acceptodds
Under review as a conference paper at ICLR 2027

Can Layer-wise Relevance Propagation Capture Competition Among Transformer Logits?

Abstract

Layer-wise Relevance Propagation (LRP) is a widely used framework for attributing Transformer predictions to input tokens. However, existing Transformer-based LRP methods typically propagate relevance from a single output logit, which implicitly assumes that the prediction is governed by that logit in isolation. This assumption is limiting, since what determines the prediction is not the absolute magnitude of any individual logit but the relative relationships among logits. Intuitively, a factor that contributes equally and positively to all logits does not affect the final decision and therefore should not be attributed as important, a property that current algorithms generally struggle to model. We formalize this limitation from two complementary perspectives in transformers: low-coupling projection heads and highly coupled parameters in deeper layers, showing that single-logit relevance cannot faithfully reflect inter-class competition. To address this, we propose a top-k differential weighting strategy that constructs scalar targets by contrasting the largest logit with its strongest competitors, thereby explicitly incorporating competition into the relevance propagation process. Extensive perturbation experiments show that our method, by down-weighting common-mode evidence that contributes equally to competing logits, helps LRP capture the truly target-specific features that primarily influence the target logit, and consistently outperforms state-of-the-art baselines. Our code is available at https://anonymous.4open.science/r/RI-LRP-6AF0.

Primary area:
interpretability and explainable AI
Keywords:
Model Interpretability, Output Attribution, Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_72V4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper changes the output target used by LRP from the winning logit to its difference from the average of selected competing logits. This aims to explain relative preference rather than absolute score, while keeping the propagation rules unchanged. Validation covers language and vision tasks, using insertion/deletion-style perturbation tests, comparisons with attribution baselines, and sensitivity analysis of the number of competitors.

Critical strengths

The paper identifies a meaningful mismatch between explaining an absolute logit and explaining relative preference. The proposed seed is easy to integrate into existing LRP methods. The perturbation results show substantial gains in some settings, such as CodeJudge generation AUAC increasing from 0.4019 for AttnLRP to 0.5060.

Critical weaknesses

Better perturbation scores do not establish that the method better explains the original decision. With a large k, the target becomes the winner’s advantage over an average of many alternatives, rather than its nearest decision boundary. The supporting reparameterization control also uses a different k = 2 seed from the main definition. Together, these leave the central explanation claim less well supported than the reported performance suggests.

Questions for authors
  1. Which seed was actually used for Table 3: z1 − z2 or z1 − 0.5z2? The main definition and Appendix B disagree. Does the reported control therefore test the stated RI-LRP method?
  2. For large k, what evidence shows that the attribution better explains the original decision, rather than a different contrastive target? Please clarify how the AUAC evaluation distinguishes these interpretations.
Reviewer_Nx3dOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a new attribution method called "Relative Influence Layer-wise Relevance Propagation (RI-LRP)" to address a fundamental limitation in Transformer-based Layer-wise Relevance Propagation (LRP) techniques. Existing methods calculate feature relevance based solely on a single target logit, failing to account for competition among output classes. We demonstrate theoretically that conventional LRP can erroneously attribute "common-mode evidence"—non-discriminative information that does not influence the model's final class prediction—as significant features specific to the predicted class. To overcome this limitation, RI-LRP introduces a "top-kk differential weighting strategy," which subtracts the average of the top k-1 competing logits from the maximum logit. This formulation suppresses the influence of information shared across classes while highlighting features that contribute to distinguishing the predicted class from other competing classes. Furthermore, RI-LRP retains existing LRP propagation rules and requires no additional backward passes, thereby maintaining high computational efficiency.

Critical strengths
  1. The paper theoretically demonstrates that single-logit LRP methods fail to account for inter-class competition, erroneously attributing non-discriminative "common-mode" evidence to the predicted class. Concrete counterexamples within the classification head effectively justify the need for RI-LRP.

