acceptodds
Under review as a conference paper at ICLR 2027

CARS: Convergence-Aware Representation Steering for Multimodal Large Language Model Safety

Abstract

Multimodal Large Language Models (MLLMs) have achieved remarkable vision-language understanding through extensive pre-training, yet remain critically vulnerable to harmful visual inputs. Existing defenses either require costly retraining or introduce substantial overhead. To address this critical vulnerability, we introduce Convergence-Aware Representation Steering (CARS), an inference-time method exploiting natural cross-modal representation evolution during generation. Through systematic empirical analysis, we discover delayed refusal—a phenomenon where text-based safety steering often initially fails on harmful images but progressively activates as visual representations converge toward text-like patterns. CARS is designed to actively monitor this convergence and strategically enhances safety steering only when visual representations sufficiently align with text-based safety patterns. Experiments across FigStep, MM-SafetyBench, ToViLaG, and Jailbreak-28K show CARS achieves 92.3% average defense success rate—substantially outperforming state-of-the-art methods—while maintaining near-baseline utility with minimal degradation (0.1-1.3%) on ScienceQA and MM-Vet. Our work demonstrates that respecting natural representation dynamics enables superior safety-utility tradeoffs without architectural modifications or retraining.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial Attacks & Robustness, Generative AI Safety, Provenance & Misuse

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

The authors propose CARS as an inference time defense for MLLMs They first prepare two sets of text instructions containing benign and harmful requests These instructions are fed into the target MLLM The hidden state of the last input token is extracted at each layer The difference between the mean hidden states of the two groups gives the safety steering direction for each layer A logistic regression classifier is also trained on text hidden states to distinguish benign from harmful content Classification F1 scores across different intervention layer ranges and generation positions are compared to select the target layers and a fixed detection position During inference the model first applies steering at the base strength to the target layers At the preset generation position the classifier reads the hidden state of the current token across the target layers and averages the predictions to obtain a harmfulness probability If the input is classified as benign further steering stops and the generated content is retained If it is classified as harmful the steering strength increases and the output before the detection position is discarded Generation then continues

Critical strengths
  1. The method considers when detection and intervention become effective during generation It starts with steering at a lower strength performs classification at a later position and increases steering when harm is detected This design follows the observed failure pattern

  2. The experiments cover three models and four safety benchmarks including both text and image attacks

  3. CARS operates during inference without updating the MLLM parameters It uses direction vectors and logistic regression which gives it some deployment value

Critical weaknesses
  1. The method modifies hidden states using text safety directions and then uses a text safety classifier to assess their classifiability and guide further intervention These operations rely on signals from a single modality and do not establish cross modal alignment or convergence of the representation distributions The improvement may also come from safety related content in the generated prefix that the classifier can recognize more easily The reasons for using only text safety directions and a text classifier are also not fully supported The poor results from image and text mean difference directions only show that this construction has limited effect in the tested setting They do not rule out the value of image and text data The paper lacks a comparison of detectors trained on text data image and text data and a mixture of both under the same conditions It also does not sufficiently test directions extracted from image and text representations at later generation positions

  2. The detection position is a fixed hyperparameter chosen separately for each model. The 10th or 15th generation position is selected using offline classification performance. The method does not check whether the current input has reached convergence during inference. Figure 4 shows that changes in early detection positions can affect defense performance. It remains unclear whether these fixed positions transfer across tasks, response lengths, and input distributions

  3. The ablations on intervention layers and steering strength are limited. All three models use 12 intervention layers with a base strength of 3.5 and an enhanced strength of 6.0. Although the paper compares two strength settings, it does not test a wider range or fully report the effects of changing the starting layer and the number of intervention layers. The experiments therefore do not establish how stable these settings are or how much they depend on calibration for each model

  4. CARS makes a decision only once at a fixed position. It does not check again after a benign decision or verify the later output after a harmful decision. Its ability to recover from an initial missed detection has therefore not been evaluated separately. Later detection also means longer intervention and a longer discarded prefix. These factors may explain part of the defense improvement, so the gain cannot be attributed entirely to convergence

Questions for authors
  1. Can the authors provide evidence of convergence that is independent of the text safety directions and text safety classifier? How can they rule out shifts relative to the classification boundary and the effect of safety-related words in the generated prefix as explanations for the F1 improvement?

  2. Under the same sample budget, how do detectors trained on text data, image and text data, and a mixture of both compare? Do safety directions extracted from image and text representations at later generation positions remain ineffective?

  3. What calibration data and selection rule are used to choose the 10th or 15th position? How are these data separated from the final test set? Do the selected positions transfer across datasets?

  4. How much does performance change when the starting layer or the number of intervention layers changes? Can the authors provide a joint sensitivity analysis of the intervention layer range and detection position?

  5. Can the authors test a wider range of steering strengths and intervention layer settings?

  6. How does CARS check whether later content is harmful after the first detection? What are the benefits and costs of repeated detection or a final output check?

