acceptodds
Under review as a conference paper at ICLR 2027

Mind the Spike: Mechanisms and Brittleness of Visual Massive Activations in Large Vision–Language Models

Abstract

Large vision–language models (LVLMs) inherit massive activations from their text-only bases: spikes where a few fixed hidden channels receive values thousands of times above the typical magnitude. The text spike systematically appears in early layers at a fixed initial position, independently of input content. Visual spikes vary across images, but whether their formation follows a consistent pattern across LVLMs and how they respond to image perturbations remain open questions. We find that some LVLMs do not form visual spikes, while others spike at different rates, typically in deeper layers. We identify the trigger direction from model weights and an interpretable location rule: before the language model decoder runs, eventual spike tokens are largely restricted to those sharing least with the rest of the image. Crucially, visual spikes are strikingly brittle. Common corruptions frequently create and relocate spikes, and less often remove them, raising overall incidence. Our trigger-guided spike attack deliberately creates or removes spikes under a small budget, with enough in nine of the ten models that spike. Finally, our preventive intervention removes only the trigger component before spikes erupt, eliminating or substantially reducing spikes on clean and perturbed images while leaving the other image tokens nearly unchanged. Our study spans 25 adapter-based LVLMs built on 18 released text-only bases from 10 families, ranging from 2B to 72B parameters.

Primary area:
interpretability and explainable AI
Keywords:
massive activations, attention sinks, large vision-language models, mechanistic interpretability, internal representations, visual tokens, adversarial perturbations, inference-time intervention

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_MjevOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies decoder-written visual massive activations in 25 adapter-based vision-language models. Ten exhibit spikes under the measured criterion. The authors identify inherited blocks and trigger directions associated with spike formation, relate eventual spike locations to the least-shared image tokens at decoder input, and demonstrate sensitivity to corruptions and small adversarial perturbations. Removing the trigger component suppresses spikes, with limited task-score changes in most models. The paper leaves their functional role unresolved.

Critical strengths

The distinction between channel concentration, massive activation and attention sinks is well supported and matters for interpreting earlier measurements. The breadth of the model survey is valuable, as are the comparisons with the corresponding text-only bases. Trigger interventions, state swaps and random-direction controls provide substantially more mechanistic evidence than a correlation study. The authors also report exceptions and attention-implementation sensitivity rather than presenting a uniform account unsupported by the measurements.

Critical weaknesses
  1. The experiments do not yet establish the value of the learned controller over simpler adaptation. The paper does tune a global positive coefficient on validation data, but that does not replace a comparison with a simple gate using one tuned signed coefficient. Such a control would distinguish the value of selecting when to intervene from that of learning a 40-bin magnitude distribution. The headline 24-point improvement on LLaVA VOC is also achieved by VCD: both reach CHAIRi 14.0. It is a valid comparison with Regular decoding, but does not establish the added value of TAVCD.

  2. The deployment features need clarification. Table 2 defines t/T using the total caption length, which is unavailable before generation terminates. If inference substitutes a maximum generation length or an estimate, the paper should specify it and use the same definition when training the predictor. As written, the online feature computation cannot be reproduced unambiguously.

  3. The oracle labels provide a weaker guarantee than the term “structurally uncorrectable” suggests. Appendix B searches a bounded, discrete grid on the stored top-50 support; failure there does not establish that the full real-valued feasible set is empty. A C3 violation excludes positive coefficients, not necessarily negative ones.

  4. The recall tradeoff is substantial for BLIP-2 on VOC: 61.6 with VCD falls to 49.0 with TAVCD. CHAIRi improves from 16.4 to 16.2 and CHAIRs from 18.0 to 14.2, but these reported gains accompany considerably less object coverage. Object-mention counts and caption lengths would help characterize this tradeoff.

  5. The reported “less than 1%” overhead concerns the head, not the two-branch decoder. End-to-end latency against Regular decoding and VCD is still missing.

