acceptodds
Under review as a conference paper at ICLR 2027

Attribution Instability in Vision Transformers: The Role of Residual Cancellation

Abstract

Vision Transformers (ViTs) are central to most modern vision models, yet obtaining input attributions that are fine-grained, faithful, and stable remains challenging. Layer-wise Relevance Propagation (LRP) has been adapted to transformer attention, but in ViTs it often produces noisy, unfaithful explanations. We show that the missing ingredient is the treatment of residual connections: cancellation effects in residual pathways lead to attribution explosion. Moreover, we find that these cancellations are substantially stronger in ViTs than in language transformers. To address this issue, we introduce Residual-aware Layer-wise Relevance Propagation (ResLRP), an extension of LRP whose propagation rules explicitly account for cancellations in residual branches, are exactly conservative, and provably bound relevance explosion. Causal channel-wise interventions confirm that residual cancellation, not a generic regularization effect, drives the instability. ResLRP substantially improves attribution quality across faithfulness and localization, evaluated on ViT architectures spanning supervised, self-supervised, contrastive, hierarchical, and multimodal families, as well as on the ground-truth-controlled FunnyBirds benchmark. The largest gains arise in modern Vision Language Models (VLMs), with +27-29% localization and up to 3.4x faithfulness scores. Beyond benchmarks, ResLRP localizes Sparse Autoencoder (SAE) features in input space, and our residual amplification measure serves as an architecture-level diagnostic predicting where attribution degrades.

Primary area:
interpretability and explainable AI
Keywords:
XAI, Computer Vision, LRP, interpretability, transformer

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_pmLgOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper is concerned with computing attributions for Vision Transformers (ViTs). Specifically, it identifies a critical failure mode of layer-wise relevance propagation (LRP) in ViTs: signals in the residual stream and the attention/MLP updates that cancel each other, can lead to attribution explosion (Sec. 3). Based on this insight, the paper proposes an improved LRP version (ResLRP) that accounts for this issue while provably satisfying desirable guarantees such as conservation (Sec. 4). Finally, the paper conducts a variety of different experiments, demonstrating consistent gains with the proposed ResLRP.

Critical strengths
  • The paper is well-written and easy to follow.
  • The paper has a clear, logical structure: problem identification, fix, empirical support.
  • The identified problem (exploding attribution scores when residual cancellation happens) and the fix are convincing.
  • The empirical analysis is extensive and nicely combines a variety of established benchmark evaluations (5.1) with an ablation to establish that the residual cancellation is indeed an important factor of the proposed method (Sec. 5.2). Further, the paper extends the experiments to sparse autoencoders and vision-language models (Sec. 5.3-5.4).
Critical weaknesses

I currently do not see critical weaknesses that would hinder me from accepting the paper. However, I have a few minor comments:

  • RoutR_{out} seems to be not defined which could help readers less familiar with LRP.
  • L. 216 "Unless stated otherwise, Rup is obtained from Rin by swapping the subscripts in and up:" - it took me a while to understand what this sentence is trying to say.
  • Do you have an intuition on why this residual cancellation would be less pronounced in language models?
  • Do you think the method could also improve LRP for language models even if the effect is less pronounced?
Questions for authors

Nothing critical. For minor points please see weaknesses.

3 comments

We would like to express our sincere gratitude for your careful review of our manuscript. We are glad that the underlying diagnosis and proposed solution resonated with you, and truly appreciate your positive assessment. Your constructive feedback has also helped us further strengthen the clarity and presentation of our work.

W1 — Undefined RoutR_\text{out}. Thank you, we agree and will add an explanation. LRP explains a prediction with a single backward pass, similar to backpropagation. It starts at the target logit, whose relevance is its value. Each layer then splits the relevance of its output among its inputs in proportion to their forward contributions, so that the total is conserved. In practice, this is a backward pass with a modified gradient g~\tilde g: the relevance of any tensor xx is Rx=x⊙g~xR_x = x \odot \tilde g_x. RoutR_\text{out} is the relevance this pass has assigned to the residual output zoutz_\text{out} after going through all layers above it. The rule in Eq. (1) splits RoutR_\text{out} into RinR_\text{in} and RupR_\text{up}, separately for every token and channel. RinR_\text{in} continues backward along the residual stream, and RupR_\text{up} goes into the attention or MLP block.

W2 — L. 216 wording. Thank you for pointing out this confusing wording. We will clarify it in the final manuscript.