10 comments
AuthorsOpenReview ↗

Response to Reviewer CifH (1/4)

Q1: Can convergence be demonstrated independently of the text safety direction and classifier boundary?

Thanks for your constructive feedback. To address this concern, we construct matched datasets following the basic idea of FigStep. We sample 128 harmful prompts from AdvBench and 128 benign prompts from Alpaca. For every prompt, we create one text-only example and one (image,text) pair by rendering the prompt itself as text in an image. All image examples use the same instruction. This produces four datasets, DvlharmD_{vl}^{\mathrm{harm}}, DtextharmD_{text}^{\mathrm{harm}}, DvlbenignD_{vl}^{\mathrm{benign}}, and DtextbenignD_{text}^{\mathrm{benign}}, so comparisons across modalities keep the semantics fixed, while comparisons across harmfulness labels keep the modality fixed. Examples are provided in Appendix A.1, Fig. 5.

We first analyze representational proximity without applying any steering on two architecturally different models, LLaVA-1.5-7B and Qwen2.5-VL. At each output position, we compute cosine similarity across layers for four dataset pairings. The two within-modality pairings compare harmful and benign semantics, whereas the two cross-modal pairings compare matched semantics. Higher cosine similarity indicates greater representational proximity. The complete curves are provided in Appendix A.2, Fig. 6, and selected token positions are reported below.

ModelPairToken 0Token 1Token 2Token 3Token 4Token 5Token 10Token 15
LLaVA-1.5-7B(Dvlharm,Dtextharm)(D_{vl}^{harm}, D_{text}^{harm})0.4640.2640.2440.2110.1790.2290.2140.267
LLaVA-1.5-7B(Dvlbenign,Dtextbenign)(D_{vl}^{benign}, D_{text}^{benign})0.4740.3400.3070.3290.2140.2380.2590.246
LLaVA-1.5-7B(Dvlharm,Dvlbenign)(D_{vl}^{harm}, D_{vl}^{benign})0.9430.9800.9510.6790.3410.3420.2510.209
LLaVA-1.5-7B(Dtextharm,Dtextbenign)(D_{text}^{harm}, D_{text}^{benign})0.3860.2320.2000.1510.1570.1510.1600.152
Qwen2.5-VL(Dvlharm,Dtextharm)(D_{vl}^{harm}, D_{text}^{harm})0.4380.2470.1920.1520.2140.2440.2700.326
Qwen2.5-VL(Dvlbenign,Dtextbenign)(D_{vl}^{benign}, D_{text}^{benign})0.4560.3430.3730.3080.2960.3160.3340.289
Qwen2.5-VL(Dvlharm,Dvlbenign)(D_{vl}^{harm}, D_{vl}^{benign})0.9920.9880.7080.3480.2770.2740.2260.212
Qwen2.5-VL(Dtextharm,Dtextbenign)(D_{text}^{harm}, D_{text}^{benign})0.4540.2740.2060.2070.2290.2190.2060.207

At token 0, the highest cosine similarities occur between harmful and benign inputs within the vision modality, even though their semantics differ: 0.943 for LLaVA-1.5-7B and 0.992 for Qwen2.5-VL. In contrast, matched semantics across modalities have substantially lower similarity. During generation, the similarities between harmful and benign visual inputs decrease to 0.209 and 0.212 by token 15, whereas cross-modal pairs with matched semantics vary comparatively less after the initial generation steps. This shows that representations become progressively less modality-specific as generation proceeds, while semantic information is comparatively preserved.

AuthorsOpenReview ↗

Response to Reviewer CifH (2/4)

We next examine whether modality and harmfulness remain separable during unsteered generation. At every token index jj, we train a new classifier using only representations from that same index. The modality classifier separates Dvlharm∪DvlbenignD_{vl}^{harm}\cup D_{vl}^{benign} from Dtextharm∪DtextbenignD_{text}^{harm}\cup D_{text}^{benign}, while the semantic classifier separates Dvlharm∪DtextharmD_{vl}^{harm}\cup D_{text}^{harm} from Dvlbenign∪DtextbenignD_{vl}^{benign}\cup D_{text}^{benign}. We use 15% of the data for training and 85% for testing, so this comparison measures modality versus semantic separability at the same token position rather than transferring a fixed classifier across positions. Higher accuracy indicates that the corresponding distinction remains more separable. The complete curves are provided in Appendix A.2, Fig. 7, and selected token positions are reported below.

