acceptodds
Under review as a conference paper at ICLR 2027

Linear Separability Is Not Input Controllability: Intermediate-Layer Jailbreak on VLMs

Abstract

Linear probes reveal safety-related information in Vision-Language Models (VLMs), but linear safety separability does not necessarily translate into input-space controllability. Across three VLM families, empirical layer sweeps find that attacks targeting intermediate layers are more effective than those targeting the final layer, even when late-layer safety probes remain highly accurate. Our local analysis distinguishes how much image perturbations can move representations along a safety direction from how that movement affects downstream behavior. It shows that perfect linear separability can coexist with arbitrarily weak input reachability. We introduce Intermediate-Layer Boundary Targeting (IBT), which combines intermediate-layer image optimization with text conditioning. IBT optimizes bounded image perturbations to move representations toward targets on the side classified as safe by a linear probe. On MM-SafetyBench, IBT outperforms final-layer targeting on average across the three source models in attack success rate (ASR), stealth attack success rate (SASR), and semantic response scores. In black-box transfer from an open-weight source model to three closed-source VLMs, IBT outperforms all compared baselines in mean ASR. These findings establish that linear safety separability does not guarantee input-space controllability and identify the supervision layer as a practical design variable for representation-guided image attacks, with intermediate-layer supervision yielding stronger VLM jailbreaks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
VLM jailbreaks, Safety representations, Linear separability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_61oGOpenReview ↗
Rating
2(+1)
Confidence
3(−1)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines whether safety information in VLM hidden states is linearly separable, and argues that separability alone does not mean those directions can be driven from the image. The evidence comes from layer sweeps on three open-weight models. Probe accuracy stays near ceiling in the late layers, yet attack success peaks much earlier, at a normalized depth of roughly 0.4–0.5. Building on this gap, the authors propose an attack called IBT. It first fits a linear safety probe at a given layer. The clean representation is then reflected across the probe's decision boundary, and the reflected point becomes the optimization target. A bounded image perturbation is optimized toward this target with an alignment loss, supplemented by a cosine term and a pixel penalty. The perturbed image is then paired with a benign-looking rewrite of the prompt, which GPT-4 generates and which is cached in advance.

Critical strengths

The paper reports a clear and practically useful empirical finding. The most linearly separable safety representation is not necessarily the best supervision target for input-space attacks. This holds across three quite different VLM families. Both the controlled layer sweep and the full IBT evaluation point to the same fairly narrow intermediate region, around normalized depth 0.4–0.5, even though late-layer probes remain highly accurate. What matters more is that the pattern does not rest on a single cherry-picked layer. A shared normalized position near 0.45 keeps almost all of each model's peak performance, with mean ASR and SASR coming very close to the independently selected optima. The finding is useful because it turns supervision-layer selection into an explicit design choice for representation-guided attacks, instead of letting the final layer, or the most separable one, serve as the default.

Critical weaknesses

The experiments show a specific phenomenon that supervision around depth 0.45 is substantially better than supervision at later layers, but the theory provides no reason for such an intermediate optimum to exist. In fact, the appendix explicitly shows that the full first-order input-to-behavior gradient is invariant to the location of the computational cut. The proposed decomposition can explain why probe accuracy alone is insufficient, but it cannot explain why L12 should outperform L27, or why three models should peak around the same normalized depth. Thus, the central theoretical story stops exactly where the paper’s main empirical claim begins.

The second concern is that the reported gains are not cleanly attributable to IBT itself. The attack uses EOT with JPEG compression, random resizing, and average pooling during optimization, but this component is never removed or separately evaluated. This is a major confound because exactly this kind of transformation-robust optimization is known to help adversarial examples survive preprocessing, transfer across models, and remain visually stable. Yet the paper attributes the strong black-box transfer and stealth results mainly to intermediate-layer targeting. The component study does not resolve this issue because EOT remains bundled inside the image optimization term.

Questions for authors

Qwen is swept only from L11 and Llama/GLM from L14, while all reported optima occur very early in the evaluated range. Can you extend the sweep substantially below d=.35?

In Table 6, why is image-only SASR below the untouched baseline, and why does the implied NIQE weight at Qwen L12 differ so much from IBT at the same layer and budget?

1 comment

After looking at the paper again, I misread how SASR is computed in my second question, and I will raise my score from 1 to 2 for this paper's strength part as I wrote with the author's clarification on my two questions. The remaining weaknesses still need to be addressed.