W3 — Why less cancellation in LLMs? We have a hypothesis, but testing it rigorously was beyond the scope of this paper unfortunately, and we plan to address it in future work. Two observations from the paper motivate it. First, cancellation is strongest in the early ViT layers (Fig. 2a). Second, within a layer it concentrates on patches that contain edges or have high saliency (Fig. 2c). Early layers of vision networks are known to learn low-level, edge-like filters (Zeiler & Fergus, 2014; Yosinski et al., 2014), and ViTs learn similar filters in their patch embedding (Dosovitskiy et al., 2021). Such filters must respond strongly to small changes in local contrast. A cancellation is one way to achieve this sensitivity: when zout=zin+zupz_\text{out} = z_\text{in} + z_\text{up} is the sum of two large branches of opposite sign, a small relative change in either branch changes zoutz_\text{out} a lot. Language models, on the other hand, operate on symbolic text manifolds, so there is no continuous low-level signal that demands such sensitivity. Since this remains a hypothesis without conclusive evidence, we did not state it in the paper.

W4 — ResLRP for LLMs. We thank the reviewer for this valuable suggestion, which prompted us to strengthen the paper further. We applied ResLRP to language models and observed no improvement over AttnLRP. This is consistent with our analysis: residual amplification is clearly weaker in LLMs than in ViTs and AttnLRP heatmaps of LLMs are already clean. With little amplification, there is little for ResLRP to correct. We will add this negative result to the paper, as it supports cancellation as the cause of the noise in ViT heatmaps. Interestingly, this does not apply to LLM backbones in VLMs. Applying ResLRP on LLMs in VLMs improves the faithfulness. We believe this is due to the fact that image domain features of the vision backbone leak into the langauge backbone.

Thank you again for taking your time reviewing our paper, we are truly encouraged by the positive review.

I thank the authors for the further clarification. After reading the reply, alongside the other reviews and their replies, I maintain my initial score and still vote for accepting the paper.

We would like to express our heartfelt gratitude for your exceptionally encouraging feedback. Your highly positive assessment of our work means a great deal to us and is truly motivating. We sincerely appreciate the time and consideration you have devoted to reviewing our manuscript, as well as your continued support.

Reviewer_GsZUOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
3
Read review
Paper summary

Work introduces new LRP rule for ViTs, that is related to residual connections. Specifically, residual cancellations make LRP divide by near-zero. So ResLRP is introduced. It is a simple fix on top of AttnLRP. It makes ResLRP as bounded, conservative γ-rule at residual additions. Experiments are on SRG, IN-S localization, FunnyBirds, using VLMs and SAE models.

Critical strengths

simple solution, to interesting phenomena (discovered by the authors)

broad and consisted gains in metrics

Different benchmarks showcasing the superiority of the approach.

Nice ablations, such as component ablations

Interesteding extension to SAEs.

Critical weaknesses

Overclaiming and confusion:

LRP-γ is a rule that favors positive weight contributions. Here it favors branches whose sign agrees with the output. So it is not true!

In abstract there is "exactly conservative". Limitations say that AttnLRP attention rule breaks end-to-end conservation. So conservation holds locally at the residuals.

The “provable bound” is (1+2/γ)^{2L}. For ViT-L that’s 3^48, so it’s vacuous as a model-level guarantee.

Causal intervention is a bit different: it ranks channels by amplification and then shows that fixing high-amplification channels helps most. That’s close to circular. So, the rule matters where the rule changes things. As a result, it is not that cancellation causes bad explanations.

Small inconsistencies: z_res in Fig. 1, z_in in the text. The Fig. 1c legend says “ResiLRP”.

Questions for authors

Why does the supervised ViT-B/16 with the strongest cancellation (Fig. 6) missing from Table 1? And does “ViT-B/16” refer to the same checkpoint in Table 1 and Fig. 4? I thing it is confused with CLIP ViT version.

Why does AttnLRP’s SRG on ViT-B/16 vary (Tables 1, 9 and 11)?

Is the 0.99 C to A correlation expected by construction? As A uses the same z_out.

Does VLMs show higher C than ViTs? Where is it in the paper? (I see it is claimed)

Can you compare ResLRP to simply increasing ε or using the z⁺-rule in the network where residual additions happen?

I wonder what is the effect of ResLRP and what is the effect of 4 averages? Can you check it on SRG and localization?

Also, DAVE is not in related works - please update them and extend them, at least extended version should be in appendix as for now they are very short and mechanistically listed.

5 comments

We sincerely appreciate the time you dedicated to reviewing our manuscript. Thank you for recognizing residual cancellation as our discovery, and for checking each formal claim against its exact scope.