ModelCategory (accuracy, %)Token 0Token 1Token 2Token 3Token 4Token 5Token 10Token 15
LLaVA-1.5-7BModality classifier100.095.495.498.499.298.493.381.4
LLaVA-1.5-7BSemantic classifier98.195.491.794.482.387.192.288.4
Qwen2.5-VLModality classifier100.096.095.796.897.097.993.076.1
Qwen2.5-VLSemantic classifier98.795.797.993.894.489.390.689.7

The modality classifier is nearly perfect at the input position, even though the examples within a modality contain different semantic labels. Its accuracy then falls from 100.0% to 81.4% for LLaVA-1.5-7B and to 76.1% for Qwen2.5-VL by token 15. Meanwhile, semantic classification remains comparatively stable and becomes more competitive than modality classification at later positions. Together with the cosine results above, this shows that the representation becomes progressively less modality-specific without any base steering. It also complements the original experiment in Section 4.1 and Fig. 2, where a classifier trained only on the final input prompt representations yields increasing F1 on visual representations over generation steps. Therefore, convergence is a naturally occurring phenomenon rather than an effect induced by steering.

AuthorsOpenReview ↗

Response to Reviewer CifH (3/4)

Q2: Under the same sample budget, how do text-only, image--text, and mixed-modality safety signals compare?

Thanks for your constructive feedback. The purpose of the base text-representation steering in CARS is not to claim that visual representations are intrinsically unsafe. Rather, the backbone LLM exhibits much stronger safety alignment in the text representation space. Without early guidance, a harmful visual input can undergo a modality shift while already generating harmful tokens, and later convergence cannot undo those tokens. The original Fig. 2 shows that applying text-derived steering makes the image-representation F1 curve rise faster and reach a higher level. We further test this explanation using the matched datasets under the same sample budget.

For each output position, we train a new semantic classifier and compare three conditions: no steering, text-representation steering, and image-representation steering. All steering directions are extracted from the matched datasets above, using vtextl=RepMeanl(Dtextbenign)−RepMeanl(Dtextharm)v_{text}^{l}=RepMean^{l}(D_{text}^{benign})-RepMean^{l}(D_{text}^{harm}) and vvll=RepMeanl(Dvlbenign)−RepMeanl(Dvlharm)v_{vl}^{l}=RepMean^{l}(D_{vl}^{benign})-RepMean^{l}(D_{vl}^{harm}). The complete curves are provided in Appendix A.3, Fig. 8, and selected positions are reported below.

ModelCategory (accuracy, %)Token 0Token 1Token 2Token 3Token 4Token 5Token 10Token 15
LLaVA-1.5-7BBaseline semantic classifier98.195.491.794.482.387.192.288.4
LLaVA-1.5-7BText-rep steering classifier98.797.997.398.185.890.691.791.1
LLaVA-1.5-7BImage-rep steering classifier99.597.696.595.783.989.590.187.1
Qwen2.5-VLBaseline semantic classifier98.795.797.993.894.489.390.689.7
Qwen2.5-VLText-rep steering classifier98.996.098.194.194.689.591.492.0
Qwen2.5-VLImage-rep steering classifier98.196.298.996.092.588.285.884.1

Text-representation steering improves the average accuracy over positions 0--15 from 90.2% to 92.8% for LLaVA-1.5-7B and from 92.0% to 93.4% for Qwen2.5-VL. Image-representation steering is not consistently better than the baseline and declines at later positions, especially for Qwen2.5-VL. These results explain why CARS uses a lightweight text-derived intervention before the detection point: it makes the visual representation more compatible with the text-trained safety direction early enough to prevent harmful-prefix generation, while still allowing the model to transition to refusal after harmfulness is detected.

We also run the complete image-representation CARS variant under the same setting. Following the CARS$_{\mathrm{avg}}$ ablation, both variants average LogReg\mathtt{LogReg} predictions over output tokens 5--15, so the only changed component is the source modality of the safety direction. The results are reported in Appendix A.3, Table 6.

ModelMethodFigStepMM-SafetyBench
LLaVA-1.5Baseline43.0%64.8%
LLaVA-1.5CARS$_{\mathrm{avg}}$ (Text Rep.)92.7%95.8%
LLaVA-1.5CARS$_{\mathrm{avg}}$ (Image Rep.)64.0%76.7%
Qwen2.5-VLBaseline67.8%76.3%
Qwen2.5-VLCARS$_{\mathrm{avg}}$ (Text Rep.)97.0%93.9%
Qwen2.5-VLCARS$_{\mathrm{avg}}$ (Image Rep.)65.0%87.6%

Image-representation CARS improves over the undefended baseline in three of the four model--benchmark settings, but is 2.8 points below the Qwen2.5-VL baseline on FigStep. It is consistently weaker than text-representation CARS: the gaps are 28.7 and 32.0 points on FigStep for LLaVA-1.5 and Qwen2.5-VL, respectively, and 19.1 and 6.3 points on MM-SafetyBench. This supports our interpretation that the backbone's text safety alignment is the useful transferable signal, while an image-derived direction may retain modality-related features and therefore encode semantic refusal behavior less reliably.

Reviewer_uy2aOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes CARS, an inference-time safety alignment method for Multimodal Large Language Models (MLLMs). Motivated by the observation of delayed refusal, CARS performs harmfulness detection after generating a certain number of tokens and adaptively adjusts representation steering to improve safety while preserving utility.