Reviewer_2KpdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether linearly separable safety representations are also easy to control from the input space. The authors find that, across three evaluated VLMs, image attacks guided by intermediate-layer representations are more effective than those targeting the final layer, even though safety information remains highly separable in later layers. Based on this observation, the paper proposes Intermediate-Layer Boundary Targeting (IBT), which combines intermediate-layer representation targeting with text rewriting. The paper also introduces input reachability and downstream influence to explain why probe accuracy alone is not sufficient for selecting an attack target.

Critical strengths
  1. The paper identifies a useful gap between representation readability and controllability.

  2. Under the experimental setup of this paper, the proposed attack shows a significant improvement in effectiveness.

Critical weaknesses
  • The paper shows that IBT examples optimized at Qwen L12 transfer well to GPT-4o, Gemini, and Claude, which is a nice result. However, this experiment does not actually show that intermediate-layer supervision itself leads to better transferability. There is no matched transfer comparison between, for example, IBT optimized at Qwen L12 and IBT optimized at the final layer under the same setup. More importantly, I do not see a clear explanation of why an intermediate-layer direction learned on one model should transfer better across models.

  • The main conclusion is based on only three open-weight checkpoints: Qwen2.5-VL-7B, Llama-3.2-11B-Vision, and GLM-4.1V-9B-Thinking. In other words, each model family is represented by only one model size, and all three models are in a fairly similar parameter range. Would the same pattern hold for a much larger model from the same family? Would the optimal normalized layer stay around 0.45 as model size increases? It is also unclear whether the effect survives across more diverse VLM architectures or training recipes.

  • The paper introduces input reachability and downstream influence to explain why probe accuracy is insufficient, but I did not find a layer-wise empirical analysis showing that these quantities actually predict the observed ASR/SASR peak. As a result, the theory explains why separability and attackability can differ in principle, but it does not yet explain why the intermediate layers are the best attack targets in the evaluated models.

  • Since both the probes and the main attack evaluation are built around the same benchmark, it is unclear whether the same layer-wise pattern would persist for different distributions of harmful requests or multimodal safety tasks.

Questions for authors
  • For the closed-source transfer experiment, could you report a matched comparison between adversarial inputs optimized at the intermediate layer and at the final layer of the same source model? For example, Qwen L12 vs. Qwen L27 with the same text and optimization settings. This would help clarify whether intermediate-layer supervision is actually responsible for the stronger transfer.

  • Do you have any evidence for why intermediate-layer supervision should transfer better across different VLMs?

  • How stable is the d≈0.45d \approx 0.45 observation with model scale? In particular, have you tested multiple parameter sizes within the same model family?

Reviewer_YaWVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper argues that linear safety separability in VLM representations does not imply input-space controllability: a probe can read off safety information from hidden states, yet an input image perturbation may be unable to push the representation across that boundary through the model's own computation. Empirically, across three VLM families, attack success peaks at normalized layer position ~0.4–0.5 even though late-layer safety probes remain highly accurate. The paper gives a first-order account and a constructive example where perfect separability coexists with vanishing reachability. It proposes IBT which fit frozen per-layer probes, construct a boundary-reflected target at an intermediate layer, and optimize a bounded image perturbation combined with a cached text rewrite.

Critical strengths

The conceptual distinction between linear separability and input controllability is clean, well-motivated, and corrects a common implicit inference in probe-based safety work; it is paired with a concrete reachability-vs-separability characterization and a constructive counterexample.

Critical weaknesses
  1. Limited theoretical and methodological novelty. The attack reuses the JailBound-style per-layer-probe framework, so the net contribution reduces to intermediate-layer targeting plus a text rewrite; the core result is the empirical layer-position observation rather than a fundamentally new mechanism. Why we can select such layers? I hope to see the deep analysis/true reasons.

  2. Evaluation leaves soundness gaps. Results rest largely on a single benchmark (MM-SafetyBench); ASR is measured by refusal-string absence, which is permissive and likely inflates success; gpt-4o both supplies rewrite context and serves as the semantic judge.

Questions for authors
  1. Does the ~0.4–0.5 ASR peak persist on benchmarks beyond MM-SafetyBench (e.g., a second multimodal safety set), and how sensitive is the optimal layer to the benchmark and the probe's training labels?
  2. How much does the refusal-string ASR overestimate true unsafe compliance relative to a stricter human or independent-judge evaluation, and does using a judge model other than gpt-4o change the semantic-score ordering?
  3. Beyond the first-order account, do ν/μ/ρ* predict the empirically optimal layer quantitatively (e.g., correlation between measured reachability and per-layer ASR), or are they primarily illustrative?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.