  2. RI-LRP captures inter-class competition by simply adjusting logit weights without altering LRP propagation rules or adding backward passes. It consistently outperforms SOTA baselines across NLP and vision tasks—notably achieving a 31.2% relative improvement on CodeJudge-Eval—delivering high attribution fidelity with minimal computational cost.

Critical weaknesses
  1. RI-LRP models inter-class competition solely by adjusting the output logit weighting, leaving the core LRP propagation rules entirely unchanged. While intuitive, this approach essentially boils down to introducing a contrastive target into existing LRP, rendering its algorithmic novelty somewhat incremental. Furthermore, while the theoretical analysis effectively demonstrates the flaws of single-logit attributions, it falls short of providing a rigorous theoretical foundation for why subtracting the uniform average of the top-k logits represents an optimal attribution strategy.

  2. The performance of RI-LRP relies heavily on the hyperparameter k, which varies drastically across tasks (e.g., k=1000 for Wikipedia and MATH-500 versus k=10 for CodeJudge-Eval). The paper selects k purely via empirical grid search without offering a theoretical rationale or an automated selection mechanism. Additionally, the empirical validation lacks fine-grained control experiments to disentangle whether the performance gains stem from genuine inter-class competition modeling or merely from general signal cancellation across multiple logits, raising concerns regarding generalizability.

Questions for authors
  1. RI-LRP propagates relevance using the difference between the maximum logit and the mean of competing logits. How does this approach fundamentally differ from existing contrastive attribution or standard logit-difference methods (e.g., k=2)? Please clarify the specific technical advantages of adopting the top-k mean difference over simple margin-based attribution within the LRP framework.

  2. The optimal value of k varies significantly across tasks (e.g., k=1000 for Wikipedia/MATH-500 vs. k=10 for CodeJudge-Eval). What specific criteria guided these choices, and how should practitioners determine k in advance for new, unseen datasets? Additionally, to what extent does performance degrade if a standard default value (such as k=2) is applied universally?

  3. The performance gains are attributed to explicit modeling of inter-class competition via uniform negative weighting (-1/(k-1)). Have the authors evaluated alternative weighting schemes, such as weighting competing logits by their Softmax probabilities or ranks? Providing such comparisons would help disentangle whether the gains stem from general multi-logit cancellation or specifically from the proposed top-k uniform weighting.

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

The paper proposes RI-LRP, which changes the attribution objective from a single predicted logit to the difference between the predicted logit and the mean of top-k competing logits, allowing hierarchical correlation propagation (LRP) to explicitly consider category competition and suppress non discriminative contributions shared by each category. The method retains the existing propagation rules and location attribution mechanism, which can be achieved through a single backpropagation.

Critical strengths
  1. The contribution of a single logit is not equivalent to its contribution to the final classification decision. The paper proposes improvements to address this difference with clear explanatory motivation.
  2. Mainly modify the output attribution target and reuse existing LRP propagation rules; The running time is close to AttnLRP and PE-LRP, and it has practical application value.
Critical weaknesses
  1. The core objective function is inconsistent between the main text and the appendix.
  2. The appendix considers SiLU and Sigmoid as positive homogeneous functions, and based on this, extends the scaling invariance of ReLU network to Transformer, but this inference does not hold.
  3. The main contribution of the method is to replace the attribution target, but the top-2 margin, log probability and other targets under the same propagation rule are not fully compared
Questions for authors
  1. Is the actual competition weight used in the main experiment -1/(k-1) or -1/k?

Comments on the paper

1 comment
Area_Chair_mEqcOpenReview ↗

Reviewer's responses to authors

Dear reviewers,

If not done already (thank you - rev. Nx3d ), could you please take a look and reply briefly to the "Official Comment(s) by Authors" / Rebuttals, possibly by an acknowledgement that you have seen it.

Feel free to respond in more detail, in particular, if you have further questions or opinions to the replies from the authors.

With best regards, an AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.