Critical strengths
  1. Insightful Analysis of Delayed Refusal. The paper provides a systematic analysis of the delayed refusal phenomenon in MLLMs, supported by multiple empirical experiments that investigate its underlying mechanism and the progressive convergence of visual and textual representations during generation.

  2. Comprehensive Experimental Evaluation. The paper provides extensive comparisons with representative defense baselines across multiple safety benchmarks, along with ablation studies examining key design choices. These experiments offer convincing evidence of the effectiveness of the proposed method.

Critical weaknesses
  1. Insufficient Mechanistic Explanation of Representation Convergence. While the experiments demonstrate that visual and textual representations progressively converge during generation, the underlying mechanism remains unclear. This convergence may simply result from the accumulation of generated text tokens, which increasingly encode the visual content, including harmful information. The current experiments do not distinguish this effect from genuine cross-modal representation alignment. Controlled experiments with identical generated prefixes (e.g., teacher forcing) would help clarify the mechanism.

  2. Limited Generalization to Recent MLLMs. The evaluation is limited to relatively older MLLMs (LLaVA-1.5, LLaVA-Next-8B, and Qwen2.5-VL). Testing newer models such as Qwen3-VL and InternVL3.5 would help establish whether the observed representation convergence and the effectiveness of CARS generalize to more recent architectures.

  3. Potential Issues with Prefix Truncation. When harmful content is detected, CARS discards the previously generated tokens from the final output but retains them in the generation context. It is unclear why the model continues generating from potentially harmful context rather than restarting generation with enhanced steering. Moreover, removing the initial tokens may disrupt the semantic coherence of the final response, as subsequent tokens are generated conditioned on the discarded prefix. The authors should investigate whether this strategy produces incomplete and incoherent responses.

Questions for authors

All questions and concerns for the authors have been included in the Weaknesses section.

6 comments
AuthorsOpenReview ↗

Response to Reviewer uy2a (1/2)

Q1: Is the observed convergence genuine cross-modal representation alignment rather than an effect of accumulated generated text?

Thanks for your constructive feedback. To distinguish representation convergence from an effect specific to the original text safety classifier, we add a controlled analysis using matched text-only and image--text inputs with identical semantics following the basic idea of FigStep. We sample 128 harmful prompts from AdvBench and 128 benign prompts from Alpaca. For every prompt, we create one text-only example and one (image,text) pair by rendering the prompt itself as text in an image. All image examples use the same instruction. This produces four datasets, DvlharmD_{vl}^{\mathrm{harm}}, DtextharmD_{text}^{\mathrm{harm}}, DvlbenignD_{vl}^{\mathrm{benign}}, and DtextbenignD_{text}^{\mathrm{benign}}, so comparisons across modalities keep the semantics fixed, while comparisons across harmfulness labels keep the modality fixed. Examples are provided in Appendix A.1, Fig. 5. Note that, following the FigStep design, matched harmful visual and text inputs usually produce different response behaviors: DvlharmD_{vl}^{\mathrm{harm}} tends to elicit harmful generation, whereas DtextharmD_{text}^{\mathrm{harm}} tends to elicit refusal. The observed convergence therefore cannot be explained merely by the two modalities accumulating the same generated text prefix, because their generated prefixes and response behaviors are typically different.

We first analyze representational proximity without applying any steering on two architecturally different models, LLaVA-1.5-7B and Qwen2.5-VL. At each output position, we compute cosine similarity across layers for four dataset pairings. The two within-modality pairings compare harmful and benign semantics, whereas the two cross-modal pairings compare matched semantics. Higher cosine similarity indicates greater representational proximity. The complete curves are provided in Appendix A.2, Fig. 6, and selected token positions are reported below.

ModelPairToken 0Token 1Token 2Token 3Token 4Token 5Token 10Token 15
LLaVA-1.5-7B(Dvlharm,Dtextharm)(D_{vl}^{harm}, D_{text}^{harm})0.4640.2640.2440.2110.1790.2290.2140.267
LLaVA-1.5-7B(Dvlbenign,Dtextbenign)(D_{vl}^{benign}, D_{text}^{benign})0.4740.3400.3070.3290.2140.2380.2590.246
LLaVA-1.5-7B(Dvlharm,Dvlbenign)(D_{vl}^{harm}, D_{vl}^{benign})0.9430.9800.9510.6790.3410.3420.2510.209
LLaVA-1.5-7B(Dtextharm,Dtextbenign)(D_{text}^{harm}, D_{text}^{benign})0.3860.2320.2000.1510.1570.1510.1600.152
Qwen2.5-VL(Dvlharm,Dtextharm)(D_{vl}^{harm}, D_{text}^{harm})0.4380.2470.1920.1520.2140.2440.2700.326
Qwen2.5-VL(Dvlbenign,Dtextbenign)(D_{vl}^{benign}, D_{text}^{benign})0.4560.3430.3730.3080.2960.3160.3340.289
Qwen2.5-VL(Dvlharm,Dvlbenign)(D_{vl}^{harm}, D_{vl}^{benign})0.9920.9880.7080.3480.2770.2740.2260.212
Qwen2.5-VL(Dtextharm,Dtextbenign)(D_{text}^{harm}, D_{text}^{benign})0.4540.2740.2060.2070.2290.2190.2060.207