W1 — What γ favors. Thank you for raising this. You describe the original LRP-γ (Montavon et al., 2019), designed for ReLU networks, where positive contributions are exactly those agreeing with the (positive) output. We use its generalization to signed inputs and outputs (introduced by Andéol et al., 2023), which upweights by (1+γ)(1+\gamma) the contributions whose sign agrees with the output. This is also the LRP-γ implemented in Zennit (Anders et al., 2021), the library we use for our experiments. Eq. (3) is exactly this rule applied to the residual addition as a linear layer with unit weights. We will add the missing citation of Andéol et al. (2023) and state the generalization explicitly.

W2 — "Exactly conservative". The reviewer is correct: the guarantee is local. ResLRP extends existing composites such as AttnLRP or CP-LRP and replaces only the rule at the residual additions, which is exactly conservative (Sec. 4.1). It therefore never breaks conservation itself, and the full propagation is conservative whenever the underlying composite is, e.g. with CP-LRP. We will rephrase the abstract to "exactly conservative at every residual addition" and state this explicitly in Sec. 4.

W3 — Bound vacuous at depth. We agree, thank you. Compounded over all residual additions, (1+2/γ)2L(1+2/\gamma)^{2L} is a worst case and vacuous as a model-level guarantee. The meaningful guarantee is only local: at each residual addition, ∣Rin∣+∣Rup∣≤(1+2/γ)∣Rout∣|R_{\text{in}}| + |R_{\text{up}}| \le (1+2/\gamma) |R_{\text{out}}| (at most 3 for γ = 1), whereas the standard rule admits no bound, as its amplification grows without limit as zout→0z_{\text{out}} \to 0. We will remove the model-level bound from the abstract and Sec. 4.1 and state the guarantee per residual addition.

W4 — Circularity of the causal test. Our goal is to demonstrate that the channels with the strongest relevance explosion are responsible for the low SRG of AttnLRP. To this end, we rank all residual channels by their relevance amplification: either the most-exploding channels, or two controls of the same size, a random subset and the least-exploding channels. The most-exploding channels recover 48–81% of the full ResLRP gain, versus 13–52% for random channels and at most 4% for the least-exploding ones. The low SRG of AttnLRP is thus caused by the most-exploding channels.

However, we agree that ranking by relevance amplification may seem to presuppose the result. We therefore repeated the experiment with a ranking that involves no LRP quantity: the forward activation cancellation of each channel, (∑∣zin∣+∑∣zup∣)/∑∣zout∣(\sum |z_{\text{in}}| + \sum |z_{\text{up}}|) / \sum |z_{\text{out}}|. Forward cancellation predicts where relevance explodes: the 1% most-cancelling channels alone account for 10–20% of the total relevance amplification under AttnLRP. Share of the full ResLRP gain recovered with 20% of channels treated:

Channels treated (20%)ViT-B/16CLIP ViT-B/16SigLIP2-B/16DINOv2-B
highest relevance amplification (Fig. 4)51%81%48%80%
highest activation cancellation (new)42%63%59%87%
random subset (control)20%37%13%52%
lowest relevance amplification (control)4%−5%3%−17%
lowest activation cancellation (control, new)3%32%4%16%

Both rankings beat the random control in all four models (activation ranking: paired Wilcoxon p<10−35p < 10^{-35}), while the lowest-ranked channels recover almost nothing. The gain thus follows the forward cancellation itself, not an LRP-derived score.

W5 — z_res/z_in, "ResiLRP". Thanks a lot, we will correct the spelling.

Q1 / Q2 — ViT-B/16 naming (Table 1, Fig. 4, Fig. 6) and AttnLRP values across Tables 1, 9, 11. Thank you for catching this ambiguity on our side. "ViT-B/16" denotes three different checkpoints of the same architecture. We now clarify the checkpoints more precisely in Table 1 and Fig. 4:

Name in the revisionCheckpointUsed inAttnLRP → ResLRP (SRG)
ViT-B/16 (OpenAI)CLIP, vit_base_patch16_clip_224.openai_ft_in12k_in1kTable 11.51 → 3.20
ViT-B/16 (LAION)CLIP, vit_base_patch16_clip_224.laion2b_ft_in1kTable 92.29 → 3.74
ViT-B/16 (AugReg2)supervised, vit_base_patch16_224.augreg2_in21k_ft_in1kFig. 4, Table 110.62 → 2.60

Table 9 uses the LAION checkpoint so that the γ-sweep covers one pretraining lineage at three sizes (B/L/H; L and H are the Table 1 models).

