARD instead acts on the resulting answer-prediction representation and selectively attempts to restore the pre-attack behavior. The alignment between F3's intervention locus and the attack surface offers a plausible explanation for its strong results here, distinct from the intervention-coverage gap measured below.

To identify the observed bottlenecks more directly, we performed an additional five-seed paired failure analysis, examining cases recovered by F3 but not AM-GARD.

Failure categoryShare (%)
Gate non-activation59.6
Trust-based abstention15.2
Failure despite selecting the pre-attack boundary16.6
Candidate unavailable or incorrect boundary selected6.3
Unclassified / unresolved (residual; pending audit)2.3

All shares use the F3-success / AM-GARD-failure paired cohort as the denominator. The final 2.3% is the arithmetic residual of the reported categories, not yet a separately verified failure mechanism; its classification requires confirmation against the per-example logs.

The analysis identifies intervention coverage as the largest source of missed recovery opportunities. Gate non-activation and trust-based abstention together account for 74.8% of the analyzed cases.

This finding helps explain the substantial difference between the two approaches. Under the evaluated configuration, F3 applies image purification to every input, whereas AM-GARD selectively applies representation-level correction. Consequently, a conservative intervention policy can preserve the model's original behavior while leaving some recoverable attacked inputs unmodified.

The remaining cases further distinguish failures associated with identifying the relevant answer boundary from those occurring despite selecting the pre-attack boundary. This analysis therefore connects the practical performance gap to the two components central to AM-GARD: intervention selection and answer correction.

AuthorsOpenReview ↗

W1, Q1, Q2: Recovery–Preservation Trade-offs and Practical Advantages (Continued)

3. Understanding the results below no-op on mathematical tasks

We also investigated the negative recovery differences on LLaVA MathVision and MathVista through paired comparisons with no-op under the original evaluation protocol.

DatasetNewly recoveredNewly disruptedNet difference
MathVision3445−11 / 535 (−2.1 pp)
MathVista14−3 / 80 (−3.75 pp)

On MathVision, AM-GARD recovers 34 additional responses but disrupts 45 previously preserved responses. Thus, the observed decrease arises from an unfavorable balance between successful recovery and unintended answer changes, rather than an absence of recovery.

This illustrates why intervention selection is essential to behavioral recovery: applying a correction is beneficial only when its recovery gain outweighs the risk of altering a response that should remain unchanged.

For MathVista, the aggregate decrease corresponds to only three net responses across 80 seed–example evaluations. Given the small evaluation set and substantial seed variability, we interpret this result cautiously rather than as evidence of systematic degradation.

These paired analyses provide a more informative interpretation of the two exceptions than aggregate recovery percentages alone.

4. Comparison with input-space defenses and practical advantages

We further extended the comparison with F3 beyond the default operating points in Table 2.

Existing defenses intervene at different levels: F3 purifies the visual input, DPS guides generation using partial visual evidence, and conventional activation-steering methods commonly apply fixed or answer-agnostic representation modifications.

In contrast, AM-GARD operates on the answer-prediction representation and computes an input-specific correction from local answer-boundary geometry. This design enables direct intervention in the model's answer decision while retaining the observed image and frozen backbone.

Our additional recovery–preservation comparison (Appendix J.1, Figure 15) examines AM-GARD and F3 across gate and purification settings.

On Qwen2-VL GQA, the methods achieve comparable recovery over all attacked inputs at selected operating points, with AM-GARD offering higher pre-attack response preservation in some settings. At approximately 80% preservation, AM-GARD and F3 achieve 73% and 71% recovery, respectively (Appendix J.1, Figure 15).

On LLaVA GQA, F3 performs better across the evaluated settings, consistent with the coverage limitations identified in our paired analysis.

We also measured computational cost (Appendix J.3, Table 23). AM-GARD achieves lower all-input latency than F3 on Qwen2-VL-7B and InternVL3.5-8B, although F3 is faster on LLaVA-1.5-7B.

Together, these results establish model-dependent practical operating conditions for selective representation-level correction, rather than suggesting a universal advantage over input purification.

5. Implications for the contribution of our work

Our broader research objective is to investigate behavioral recovery through the model's answer-prediction state without designing a separate correction rule for each attack family.

AM-GARD operationalizes this objective through an input-specific answer-margin update, while its intervention gate determines whether correction should be attempted. The correction mechanism is not conditioned on a predefined attack category.

To further examine this principle, we extended the analysis to sequence-NLL attacks, continued margin attacks, adversarial patches, and universal perturbations (Section 6.3 and Appendix I). These experiments identify how recovery depends on attack-induced changes in intervention coverage, target accessibility, and correction reach.

The resulting contribution is therefore twofold: we identify the distinct mechanisms governing behavioral recovery and develop an inference-time method that uses this understanding to recover pre-attack responses.

Rather than positioning AM-GARD solely as a replacement for image purification, we view selective representation-level correction as a complementary intervention mechanism whose effectiveness can be analyzed and controlled at the answer-prediction level.