At token 0, the highest cosine similarities occur between harmful and benign inputs within the vision modality, even though their semantics differ: 0.943 for LLaVA-1.5-7B and 0.992 for Qwen2.5-VL. In contrast, matched semantics across modalities have substantially lower similarity. During generation, the similarities between harmful and benign visual inputs decrease to 0.209 and 0.212 by token 15, whereas cross-modal pairs with matched semantics vary comparatively less after the initial generation steps. This shows that representations become progressively less modality-specific as generation proceeds, while semantic information is comparatively preserved.

AuthorsOpenReview ↗

Response to Reviewer uy2a (2/2)

We next examine whether modality and harmfulness remain separable during unsteered generation. At every token index jj, we train a new classifier using only representations from that same index. The modality classifier separates Dvlharm∪DvlbenignD_{vl}^{harm}\cup D_{vl}^{benign} from Dtextharm∪DtextbenignD_{text}^{harm}\cup D_{text}^{benign}, while the semantic classifier separates Dvlharm∪DtextharmD_{vl}^{harm}\cup D_{text}^{harm} from Dvlbenign∪DtextbenignD_{vl}^{benign}\cup D_{text}^{benign}. We use 15% of the data for training and 85% for testing, so this comparison measures modality versus semantic separability at the same token position rather than transferring a fixed classifier across positions. Higher accuracy indicates that the corresponding distinction remains more separable. The complete curves are provided in Appendix A.2, Fig. 7, and selected token positions are reported below.

ModelCategory (accuracy, %)Token 0Token 1Token 2Token 3Token 4Token 5Token 10Token 15
LLaVA-1.5-7BModality classifier100.095.495.498.499.298.493.381.4
LLaVA-1.5-7BSemantic classifier98.195.491.794.482.387.192.288.4
Qwen2.5-VLModality classifier100.096.095.796.897.097.993.076.1
Qwen2.5-VLSemantic classifier98.795.797.993.894.489.390.689.7

The modality classifier is nearly perfect at the input position, even though the examples within a modality contain different semantic labels. Its accuracy then falls from 100.0% to 81.4% for LLaVA-1.5-7B and to 76.1% for Qwen2.5-VL by token 15. Meanwhile, semantic classification remains comparatively stable and becomes more competitive than modality classification at later positions. Together with the cosine results above, this shows that the representation becomes progressively less modality-specific without any base steering. It also complements the original experiment in Section 4.1 and Fig. 2, where a classifier trained only on the final input prompt representations yields increasing F1 on visual representations over generation steps. Therefore, convergence is a naturally occurring phenomenon rather than an effect induced by steering.

Q2: Does CARS generalize to more recent MLLMs?

Thanks for your constructive feedback. We evaluate CARS on two more recent and architecturally different models, Qwen3-VL-8B and InternVL3.5-8B, using FigStep and MM-SafetyBench for safety and ScienceQA for utility. Results are shown below:

ModelMethodFigStep DSRMM-SafetyBench DSRScienceQA
Qwen3-VL-8BBaseline DSR64.6%71.5%89.7%
Qwen3-VL-8BCARS95.4%97.6%89.2%
InternVL3.5-8BBaseline DSR46.0%85.3%89.5%
InternVL3.5-8BCARS98.4%99.4%88.4%

On Qwen3-VL-8B, CARS improves DSR from 64.6% to 95.4% on FigStep and from 71.5% to 97.6% on MM-SafetyBench, while ScienceQA changes from 89.7% to 89.2%. On InternVL3.5-8B, CARS improves DSR from 46.0% to 98.4% on FigStep and from 85.3% to 99.4% on MM-SafetyBench, while ScienceQA changes from 89.5% to 88.4%. Full results are provided in Appendix B. These results support the transferability of CARS beyond the original LLaVA and Qwen2.5-VL evaluation while preserving most of the utility.

Q3: Why does CARS retain the generated prefix in the context after cutoff, and does cutoff harm response coherence or utility?

Thanks for your constructive feedback. CARS discards the pre-detection tokens from the returned response but retains them in the autoregressive context so that generation can continue from the semantic state already established by the model. This design preserves autoregressive continuity: restarting would discard the accumulated visual and semantic state, repeat the prefix computation, and require the model to reconstruct the same context. Once harmfulness is detected, CARS instead increases the steering strength to αenhanced\alpha_{\mathrm{enhanced}}, allowing the existing trajectory to transition toward refusal, while cutoff ensures that potentially harmful pre-detection tokens are not exposed to the user.