Fig. 6. The figure is a qualitative illustration with models picked from a broad set of families. Its supervised checkpoint (vit_base_patch16_224.augreg_in21k) only has an ImageNet-21k head and cannot be evaluated with the ImageNet-1k protocol of Table 1. We therefore evaluated its ImageNet-1k fine-tuned counterpart, ViT-B/16 (AugReg2), with the Table 1 protocol.

We now name each checkpoint unambiguously in the text and all tables, and Table 5 lists the identifier of each.

Q3 — C–A correlation by construction? A strong correlation is not expected by construction. As we note in Sec. 3 (l. 188–196), under the standard rule Rin=zinzout Rout,R_\text{in} = \frac{z_\text{in}}{z_\text{out}}\,R_\text{out}, the factor zin/zoutz_\text{in}/z_\text{out} captures the forward-pass cancellation, but the resulting amplification also depends on RoutR_\text{out}, which is determined by the backward pass and is not determined by the forward activations at that site. Thus, even strong cancellation does not necessarily lead to strong amplification if the corresponding entry carries little relevance (Rout≈0R_\text{out}\approx 0). Consequently, C\mathcal{C} and A\mathcal{A} do not need to agree. For example, Qwen2.5 exhibits strong cancellation but little amplification (Fig. 2b).

Q4 — VLM cancellation.

Thank you for this question. We measured and report now the residual cancellation directly inside VLMs. The measurement confirms that the mechanism of Sec. 3 carries over to VLMs and that ResLRP bounds the resulting amplification in every vision tower. It also lets us state the comparison to standalone ViTs more precisely: VLM vision towers span the same range of cancellation as ViTs rather than exceeding it in general. We have refined the claim in Sec. 1 accordingly.

Measurement. Same C\mathcal{C}/A\mathcal{A} definitions and residual probe as Sec. 3, at every residual addition of the vision tower and the LLM, on the images of the VLM table (forced-choice ImageNet-S, 400–500 images per model). As reference we re-measured 20 standalone ViTs (the 13 CLIP ViTs of Fig. 2, DeiT3, DINOv2 with and without registers, SigLIP2, SwinV2) on exactly the same images.

vision C\mathcal{C}vision C\mathcal{C}, first 20% of blocksvision A\mathcal{A}: AttnLRP → ResLRPLLM C\mathcal{C}LLM C\mathcal{C}, first 20% of blocksLLM A\mathcal{A}: AttnLRP → ResLRP*
ViTs (20, range)1.17–1.561.25–1.851.14–1.53 → 1.04–1.12–––
Qwen2.5-VL-3B1.261.331.27 → 1.061.221.121.24 → 1.05
Gemma-3-4B1.291.301.28 → 1.071.211.181.22 → 1.05
Qwen3-VL-4B1.481.661.48 → 1.121.251.151.23 → 1.04
Qwen3-VL-8B1.611.801.60 → 1.131.261.181.23 → 1.04

*γ\gamma applied to the LLM residuals as well.

Within our three VLMs, higher C\mathcal{C} coincides with larger faithfulness gains, with Qwen3-VL-4B achieving a 3.4×3.4\times improvement over AttnLRP, compared with 1.9×1.9\times for the other models. While this trend is not conclusive given the small sample, it supports our analysis. Accordingly, Sec. 1 will now highlight residual cancellation in VLM vision towers and the effectiveness of ResLRP, with supporting results added to the Appendix.

Q5 — ε and z⁺ at residuals. We applied both rules at all residual additions, keeping AttnLRP elsewhere, and evaluated them with the Table 1 protocol (logit SRG, 500 images).

The z⁺-rule was designed for non-negative outputs, e.g. after a ReLU. Residual streams are signed, and for negative outputs the rule is ill-defined: if both branches are negative, the denominator vanishes and the relevance is lost. On 20 checkpoints, z⁺ therefore drops to the level of a random heatmap (SRG −0.09 to 0.57).

We also swept the classical ε-rule, Rin=zinzout+sign(zout)εRoutR_{\text{in}} = \frac{z_{\text{in}}}{z_{\text{out}} + \text{sign}(z_{\text{out}}) \varepsilon} R_{\text{out}}, over ε ∈ {0.01, 0.1, 1, 10}:

ModelAttnLRPε = 0.01ε = 0.1ε = 1ε = 10ResLRP
ViT-B/16$_\text{OpenAI}$1.512.522.510.00−0.093.20
DeiT3-L/162.692.913.272.81−0.233.38
DINOv2-L1.653.560.46−0.160.054.72

No value of ε reaches ResLRP, and the rule is sensitive to ε: the best value differs between models, and large values collapse SRG to random level. The ε-rule absorbs relevance at every entry, not only where the branches cancel, and does not conserve it. ResLRP changes relevance only at cancellations, where the amplification arises, conserves it at each residual addition, and does not need model-specific tuning.