Questions for authors

Where are the residual spikes written in InternVL3.5-8B after trigger removal, and do they involve the same tokens and blocks as before intervention?

2 comments

Thank you for reviewing our paper. We appreciate your recognition of the breadth of our study, the mechanistic evidence provided by our experiments, and our candid reporting of model-specific exceptions. We address your question below.

Residual spikes after trigger removal

We appreciate the reviewer’s thoughtful question, which led us to investigate the residual spikes further.

In InternVL3.5-8B, every residual spike after trigger removal is written at the output of the visual switch block, the same block that writes the clean spike, and in all 66 forward passes with residual spikes, the spike occurs at the same image token as in the clean forward. We extended the analysis to the four models where the trigger intervention leaves residual spikes. The same token spikes in 6 of 6 residual cases in InternVL3.5-4B, 25 of 31 in InternVL3.5-14B, and 86 of 91 in Molmo2-8B, and every residual spike peaks at the switch's output. We found that the residual spikes arise from the other dominant eigendirection of the switch, the one of the opposite sign, which becomes prominent once the trigger direction is projected out. Projecting out this second direction as well eliminates them completely.

Based on this finding, we have updated our intervention to project out both the trigger tt and t−t_-, obtained by running the trigger search with the write sign reversed. In these four models, tt matches v+v_+ and t−t_- matches v−v_-, the eigenvectors associated with the largest positive and most negative eigenvalues of the switch's quadratic form, respectively. Removing both directions eliminates all residual spikes, without changing our overall results and conclusions (Table 1; ablation in Table 50).

Table: Origin of residual spikes and effectiveness of the updated intervention. We re-evaluate the spiking POPE forwards from Table 1, first projecting out tt at the visual switch, then projecting out both tt and t−t_-. Every residual spike peaks at the output of the original visual switch block. Values leaving the switch are medians over forwards with residual spikes; the cosine similarity between tt and v+v_+ is at least 0.999.

ModelSpiking POPE forwardsStill spiking, tt removedOn the clean spike's tokenValue leaving switch, clean / tt removedλ(v+)\lambda(v_+) / λ(v−)\lambda(v_-)Still spiking, tt and t−t_- removedAnswers unchanged, tt and t−t_- removed
InternVL3.5-4B3266 of 6+6,552 / −632+14.6 / −33.6032 of 32
InternVL3.5-8B1196666 of 66+14,154 / −1,108+17.1 / −33.50119 of 119
InternVL3.5-14B1223125 of 31+10,040 / −976+30.9 / −37.80121 of 122
Molmo2-8B2009186 of 91+14,423 / −756+21.6 / −40.80200 of 200

! The critical weaknesses section appears to refer to another paper. Please let us know if anything about our work needs clarification.

Revisions to the manuscript are highlighted in blue.

Dear Reviewer Mjev,

We appreciate your initial positive feedback. We would be happy to provide any further clarification before the discussion period ends.

Thank you for your time and consideration.

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

This paper systematically analyzes the formation, location, and sensitivity to image perturbations of visual spikes in LVLMs. Among the 25 LVLMs evaluated, 10 models spike and and follow an interpretable rule: eventual spike tokens are already largely restricted to those sharing least with the rest of the image. Based on this finding, the authors could identify the trigger direction. Moreover, the authors find that visual spikes are brittle to image perturbations and propose preventive intervention to eliminate or significantly reduce spikes on clean images and perturbed images while leaving the other image tokens nearly unchanged.

Critical strengths
  1. Previous studies only analyze one or two LVLMs and then generalize the conclusions to other models. In contrast, this paper comprehensively evaluates 25 adapter-based LVLMs built on 18 released text-only bases from 10 families, ranging from 2B to 72B parameters, drawing more convincing conclusions.
  2. The paper is logically solid, and the explanation for the visual spike phenomenon has been validated through a lot of experiments. It provides interpretable rules for the location of visual spikes.
Critical weaknesses

The function of visual spikes remains insufficiently explained. Although the paper evaluates task performance after intervention, it remains unclear under what conditions visual spikes help or affect the task performance of models. Further analyzing their impact on specific model capabilities can enhance the significance of the research.