Cutoff is activated only when the classifier predicts a harmful input at jdetectj_{\mathrm{detect}} or when the token-averaged prediction exceeds the threshold. Otherwise, enhanced steering and cutoff are never applied. To isolate whether cutoff harms benign responses, we measure its activation rate on ScienceQA and MM-Vet. It is never activated on ScienceQA and is activated for fewer than 2% of MM-Vet inputs: 1.84% for LLaVA-1.5-7B and 0.75% for Qwen2.5-VL-7B (Appendix C.2). Together with the near-baseline utility in Figure 3, this shows that cutoff does not affect benign response quality. The small remaining utility trade-off is primarily attributable to the lightweight base steering applied before detection, rather than prefix deletion or enhanced steering.

AuthorsOpenReview ↗

Further experiment on controlled prefixes

Q: Could the observed separability trends come from the generated prefix rather than cross-modal convergence?

Thank you for raising this concern. To test whether the observed trends depend on harmful content generated from visual inputs, we apply teacher forcing only to image–text inputs using the same fixed, harmless sentence for all harmful and benign examples. Text-only inputs retain the original free-generation setting. Details are provided in Appendix A.4.

Let me carefully consider the information provided before giving a clear and detailed response.

This sentence contains 15 tokens for both models under our experimental tokenization. Let s=(s1,…,s15)\mathbf{s}=(s_1,\ldots,s_{15}) denote this fixed sequence. For each image–text input xvlx_{vl}, we apply teacher forcing by setting the next token to sjs_j and extract the last-token representation:

h_j^\ell(x_{vl})=R_j^\ell(x_{vl},s_{1:j}),\qquad j=1,\ldots,15.$$ Token 0 denotes the last-token representation of the original input before adding the prefix. The prefix is identical across harmful and benign image–text inputs, preventing its wording from revealing their harmfulness labels. Following our previous protocol, we train a new modality or semantic classifier at each position using representations from both modalities and a 15%/85% train/test split grouped by the underlying prompt. These classifiers measure representation separability and are distinct from the single text-trained CARS detector. We first compare free generation and teacher forcing without steering. | Model | Category (accuracy, %) | Token 0 | Token 1 | Token 2 | Token 3 | Token 4 | Token 5 | Token 10 | Token 15 | |---|---|---:|---:|---:|---:|---:|---:|---:|---:| | LLaVA-1.5-7B | Modality, free generation | 100.0 | 95.4 | 95.4 | 98.4 | 99.2 | 98.4 | 93.3 | 81.4 | | LLaVA-1.5-7B | Modality, teacher forcing | 100.0 | 94.8 | 95.8 | 97.6 | 96.5 | 95.7 | 93.7 | 80.6 | | LLaVA-1.5-7B | Semantic, free generation | 98.1 | 95.4 | 91.7 | 94.4 | 82.3 | 87.1 | 92.2 | 88.4 | | LLaVA-1.5-7B | Semantic, teacher forcing | 98.1 | 91.6 | 90.2 | 89.2 | 87.1 | 88.0 | 89.2 | 90.8 | | Qwen2.5-VL | Modality, free generation | 100.0 | 96.0 | 95.7 | 96.8 | 97.0 | 97.9 | 93.0 | 76.1 | | Qwen2.5-VL | Modality, teacher forcing | 100.0 | 94.8 | 93.9 | 95.9 | 95.6 | 95.9 | 94.8 | 76.1 | | Qwen2.5-VL | Semantic, free generation | 98.7 | 95.7 | 97.9 | 93.8 | 94.4 | 89.3 | 90.6 | 89.7 | | Qwen2.5-VL | Semantic, teacher forcing | 98.7 | 94.5 | 96.1 | 93.0 | 92.9 | 88.2 | 93.3 | 90.5 | The same overall trends persist when the visual-input prefix is controlled. By token 15, modality accuracy falls from 100.0% to 80.6% for LLaVA-1.5-7B and to 76.1% for Qwen2.5-VL, close to 81.4% and 76.1% under free generation. Semantic accuracy remains high, reaching 90.8% and 90.5%, compared with 88.4% and 89.7%. Thus, modality separability decreases while substantial semantic separability remains across the combined datasets, even when image–text inputs are followed by an identical harmless prefix. This is consistent with our observation that most Qwen2.5-VL responses begin with the generic opening "Certainly! Here is the detailed content for each item on the list", which already occupies 13 tokens under our experimental tokenization. We next apply text-derived steering and compare semantic accuracy across both modalities, with teacher forcing again applied only to image–text inputs. | Model | Condition (accuracy, %) | Token 0 | Token 1 | Token 2 | Token 3 | Token 4 | Token 5 | Token 10 | Token 15 | |---|---|---:|---:|---:|---:|---:|---:|---:|---:| | LLaVA-1.5-7B | Text steering, free generation | 98.7 | 97.9 | 97.3 | 98.1 | 85.8 | 90.6 | 91.7 | 91.1 | | LLaVA-1.5-7B | Text steering, teacher forcing | 98.7 | 94.1 | 93.4 | 95.3 | 87.2 | 88.8 | 90.5 | 89.4 | | Qwen2.5-VL | Text steering, free generation | 98.9 | 96.0 | 98.1 | 94.1 | 94.6 | 89.5 | 91.4 | 92.0 | | Qwen2.5-VL | Text steering, teacher forcing | 98.9 | 95.9 | 97.2 | 93.9 | 93.8 | 89.2 | 90.4 | 91.8 | With text steering, semantic accuracy at token 15 is 89.4% under teacher forcing versus 91.1% under free generation for LLaVA-1.5-7B, and 91.8% versus 92.0% for Qwen2.5-VL. Thus, late-position semantic separability remains comparable when the visual-input prefix is fixed to harmless text. Together, these results are consistent with natural representation convergence and show that the observed trends persist without harmful text accumulating in the visual-input prefix. They complement our earlier detection and defense results in motivating delayed detection.
Reviewer_5y9HOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces an inference-time defense for VLMs, motivated by the delayed refusal: text-derived safety steering can produce harmful content before eventually triggering refusal on visual inputs. CARS initially applies lightweight steering, then uses a text-trained logistic classifier at a predetermined generation step to assess harmfulness. For flagged inputs, it strengthens steering and removes the preceding output tokens. The paper evaluates three models across four safety benchmarks.