Q6 — ResLRP vs. 4 averages. As suggested, we evaluated the different residual rules (AttnLRP vs. ResLRP) with and without a 4-view averaging (off vs. on) on the 7 backbones of the main table with the main-table protocols. Cells show SRG / Loc. as the mean over the 7 backbones (higher is better).

no averaging4-view averagingΔ averaging
AttnLRP1.50 / 0.3861.55 / 0.409+0.05 / +0.023
ResLRP3.68 / 0.5903.47 / 0.612−0.21 / +0.021
Δ rule+2.18 / +0.205+1.92 / +0.203

We find that the rule is the dominant effect. Switching AttnLRP → ResLRP raises SRG by about +2 and localization by about +0.20, with or without averaging. Per backbone, the gain is positive and significant on all 7 for both metrics (paired per-image test, 95% CI). Averaging is a small, metric-dependent effect. It adds about +0.02 localization under both rules (on 7/7 backbones). Its effect on SRG depends on the rule: mixed on AttnLRP (+0.19 to −0.14), but negative on ResLRP for 7/7 backbones (significant on 6/7). Averaging smooths the maps. This helps mask overlap but costs patch-level faithfulness once the underlying map is already sharp.

ResLRP+ is therefore a trade-off variant: it gains localization and loses SRG relative to ResLRP. The improvement over AttnLRP comes from the residual rule, not from the averaging. Per-backbone numbers with paired confidence intervals will be added to the appendix.

Q7 — DAVE, related work. Thank you for raising this. We will add DAVE (Wróbel et al., 2026) to Sec. 2 and include an extended related-work section.

We thank the reviewer for the careful review, which made our formal claims more precise and strengthened the evidence through the forward-only causal test, the VLM measurement, and the new rule comparisons. We are happy to answer any remaining questions.

Reviewer_pG3hOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies explanations in Vision Transformers, VLM made with LRP. They motivate this problem by finding the cause of noise, the residual addition. Then they explain inside each block, the stream and the update often cancel, so their net sum is almost 0. Which LRP divides the sum by that sum when tracing back, so very less evidence turns into huge values, it could be positive or negative. They call this relevance explosion and measure it with C (forward cancellation) and A (backward blowup). They fix it by presenting their method ResLRP, which applies γ rule at each addition. This not only gives more relevance to the branch that agrees with the sum but also keeps the relevance constant. They test this on different Vit, datasets, SAE eatures and VLMs.

Critical strengths

The diagnosis is new and novel, the authors identifyi one mechanism, cancellation at residual additions. The method is nicely motivated and supported using measure, bound and intervention. This mechanism itself is a big contribution to the attribution community irrespective of the results and the analysis. The experimental analysis is broad and is done on multiple Vit, VLM families and on multiple datasets which is supported by intense statistical and qualitative analysis throughout the paper.

Critical weaknesses

The manuscript over all is good, but there are a few concerns and those are as follows:

  1. The title uses the word stability, and authors claimed they made the method more stable but stability is not measured anywhere in the paper no metric, no heatmap or anything. The authors also claim specalised rules such as ε and γ rules improve stability, they briefly discuss about γ-rules in appendix but there is not discuss about ε.
  2. ResLRP+ is the best looking method in qualitative analysis, Fig.1 and Fig. 3 but never talked about in the whole paper except for the Appendix 7.4. No quant. analysis anywhere else in the paper.
  3. Section 5.2 abalation is overclaimed when compared to Table 11 in appendix, and to strengthen the claim authors could do additional analysis/ abalation on other models/datasets.
  4. VLM cancellation is claimed in Section 1 but never measured in the paper, also the VLM is headlined as the strongest result in the abstract but it has the weakest evidence. The claim is only tested on 3 models with max 4B parameters, with only 2 baselines no calculating C or A.
  5. DAVE looks clean in Fig. 3, and still in not included in any tables. The methods such as LeGrad and CheferLRP which dont look clean tie or beat ResLRP as shown in Table 1 and Table 8. Which brings me on to next point.
  6. Table 1 is a more stronger subset of given Table 8 in the appendix.
  7. The causal test in 5.2 could be more detailed, as it currently dont use a control set and neither do they show layerwise version when the authors claim cancellation is concentrated in early layers.
  8. Authors imply one backward pass for ResLRP but it doesnt do any time or memory analysis.
  9. Every table uses 500 samples to test the claim, for example ImageNet 1k val has 50000 images over 1000 classes, a random 500 sample leaves around 40% of classes which may or may not be discusses or are a part of experimental analysis. This is especially important when we consider results in Table 1 and Table 8 when one of the baseline is competitive or outperforming the proposed method.
Questions for authors
  1. A small experiment or a stability analysis to back the claim would make the claim and paper more strong along with analysis and values varying ε.
  2. Authors need to decide the relevance of ResLRP+, if they want to exclude, exclude it completely but if they want to include as a stronger/ finer version they would need to justifyi why it is stronger using both qualitative and quantitative analysis.
  3. Authors would benefit by either strengthening their abalation based on their claim or make a better claim suited to the emperical analysis.
  4. Authors would benefit by strengthening their VLM claim, which is not justified as a main claim in current manuscript.
  5. Inclusion of DAVE in different analysis would be beneficial, along with a reason to justifyi how are methods that dont do a good job explaning in qualitative analysis but somehow do good in experical analysis specially in Table 8 in appendix.
  6. The causal experiment could also be benefitted by using a controlled set along with a layer wise analysis to show the cancellation claim.
  7. A time/ memory analysis would also benefit the paper.
  8. Authors would also benefit by explaining why AttnLRP has 3 different values for "Vit-B/16" across 3 different tables ourside the error given in those tables 7,table 9, table 11.
  9. Authors would also benefit by improving the sample size that make sure atleast all of the classes are used while coming to a particular conclusion.
6 comments