Revisions: We strengthened the discussion of model-dependent recovery in Section 6.1, added recovery–preservation operating-point comparisons in Appendix J.1 (Figure 15), and provided computational profiling in Appendix J.3 (Table 23). Additional paired failure analyses are reported above to directly address the reviewer's concerns.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Common Response to All Reviewers

Dear Reviewers,

We sincerely thank all reviewers for their thoughtful and constructive feedback. We appreciate your recognition of the distinction between attack detection and behavioral recovery, the motivation for separating intervention selection from answer correction, and the extensive experimental analyses.

In response to your concerns, we have substantially expanded our theoretical analysis, experimental evaluation, and discussion of practical limitations. We summarize the main additions below; detailed explanations are provided in our individual responses.

1. Analyzing Nearest-Boundary Selection and Its Relationship to Recovery (BkmS Q1; 2ugx Q3).

We introduced Proposition 3.3 to characterize the conditions under which nearest-boundary selection identifies the pre-attack leading token's boundary. Additional experiments examine candidate availability, correct-boundary selection, and their relationship to actual recovery. We also compare normalized boundary-distance selection with simpler alternatives, distinguishing successful recovery from unintended answer changes. Together, these analyses connect the geometric criterion for boundary selection to behavioral recovery while clarifying why crossing the nearest boundary does not necessarily restore the pre-attack response (Section 3.2, Section 6.3, Appendix H).

2. Recovery–Preservation Trade-offs and Computational Efficiency (2ugx W1, Q1–Q2).

We expanded comparisons with F3 and other input-space defenses across recovery–preservation operating points, investigated model-dependent performance differences and below-no-op results, and measured inference latency and GPU memory across three LVLMs. These analyses clarify the recovery–preservation–efficiency trade-offs, highlighting the practical benefits of selective representation-level correction relative to input-space defenses across different operating conditions (Section 6.1, Appendix J).

3. Recovery Behavior Across Models and Attack Mechanisms (MLrf W1–W2; BkmS Q2).

We extended the evaluation to InternVL3.5-8B and investigated sequence-NLL attacks, continued margin optimization, adversarial patches, universal perturbations, and gate-aware attacks. By separating deployed recovery from oracle-assisted correction, we identify how the dominant recovery bottleneck shifts between intervention coverage, target identification, and correction effectiveness across attack conditions. These findings clarify when representation-level recovery remains feasible and why its effectiveness varies across attacks (Appendix F.1, Appendix I).

4. Analyzing the Limitations of First-Token Interventions in Sentence-Form Recovery (2ugx W2; BkmS W2).

Our method primarily operates on the first response-token prediction state, limiting its applicability when answers appear at later positions. To examine this limitation, we evaluated two non-oracle localization strategies and analyzed how intervention placement affects sentence-form recovery (Section 6.4, Appendix K).

We also clarified semantic-agreement scoring to account for equivalent answer expressions beyond exact string matching (Appendix E.5; MLrf Q).

Together, these additions strengthen our central contribution: 1) understanding how activation-level interventions recover pre-attack behavior, 2)distinguishing intervention selection from answer correction, and 3) identifying the key factors that govern effective behavioral recovery.

We appreciate your continued consideration and welcome any further questions.

Best regards,
The Authors

AuthorsOpenReview ↗

Revised Paper Uploaded

Dear Reviewers,

We have uploaded a revised version of the manuscript. All revisions are highlighted in blue, and the main text remains within the 10-page limit.

The major revisions are summarized below, together with their corresponding reviewer concerns.

Main Text

  • Section 3.2: Added Proposition 3.3, establishing a condition under which nearest-boundary selection identifies the pre-attack answer boundary (BkmS Q1).
  • Section 6.1: Clarified model-dependent recovery performance, recovery–preservation trade-offs, and computational efficiency (2ugx W1, Q1–Q2).
  • Section 6.3: Expanded controlled competitor-selection experiments and analyses of attack strength, boundary accessibility, and correct-boundary selection (Table 3, Figure 6; BkmS Q1, 2ugx Q3).
  • Section 6.4: Analyzed the limitations of first-token interventions through non-oracle sentence-form recovery experiments (2ugx W2; BkmS W2).

Appendix

  • Appendix C: Added the proof of Proposition 3.3 and a counterexample illustrating the limitations of nearest-boundary selection.
  • Appendix E: Clarified evaluation protocols, semantic-agreement scoring, and inference-time information access.
  • Appendix F.1: Added recovery and preservation results for InternVL3.5-8B (Table 11; MLrf W2).
  • Appendix H: Expanded candidate-accessibility, conditional-recovery, and recovery–harm analyses (Tables 14–17, Figure 11; BkmS Q1, 2ugx Q3).
  • Appendix I: Added evaluations across alternative attack objectives, perturbation mechanisms, and gate-transfer settings (Tables 18–21, Figures 13–14; BkmS Q2, MLrf W1).
  • Appendix J: Expanded recovery–preservation comparisons with input-space defenses and computational profiling across three LVLMs (Figure 15, Table 23; 2ugx W1–W2).
  • Appendix K: Extended sentence-form recovery experiments to analyze the effects of non-oracle answer-position localization (Figure 16; 2ugx W2; BkmS W2)

We thank all reviewers again for their valuable feedback and consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.