Critical strengths
  1. Delayed refusal is a useful observation. Table 1 shows a substantial discrepancy between refusal-keyword detection and response safety. This finding can motivate more research into the response safety detection.

  2. CARS leaves the backbone weights unchanged and requires only direction extraction, lightweight classifier fitting, and calibration. The reported 7.3-8.7% inference overhead is practical.

  3. The experiments covers various models. On LLaVA-1.5, LLaVA-Nextm and Qwen2.5-VL, CARS achieves consistent improvement.

Critical weaknesses
  1. The evidence does not directly establish convergence as the cause of delayed refusal. The PCA overlap and classifier F1 may only show the refusal behavior, rather than the overall safety of responses.

  2. The ablations do not isolate the contribution of the proposed mechanism. CARS combines early steering, delayed classification, stronger subsequent steering, and prefix deletion, but the experiments mainly vary strength and detection position.

Questions for authors

Please refer to the questions.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 5y9H (1/2)

Q1: Does the evidence establish representation convergence as the cause of delayed refusal?

Thanks for your constructive feedback. We agree that the original PCA plots and classifier F1 scores alone did not provide enough controlled evidence for representation convergence. To address this concern, we construct matched datasets following the basic idea of FigStep. We sample 128 harmful prompts from AdvBench and 128 benign prompts from Alpaca. For every prompt, we create one text-only example and one (image,text) pair by rendering the prompt itself as text in an image. All image examples use the same instruction. This produces four datasets, DvlharmD_{vl}^{\mathrm{harm}}, DtextharmD_{text}^{\mathrm{harm}}, DvlbenignD_{vl}^{\mathrm{benign}}, and DtextbenignD_{text}^{\mathrm{benign}}, so comparisons across modalities keep the semantics fixed, while comparisons across harmfulness labels keep the modality fixed. Examples are provided in Appendix A.1, Fig. 5.

We first analyze representational proximity without applying any steering on two architecturally different models, LLaVA-1.5-7B and Qwen2.5-VL. At each output position, we compute cosine similarity across layers for four dataset pairings. The two within-modality pairings compare harmful and benign semantics, whereas the two cross-modal pairings compare matched semantics. Higher cosine similarity indicates greater representational proximity. The complete curves are provided in Appendix A.2, Fig. 6, and selected token positions are reported below.

ModelPairToken 0Token 1Token 2Token 3Token 4Token 5Token 10Token 15
LLaVA-1.5-7B(Dvlharm,Dtextharm)(D_{vl}^{harm}, D_{text}^{harm})0.4640.2640.2440.2110.1790.2290.2140.267
LLaVA-1.5-7B(Dvlbenign,Dtextbenign)(D_{vl}^{benign}, D_{text}^{benign})0.4740.3400.3070.3290.2140.2380.2590.246
LLaVA-1.5-7B(Dvlharm,Dvlbenign)(D_{vl}^{harm}, D_{vl}^{benign})0.9430.9800.9510.6790.3410.3420.2510.209
LLaVA-1.5-7B(Dtextharm,Dtextbenign)(D_{text}^{harm}, D_{text}^{benign})0.3860.2320.2000.1510.1570.1510.1600.152
Qwen2.5-VL(Dvlharm,Dtextharm)(D_{vl}^{harm}, D_{text}^{harm})0.4380.2470.1920.1520.2140.2440.2700.326
Qwen2.5-VL(Dvlbenign,Dtextbenign)(D_{vl}^{benign}, D_{text}^{benign})0.4560.3430.3730.3080.2960.3160.3340.289
Qwen2.5-VL(Dvlharm,Dvlbenign)(D_{vl}^{harm}, D_{vl}^{benign})0.9920.9880.7080.3480.2770.2740.2260.212
Qwen2.5-VL(Dtextharm,Dtextbenign)(D_{text}^{harm}, D_{text}^{benign})0.4540.2740.2060.2070.2290.2190.2060.207

At token 0, the highest cosine similarities occur between harmful and benign inputs within the vision modality, even though their semantics differ: 0.943 for LLaVA-1.5-7B and 0.992 for Qwen2.5-VL. In contrast, matched semantics across modalities have substantially lower similarity. During generation, the similarities between harmful and benign visual inputs decrease to 0.209 and 0.212 by token 15, whereas cross-modal pairs with matched semantics vary comparatively less after the initial generation steps. This shows that representations become progressively less modality-specific as generation proceeds, while semantic information is comparatively preserved.