We are very grateful for the time you devoted to evaluating our paper. Thank you for seeing residual cancellation as a contribution in its own right, and for cross-checking our appendix tables so closely.

W1 / Q1 — Stability and ε. By stability we mean the numerical stability of relevance propagation: at a cancellation, the standard rule divides by a near-zero zoutz_\text{out}, so small changes in the activations produce large, sign-flipping relevance (Sec. 3). We measure this amplification inside the network via A\mathcal{A} (Fig. 2), bound it per residual addition (Sec. 4.1), and show its effect on the explanations in the heatmaps of Figs. 1 and 3, where the unstable propagation appears as high-frequency noise of alternating sign.

To quantify the latter, we now additionally report the relevance explosion of the input heatmap, G=∑∣R∣/∣∑R∣G = \sum|R| / |\sum R|, i.e., the total absolute relevance relative to the net relevance (1 = no cancelling mass), computed on the heatmaps of Table 7. AttnLRP reaches GG = 1.1–39, while ResLRP stays at 1.00–1.25. We also swept the classical ε-rule at all residual additions, Rin=zinzout+sign(zout)εRoutR_{\text{in}} = \frac{z_{\text{in}}}{z_{\text{out}} + \text{sign}(z_{\text{out}}) \varepsilon} R_{\text{out}} (logit SRG, Table 1 protocol):

ModelAttnLRPε = 0.01ε = 0.1ε = 1ε = 10ResLRP
ViT-B/16$_\text{OpenAI}$1.512.522.510.00−0.093.20
DeiT3-L/162.692.913.272.81−0.233.38
DINOv2-L1.653.560.46−0.160.054.72

No value of ε reaches ResLRP, while large values collapse SRG to random level. ε absorbs relevance at every entry and does not conserve it. ResLRP changes relevance only at cancellations and needs no tuning.

W2 / Q2 — Role of ResLRP+. As the reviewer noticed, ResLRP+ yields the visually best heatmaps, which is why we would like to keep highlighting it. ResLRP+ shows that ResLRP is orthogonal to the recent DAVE (Wróbel et al., 2026): DAVE removes patch-grid artifacts at the input, while ResLRP corrects the residual additions inside the network, and the two compose directly. We apologize that, due to space constraints, we focused on our core contribution of explaining and correcting residual cancellation.

We now include quantitative evaluations and comparisons of both ResLRP+ and DAVE in the paper. See GsZU Q6 for the details of ResLRP, ResLRP+ and AttnLRP+Avg. DAVE and ResLRP+ are now included in the quantitative localization and faithfulness tables in the paper. Across all 10 backbones, DAVE is never the best or second-best method on either metric. Its SRG is below ResLRP on all 10 models and roughly at AttnLRP level (higher on 5, lower on 4, tied on SwinV2-L). Its localization is below ResLRP on 9 of 10 models; the exception is DeiT3-L (0.60 vs. 0.55). ResLRP+ outperforms DAVE on SRG on all 10 models and on localization on 9 of 10, consistent with the two methods being complementary: input-side averaging removes grid artifacts but does not address residual cancellation, which is where most of the faithfulness gain comes from.

We will add this analysis to App. 7.4 and clearly label ResLRP+ in Fig. 1.