Questions for authors
  1. The experiment is based on the COCO dataset. Could the authors provide further analysis on more datasets?
  2. Does the preventive intervention for spikes have any effect on attention sinks? I am curious whether a similar approach could eliminate or substantially reduce attention sinks.
3 comments

Thank you for reviewing our paper. We appreciate your recognition of the breadth of our evaluation and the interpretability of our findings. We address your questions below.

1. Functional role of visual spikes

We agree that the functional role of visual spikes is an important open question, as our conclusion and the limitations and future work sections already state (Appendices I and J). Prior work, including Kang et al., reports that removing visual sink tokens has little effect on the evaluated tasks. Our findings suggest that this question is far from settled: although removing visual spikes largely preserves task scores and POPE answers, it changes generated captions to varying degrees across models. Changes in generated wording do not, by themselves, establish differences in correctness or functional behaviour. A comprehensive characterization would require a principled study across model capabilities, tasks, and safety-relevant behaviours, which we leave to future work.

2. Generalization beyond COCO

We have added experiments on 300 images each from ImageNet and Open Images (new Appendix G). The results largely confirm our findings on COCO: the same ten models exhibit visual spikes, while the others do not. The trigger alignment, attention sinks, spike relocation, brittleness under perturbations, and effectiveness of our preventive intervention are also consistent across datasets.

3. Effect of the intervention on attention sinks

Image attention sinks come mainly from spikes, as Figure 3 illustrates. The exceptions are three non-spiking models, LLaVA-1.5-13B, Molmo-7B-D and PaliGemma 2, on at most 17% of a task's forwards (Table 7). Once the intervention removes the spike, the attention sink goes with it.

Revisions to the manuscript are highlighted in blue.

The authors successfully addressed my concerns, so I keep my positive score.

Dear reviewer TdxN, thank you for your reply. We thank you for taking the time to review our work and for your positive feedback.

Reviewer_9re2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies when and where visual activation spikes arise in 25 LVLMs. It identifies a trigger direction, shows that small image perturbations can change the spikes, and tests an intervention that suppresses them with limited effects on task scores in most models.

Critical strengths

1.The distinction between channel concentration and unusually large activations is useful and well supported. The threshold comparisons and separate attention measurements make the cross-model survey more convincing than a collection of activation plots. 2.The paper goes beyond observation. The base-model comparisons, state swaps, and trigger removal provide useful evidence about spike formation. Reporting the less successful cases, particularly Molmo2 and Pixtral, also helps establish the limits of the findings.

Critical weaknesses

1.I am not convinced that the interventions establish the full causal account. Tables 31–32 track whether the original spike token still spikes, leaving open whether the spike moves elsewhere. The shared-part result is also stronger as a predictor than as a mechanism: scaling changes both norm and shared part, while the fixed-norm edits give mixed results in Table 33. 2.The added value of the nonlinear trigger estimate remains unclear. It nearly matches the existing quadratic direction (Table 8), but the paper does not compare their attack or intervention results, or ablate the attack objective. Missing code and exact sampling settings make this harder to check, especially given the reported sensitivity to attention implementations.

Questions for authors

1.After blocking attention to the text spike token, does the image contain no spike, or does only the original token stop spiking? Please report image-level incidence. 2.How should the shared-part mechanism explain Qwen3-VL-32B’s opposite response and Pixtral’s lack of response in Table 33? Can relocation be separated from spike creation? 3.Do the nonlinear direction and combined attack objective outperform the quadratic direction and a value-only objective under the same evaluation settings? 4.At matched perturbation and optimization budgets, how does the proposed attack compare with random perturbations and with the spike-value term alone?