We next examine whether modality and harmfulness remain separable during unsteered generation. At every token index jj, we train a new classifier using only representations from that same index. The modality classifier separates Dvlharm∪DvlbenignD_{vl}^{harm}\cup D_{vl}^{benign} from Dtextharm∪DtextbenignD_{text}^{harm}\cup D_{text}^{benign}, while the semantic classifier separates Dvlharm∪DtextharmD_{vl}^{harm}\cup D_{text}^{harm} from Dvlbenign∪DtextbenignD_{vl}^{benign}\cup D_{text}^{benign}. We use 15% of the data for training and 85% for testing, so this comparison measures modality versus semantic separability at the same token position rather than transferring a fixed classifier across positions. Higher accuracy indicates that the corresponding distinction remains more separable. The complete curves are provided in Appendix A.2, Fig. 7, and selected token positions are reported below.

ModelCategory (accuracy, %)Token 0Token 1Token 2Token 3Token 4Token 5Token 10Token 15
LLaVA-1.5-7BModality classifier100.095.495.498.499.298.493.381.4
LLaVA-1.5-7BSemantic classifier98.195.491.794.482.387.192.288.4
Qwen2.5-VLModality classifier100.096.095.796.897.097.993.076.1
Qwen2.5-VLSemantic classifier98.795.797.993.894.489.390.689.7
AuthorsOpenReview ↗

Response to Reviewer 5y9H (2/2)

The modality classifier is nearly perfect at the input position, even though the examples within a modality contain different semantic labels. Its accuracy then falls from 100.0% to 81.4% for LLaVA-1.5-7B and to 76.1% for Qwen2.5-VL by token 15. Meanwhile, semantic classification remains comparatively stable and becomes more competitive than modality classification at later positions. Together with the cosine results above, this shows that the representation becomes progressively less modality-specific without any base steering. It also complements the original experiment in Section 4.1 and Fig. 2, where a classifier trained only on the final input prompt representations yields increasing F1 on visual representations over generation steps. Therefore, convergence is a naturally occurring phenomenon rather than an effect induced by steering.

Q2: Do the ablations isolate the contributions of CARS's detection and intervention choices?

Thanks for your constructive feedback. In the original submission, Figure 4 already ablates jdetectj_{\mathrm{detect}} under different steering strengths, and CARS$_{\mathrm{avg}}$ averages classifier predictions over tokens 5--15 to verify that the method does not depend on a single selected position. In this rebuttal, we further add one-variable-at-a-time experiments on the base steering strength, enhanced steering strength, and 12-layer intervention range. Full settings and results are reported in Appendix C.3.

ModelSettingScienceQAMM-SafetyBench DSRFigStep DSR
LLaVA-1.5-7BDefault CARS$_{\mathrm{avg}}$61.1%95.8%92.7%
Base α=1.0\alpha=1.061.5%80.6%34.5%
Base α=6.0\alpha=6.055.9%99.6%99.5%
Enhanced α=3.5\alpha=3.561.1%82.8%65.0%
Enhanced α=12.0\alpha=12.061.1%95.0%95.3%
Layers 0--110.0%62.6%30.0%
Layers 20--3161.7%80.3%32.0%
Qwen2.5-VL-7BDefault CARS$_{\mathrm{avg}}$85.6%93.9%97.0%
Base α=1.0\alpha=1.086.7%93.2%71.5%
Base α=6.0\alpha=6.075.8%100.0%100.0%
Enhanced α=3.5\alpha=3.585.6%91.6%92.8%
Enhanced α=12.0\alpha=12.085.2%94.6%98.8%
Layers 0--110.0%63.2%29.8%
Layers 16--2784.5%98.8%90.0%

The intervention layers are the most sensitive choice: moving steering to early layers collapses both safety and utility, while other shifted ranges also substantially reduce at least one safety metric. Base strength exhibits the expected safety--utility trade-off: increasing it improves DSR but lowers ScienceQA, whereas decreasing it preserves utility but weakens defense, particularly on FigStep. Enhanced strength has much less effect on utility because it is applied only after a harmful prediction and increasing it to 12.0 maintains ScienceQA and slightly improves DSR relative to the default.

Thank the authors for the rebuttal. The added experiments addressed my concerns. I will maintain the positive score and increase my confidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.