W3 / Q3 — Sec. 5.2 ablation vs. Table 11. Thank you for the careful comparison. Could the reviewer please clarify whether this refers to the component-ablation paragraph in Sec. 5.2 (l. 410–417)? If so, we would like to clarify its purpose. ResLRP does not replace the attention and LayerNorm rules of AttnLRP. These rules fix other problems, and ResLRP keeps them, changing only the rule at the residual additions. The ablation is therefore a sanity check: does the γ-rule still improve over AttnLRP when any one of these rules is removed? Table 11 confirms this: ResLRP exceeds AttnLRP in 11 of 12 ablated settings. That removing the attention or LayerNorm rule lowers SRG for both methods is expected, since each rule fixes its own problem. The ablation is not central to our claims, which rest on the benchmarks in Sec. 5.1 and the targeted interventions in Fig. 4. We will rephrase the paragraph to state this purpose explicitly.

W4 / Q4 — VLM evidence. Thank you, you are right: the claim was not measured. We have now measured it. VLM vision towers cancel about as strongly as standalone ViTs, not more, and ResLRP bounds the resulting amplification in every tower. We measured C\mathcal{C} and A\mathcal{A} at every residual addition of four VLMs (same probe as Sec. 3, 400–500 ImageNet-S images per model) and of 20 standalone ViTs on the same images:

vision C\mathcal{C}vision C\mathcal{C}, first 20% of blocksvision A\mathcal{A}: AttnLRP → ResLRPLLM C\mathcal{C}LLM C\mathcal{C}, first 20% of blocksLLM A\mathcal{A}: AttnLRP → ResLRP*
ViTs (20, range)1.17–1.561.25–1.851.14–1.53 → 1.04–1.12n.a.n.a.n.a.
Qwen2.5-VL-3B1.261.331.27 → 1.061.221.121.24 → 1.05
Gemma-3-4B1.291.301.28 → 1.071.211.181.22 → 1.05
Qwen3-VL-4B1.481.661.48 → 1.121.251.151.23 → 1.04
Qwen3-VL-8B1.611.801.60 → 1.131.261.181.23 → 1.04

*γ applied to the LLM residuals as well.

Among the three VLMs we evaluated, the one with the highest C\mathcal{C} (Qwen3-VL-4B) also shows the largest faithfulness gain (3.4× vs. 1.9×). We will revise Sec. 1 and the abstract to state that VLMs show cancellation comparable to ViTs rather than the strongest.

W5 / Q5 — DAVE; LeGrad/CheferLRP in Tables 1, 8. DAVE is now included in the quantitative localization and faithfulness tables; please see W2 / Q2 above for the comparison across all 10 backbones, and our reply to Reviewer GsZU, Q6, for the 2×2 analysis separating the residual rule from the 4-view averaging. In short, DAVE's clean appearance does not translate into faithfulness: its SRG is below ResLRP on all 10 models and roughly at AttnLRP level, because input-side averaging removes grid artifacts but leaves residual cancellation untreated.

Regarding LeGrad and CheferLRP: the reviewer's observation is correct, and we think the tables explain it. Our two metrics capture different aspects of a heatmap. SRG occludes whole patch tokens, so it only tests whether patches are ranked correctly; noise, blur or grid artifacts within and across patches do not affect it. Localization measures how much relevance falls on the object, which is closer to what one judges visually. The baselines split along this line. LeGrad ranks patches well (SRG within the standard error of ResLRP on ViT-B/16 and ViT-L/14) but its localization is below ResLRP on every backbone it is defined for (0.33–0.53 vs. 0.51–0.62). CheferLRP shows the opposite: competitive localization on ViT-B/16 and DeiT3-L, but lower SRG on every backbone. DAVE is visually clean yet at AttnLRP level on SRG. ResLRP is the only method that is best or second-best on SRG on all seven backbones of Table 1 and on localization on six of them. We will add this discussion to Sec. 5.1 and mark close cells as ties.

W6 — Table 1 vs. Table 8. Thank you for reading our manuscript carefully. Table 1 shows indeed 7 of the 10 backbones; we will include more in the revised paper. On the three omitted backbones, ResLRP still achieves the best SRG, our primary faithfulness metric, on DINOv2-S and DINOv2-S (Reg.), and more than doubles the SRG of AttnLRP on DINOv2-L (Reg.). It is also the only strong method defined for every backbone (e.g., LeGrad is not defined for DINOv2 and SwinV2).