5 comments

Thank you for reviewing our paper. We appreciate your recognition of the value of our cross-model analysis, mechanistic experiments, and candid reporting of less successful cases. We address your concerns below.

1. Role of text spikes in visual spike formation and relocation

Thank you for pointing out the missing image-level incidence in Tables 31 and 32. We have added these results (Tables 34–36 in the revision). We agree that suppressing one spike through attention denial does not necessarily eliminate image-level spiking. However, the purpose of this intervention is to establish a mechanistic dependence of spike formation at a specific image token on attention to text spike tokens. Denying a spiking image token attention to the text spike tokens suppresses its spike on 75–100% of images, regardless of whether another token spikes. The intervention also sometimes suppresses other existing spikes, although a new token subsequently spikes (Table 1A). Applying the same denial to the newly spiking token suppresses its spike as well on 63–100% of the affected images (Table 1B).

Table 1A: Counts per image. The denial covers every text position that spikes: one in six models, four in Qwen3-VL-32B, InternVL3.5-14B and InternVL3.5-38B. In brackets, the number of spiking image tokens per image in the clean forward. na: a single spiking token, so there is no other spike to turn off.

ModelImages# images where the spike token turned off# images: still spiking, a new spike token appeared, another natural spike turned off
Qwen3-VL-4B (1–3)595637, 25, 0
Qwen3-VL-8B (1)32328, 8, na
Qwen3-VL-32B (1–3)606022, 13, 1
Qwen2.5-VL-32B (1)60600, 0, na
InternVL3.5-4B (1–2)494918, 12, 0
InternVL3.5-8B (1–5)604538, 23, 1
InternVL3.5-14B (1–6)605938, 26, 2
InternVL3.5-38B (1)330, 0, na
Molmo2-8B (2–11)605160, 47, 11

Table 1B: In brackets, the number of other spiking tokens after the first denial.

ModelImages where another token spikes after the first denial# images where that token turned off# images: still spiking, a new spike token appeared, another spike turned off
Qwen3-VL-4B (0–3)373627, 15, 0
Qwen3-VL-8B (0–1)883, 3, 0
Qwen3-VL-32B (0–2)22229, 6, 0
Qwen2.5-VL-32B (0)0nana
InternVL3.5-4B (0–1)18186, 6, 0
InternVL3.5-8B (0–4)322026, 13, 0
InternVL3.5-14B (0–6)383832, 11, 1
InternVL3.5-38B (0)0nana
Molmo2-8B (2–11)605560, 38, 11

2. Least shared-part rule and causal mechanism

We agree that the least-shared-part rule is better supported as a predictor of spike location than as an established causal mechanism, and have revised Section 4 accordingly. To investigate its causal role, we separated the two degrees of freedom of the shared part (p = n\cos(\theta)), the norm (n) and angle (\theta), and swept each while holding the other fixed, across every image quadrant in all ten spiking models (submitted Table 33, revised Table 37; new Figure 12 and Tables 38–39).

We find that changing either component can create, relocate, or suppress spikes, but not always in the direction predicted by the least-shared-part rule (Tables 2A–2B). The relative influence of norm and angle varies across models, explaining why changes to the shared part can produce opposite effects, as in Qwen3-VL-32B. Pixtral-12B is a special case: its spike consistently appears at one fixed cell, in the last row of the image grid (97–99% of spiking images across COCO, ImageNet, and Open Images), explaining why regional edits do not relocate it. Thus, the shared part influences spike location, but the mechanism determining that location remains an open question.

Tables 2A–2B. Entries report created / moved / removed spikes (%), averaged over the four quadrant edits per image.

  • Created: Among images without a clean spike, the percentage that develop a spike inside the edited quadrant.

  • Moved: Among images whose clean spike lies outside the edited quadrant, the percentage whose spike moves inside it after editing.

  • Removed: Among images with a clean spike, the percentage that stop spiking after editing.

The images column reports the number of eligible image-edit pairs (non-spiking / spike outside the edited quadrant/spiking), summed across the four quadrant edits of 150 images. na indicates no eligible images. (s=1) and (c=1) correspond to the unmodified input (0/0/0).

Table 2A: Norm sweep, x' = s x (angle fixed)

modelimagess = 0s = 0.01s = 0.03s = 0.1s = 0.3s = 0.5s = 2s = 4s = 8
Qwen3-VL-4B136/348/46478/54/379/54/278/54/263/43/333/29/222/16/39/4/92/3/161/1/19
Qwen3-VL-8B564/27/3629/39/1133/30/1128/23/818/17/147/7/142/4/140/0/330/0/470/0/44
Qwen3-VL-32B372/171/22831/23/2631/21/2526/18/2522/17/2213/11/209/6/143/2/163/1/242/2/22
Qwen2.5-VL-32B312/216/2880/0/260/0/260/0/260/0/260/0/250/0/201/0/07/8/012/13/2
InternVL3.5-4B492/81/10820/20/3119/17/3617/8/3610/21/347/8/224/0/121/0/250/0/350/0/35
InternVL3.5-8B272/246/32859/44/1694/63/196/62/182/53/551/35/325/18/53/4/181/1/254/3/25
InternVL3.5-14B252/261/34833/29/1138/20/738/21/734/17/715/11/49/5/37/5/56/5/131/2/19
InternVL3.5-38B600/0/029/na/na15/na/na10/na/na4/na/na2/na/na1/na/na0/na/na0/na/na0/na/na
Molmo2-8B0/450/600na/0/0na/0/0na/51/0na/89/0na/76/0na/52/0na/0/0na/0/0na/0/0
Pixtral-12B0/450/600na/0/14na/0/7na/0/2na/0/0na/0/0na/0/0na/0/0na/0/0na/0/0

Table 2B: Angle sweep, shared part set to c p (norm fixed)

modelimagesc = 0c = 0.1c = 0.3c = 0.5c = 2c = 4c = 8
Qwen3-VL-4B136/348/46432/21/527/16/517/11/515/7/38/8/67/9/94/7/9
Qwen3-VL-8B564/27/364/4/222/4/251/0/191/0/142/0/113/10/363/10/33
Qwen3-VL-32B372/171/2286/6/156/5/144/4/134/2/1110/6/1623/10/2127/13/20
Qwen2.5-VL-32B312/216/28894/87/050/40/01/0/00/0/00/0/20/0/140/0/25
InternVL3.5-4B492/81/1085/2/174/2/154/9/72/1/60/1/160/1/270/1/31
InternVL3.5-8B272/246/32827/24/1119/19/1013/11/107/6/76/2/98/4/1418/13/15
InternVL3.5-14B252/261/34811/11/410/10/46/6/34/4/36/3/69/6/913/7/12
InternVL3.5-38B600/0/03/na/na2/na/na2/na/na1/na/na0/na/na0/na/na0/na/na
Molmo2-8B0/450/600na/98/0na/71/0na/5/0na/4/0na/16/0na/34/0na/43/0
Pixtral-12B0/450/600na/0/0na/0/0na/0/0na/0/0na/0/0na/0/0na/0/0

3. Added value of our nonlinear trigger search

We acknowledge that the added value of our nonlinear trigger search was not sufficiently clear. Indeed, the trigger found by our search matches the eigendirection of the quadratic form selected according to the observed spike write sign, and both trigger estimates yield nearly identical attack and intervention results. However, we have now determined three advantages of our nonlinear formulation over the original quadratic approach:

  • Direction selection. Sun et al.'s criterion selects the eigenvector associated with the largest-magnitude eigenvalue of the quadratic form. Let v−v_- and v+v_+ denote the eigenvectors corresponding to the most negative and most positive eigenvalues, respectively. At the visual switch, this criterion selects v−v_- in all ten models, but this selection is incorrect in nine of them, where the observed spike is positive, and the trigger instead aligns with v+v_+. Pixtral-12B is the exception: its spike is negative, and v−v_- is the correct trigger direction. By optimizing the MLP write according to the observed spike sign, our search identifies the correct direction for both positive and negative spikes. (Table 3A).