W7 / Q6 — Controls and layer-wise causal test. Thank you for the valuable suggestion, which strengthens our paper. Fig. 4 already contains two controls: a random channel subset of the same size and treating the lowest-amplification channels first. With 20% of channels treated, the highest-amplification channels recover 48–81% of the ResLRP gain, a random subset 13–52%, and the lowest-amplification channels at most 4%.

We now added a third control. One could argue that any change at the high-amplification channels improves SRG, regardless of the rule. We therefore keep the channels and budget fixed and replace only the rule applied to them: a uniform ½ split, which is conservative and bounded but ignores each branch's contribution, and the ε-rule at two strengths, which damps the denominator. The ½ split collapses SRG to chance level on all four models, and the ε-rule recovers the gain only partially. The gain therefore stems from how relevance is redistributed at the cancellation sites, not from altering these channels per se.

Moreover, we applied the γ-rule to one depth range at a time, keeping AttnLRP at all other residual additions. Share of the ResLRP SRG gain recovered:

Blocks treatedViT-B/16CLIP ViT-B/16DINOv2-B
0–366%81%93%
4–714%57%94%
0–7101%112%133%
8–1122%8%7%

Cancellation per site is highest in blocks 0–3, and the gain follows: blocks 0–7 recover the full gain, blocks 8–11 at most 22%. We will add both analyses to the paper. Values above 100% mean that treating blocks 0–7 alone is slightly more faithful than treating all blocks. We nonetheless apply γ at all residual additions, as this requires no per-model choice of depth.

W8 / Q7 — Time and memory. Like a gradient, LRP needs one forward and one backward pass. Implemented via .detach() (Table 4), the LRP rules only change what is propagated, so LRP is computed by a standard .backward(). ResLRP only adds an elementwise rule at the residual additions. We measured the time relative to one plain gradient pass (forward + backward, batch 64, one H200 GPU, median over 10 runs):

Relative to one gradient passViT-B/16ViT-L/14
AttnLRP1.14×1.10×
ResLRP1.29×1.22×
AttnLRP peak memory0.92×0.87×
ResLRP peak memory1.22×1.20×

The remaining overhead stems from our implementation, which is not yet optimized (e.g., the elementwise rules are not fused into single kernels). ResLRP thus stays within 1.3× of a single gradient computation, whereas perturbation- and sampling-based methods require 10 to 1,000 of forward passes.

Q8 — AttnLRP ViT-B/16 across Tables 7, 9, 11 Thank you for catching this; it is a labelling error on our side, and we apologize. The three values come from three different checkpoints that share the ViT-B/16 architecture: Table 7 uses the CLIP checkpoint of Table 1 (openai_ft_in12k_in1k), Table 9 the CLIP laion2b_ft_in1k checkpoint (same pretraining lineage as the L and H models of the γ-sweep), and Table 11 the supervised augreg2_in21k_ft_in1k checkpoint. On identical images, ResLRP is above AttnLRP on all three. We now name every checkpoint explicitly in the text and tables and extended Table 11 with the Table 1 checkpoint. Please see our reply to Reviewer GsZU, Q2, for the details.

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

This paper studies amplification in Layer-wise Relevance Propagation (LRP) for Vision Transformers and relates it to cancellation between residual branches. It proposes ResLRP, which uses a sign-aware gamma propagation rule at residual additions. The rule preserves local relevance conservation and provides an upper bound on relevance amplification at an individual residual node.

Critical strengths
  1. The analysis of residual cancellation and relevance propagation is intuitive, and the proposed modification is straightforward to implement without retraining the prediction model.

  2. The experiments cover different pretraining setups and model architectures, and the reported gains are not confined to a single backbone.

Critical weaknesses
  1. When a residual addition is represented as a bias-free linear node with weights [1, 1], the proposed rule can be recovered as a special case of the existing generalized LRP-gamma rule, under the stated nonzero-output and numerical-stabilizer assumptions. The contribution is therefore better characterized as identifying and studying the application of this existing rule at Transformer residual nodes, rather than introducing an entirely new propagation principle.

  2. Bounded relevance amplification does not automatically imply more faithful attribution. Positive and negative contributions can genuinely cancel. The targeted-channel intervention supports focusing on problematic channels, but does not rule out a more general adaptive-stabilization explanation.

Questions for authors
  1. Does Equation (3) differ mathematically from generalized LRP-gamma applied to an additive node, beyond the selected application site and numerical implementation?

  2. Can alternative stabilization rules be applied to the same high-amplification channels under matched channel rankings and intervention budgets, to determine whether the reported gain is specific to ResLRP?

  3. Does the actual localization implementation use absolute values or positive-only relevance, and does this choice change the method ranking?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.