  • Orientation selection. Even with the correct eigendirection, the quadratic form q is even, and therefore sign-blind, q(v)=q(−v)q(v)=q(-v), whereas the actual MLP is not: only one orientation produces the observed spike, while the opposite orientation is strongly suppressed by the MLP's SiLU gate. By retaining the full nonlinearity in the optimization, our search directly identifies the effective orientation without separately testing both signs (Table 3B).

  • Memory efficiency. Our search requires approximately 200 times less GPU memory beyond the model weights than eigendecomposition, while both methods complete in under a second on an NVIDIA B200 GPU (Table 3C).

We have clarified these distinctions in the introduction and Section 3.3 of the revision. Tables 3A and 3B are combined in the new Table 9 of Appendix C.2, while Table 3C corresponds to the new Table 10.

Table 3A: Direction selection. Eigenvalues of the quadratic form at each model's visual switch block, using the model's own weights with the normalization gains folded in. v−v_- is the eigenvector associated with the most negative eigenvalue (largest in magnitude in all ten models), v+v_+ is the top positive eigenvector, and tt is our estimated trigger.

Modelλ\lambda of v−v_-λ\lambda of v+v_+Observed spike sign∣cos⁡(t,v−)∣\lvert\cos(t,v_-)\rvert∣cos⁡(t,v+)∣\lvert\cos(t,v_+)\rvert
Qwen3-VL-4B−34.09+15.00+0.0001.000
Qwen3-VL-8B−34.29+18.41+0.0001.000
Qwen3-VL-32B−69.90+39.77+0.0011.000
Qwen2.5-VL-32B−34.12+17.89+0.0001.000
InternVL3.5-4B−33.55+14.63+0.0001.000
InternVL3.5-8B−33.52+17.10+0.0001.000
InternVL3.5-14B−37.75+30.91+0.0210.999
InternVL3.5-38B−109.73+65.32+0.0001.000
Molmo2-8B−40.82+21.56+0.0001.000
Pixtral-12B−1.37+0.07−1.0000.016

Table 3B: Orientation selection. Spike-channel write for a unit-norm input, comparing the quadratic form along the trigger's eigendirection with the actual MLP evaluated at both orientations of tt.

ModelQuadratic form, either orientationMLP write along ttMLP write along −t-t
Qwen3-VL-4B+38,409+38,356+52
Qwen3-VL-8B+75,389+75,293+92
Qwen3-VL-32B+203,618+202,643+957
Qwen2.5-VL-32B+91,580+91,409+163
InternVL3.5-4B+37,462+37,410+51
InternVL3.5-8B+70,029+69,951+76
InternVL3.5-14B+158,281+156,534+1,471
InternVL3.5-38B+334,423+333,273+1,093
Molmo2-8B+88,296+88,165+127
Pixtral-12B−7,003−7,207+208

Table 3C: Computational cost. Mean and range across the ten models, with model weights preloaded. GPU measurements use an NVIDIA B200 and float64 full eigendecomposition; CPU measurements use 32 threads and Lanczos iteration for the four leading eigenvectors. The nonlinear search uses 500 Adam steps per start, with eight starts batched.

Mean over ten models (range)Quadratic form and eigenvectorsNonlinear search, 8 starts batched
GPU time0.14 s (0.04–0.20)0.46 s (0.29–0.78)
CPU time0.36 s (0.07–0.65)9.3 s (3.0–17.0)
GPU memory beyond weights970 MiB (329–1,306)4.4 MiB (2.1–10.2)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.