acceptodds
Withdrawn from ICLR 2027

DECAF: Decomposition of Evidence, Contradiction, and Fragility in Perturbation Responses

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Perturbation methods explain model decisions by measuring prediction changes under altered inputs, but response magnitude tells us only how much a model reacts, not what that reaction means. The same magnitude can support the final factual–counterfactual difference, oppose it, or arise strongly along the perturbation path yet vanish at the endpoint. We therefore track how the contrast develops as paired inputs are progressively revealed, using the final contrast to interpret the trajectory. We introduce DECAF (Decomposition of Evidence, Contradiction, And Fragility), which routes aligned, opposed, and endpoint-null responses into evidence E, contradiction C, and fragility F. The decomposition preserves ordinary magnitude exactly, Abs = E + C + F, and is unique under endpoint-relative axioms. Across controlled vision and tabular settings, the three components track independently measured behavior. In a 72-model ImageNet-9 audit, we compare cases with nearly identical response magnitude but different independently measured behaviors. The largest DECAF component agrees with an observed behavior in 96.4% of cases, compared with 35.0% for magnitude alone. Changing only the reveal path increases total response by nearly 80%, yet evidence barely changes while fragility grows by more than 4×. On FunnyBirds and ImageNet-1k, short forward-only DECAF trajectories outperform the tested general-purpose attribution baselines. On a 1B-scale DINOv2 model, a short trajectory matches a strong gradient-based baseline with 4.75× lower wall time and 2.36× lower peak memory.

Primary area:
interpretability and explainable AI
Keywords:
Feature Attribution, Counterfactual Explanations, Model Explainability, Sensitivity Analysis

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

The paper proposes new explanation method called DECAF which consists of evidence, contradiction, and fragility scores. DECAF only requires multiple forward passes to calculate its scores unlike previous method requiring internal activation or gradient. It also reveals the direction of the model reaction change instead of simply showing the magnitude. The paper not only provides the theoretical grounding but also specific correlation, showing that DECAF’s E,C, and F is related to independent model behaviors in control experiments. DECAF is also shown valid in natural images, external attribution benchmark, etc.

Critical strengths
  1. E,C, and F elements of DECAF are carefully defined and intuitive to understand.

  2. DECAF solely depends on forward passes while being much faster than previous baselines.

  3. Comprehensive analysis and experiments are conducted (control diagnostic, natural image, external attribution, etc), which allows the reader to assess the DECAF in various perspective.

Critical weaknesses
  1. It’s a bit overclaim that DECAF is model-agnostic. L78-80 argue that DECAF can be applied to any model that returns a scalar but mainly tested on image classifier based on Resnet and ViT. Whether it also works for more advanced architecture based on LLM or VLM (such as Jev [1] or VLM embedding models [2,3]) are still uncertain.

  2. Theory is actually not showing anything special. It does not prove theoretical bound when DECAF become useful compared to Abs value or explain why DECAF is better aligned with real explanation compared to other baselines. It reiterates the definition in other ways with some case studies. I would suggests to move Sec 4 (Theoretical Foundation) to the Appendix while include more details in other sections. Due to the space occupied by Sec 4, lots of details are missed in other section which highly reduce the readability.

  3. The scope of experiments on natural images is very limited. Only two interventions are tested in the paper: background replacement (ImageNet-9) and patch or part deletion (Sec. 7), which is either too simple or bit artificial. No concept or attribute level counterfactual is evaluated on real images. Can author test it on more fine-grained counterfactual set based on CUB dataset (bird attribute) [4]?

  4. No single experiment tests all three components together in Sec 5. Each component gets its own task, model set, and behavior, which seems bit too much for controlled experiment. So it is bit unclear whether the similar controlled setup for Evidence would also show high correlation score for Fragility and Contradiction.

[1] Deußer, T., Sparrenberg, L., & Sifa, R. (2026). Evaluating and benchmarking the system one model Jev. arXiv preprint arXiv:2609.37647.

[2] Jiang, Z., Meng, R., Yang, X., Yavuz, S., Zhou, Y., & Chen, W. (2024). Vlm2vec: Training vision-language models for massive multimodal embedding tasks. arXiv preprint arXiv:2410.05160.

[3] Li, M., Zhang, Y., Long, D., Chen, K., Song, S., Bai, S., ... & Lin, J. (2026). Qwen3-vl-embedding and qwen3-vl-reranker: A unified framework for state-of-the-art multimodal retrieval and ranking. arXiv preprint arXiv:2601.04720.

[4] Bader, J., Girrbach, L., Alaniz, S., & Akata, Z. (2025, October). Sub: Benchmarking cbm generalization via synthetic attribute substitutions. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 23188-23198). IEEE.

Questions for authors
  1. What is the effective strategy to choose threshold (epsilon) for distance d?

  2. How the practitioner know when to use DECAF and when Abs score is sufficient?

Reviewer_t7mYOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Given a model score function q:X→Rq : \mathcal{X} \rightarrow \mathbb{R}, an input x+x^+ and a paired sample x−x^-, obtained from a counterfactual intervention, the contrast d=q(x+)−q(x−)d = q(x^+) - q(x^-) quantifies the response of the model to a factor-level change. In this context, the Authors define reveal paths for the two samples x+(t),x−(t),t∈[0,1]x^+(t), x^-(t), t \in [0,1], with x+(0)=x−(0)=x0;x+(1)=x+x^+(0)=x^-(0) = x_0; x^+(1) = x^+ and x−(1)=x−x^-(1)=x^-, through perturbations with magnitude dependant on tt. The Authors then decompose the response r(t)=q(x+(t))−q(x−(t))r(t) = q(x^+(t)) - q(x^-(t)) along the reveal path into Evidence, Contradiction and Fragility based on its agreement with the final contrast dd, dubbing this procedure as DECAF.

If the final contrast dd exceeds a predefined threshold $\epsilon $, Evidence integrates the magnitude of responses with the same sign as dd, while Contradiction integrates the magnitude of those with the opposite sign. For pairs where the $d<\epsilon $, the magnitude of all intermediate responses is integrated into the Fragility measure. DECAF thus measures the alignment between the response to a counterfactual intervention, and the response to said intervention, combined with an additional, generic perturbation.

In controlled settings, the Authors show that the three components of DECAF correlate well with different model behaviours (e.g., background reliance). When simple response magnitude is similar, their decomposition still recovers the inherent behaviours with high accuracy. The choice of the reveal path can disproportionately alter component magnitudes, but model rankings remain relatively consistent (Fragility is the most affected metric).

On the Imagenet-1k IDSDS and FunnyBirds datasets, the Evidence metric produces patch/feature rankings that correlate with the ground-truth attribution as well as or better than rankings produced by other baselines (KernelSHAP is the only exception on Imagenet-1k). The number of steps in the reveal path controls a trade-off between processing speed and ranking correlation, saturated after 5 steps.

Critical strengths

None

Critical weaknesses

While the presentation of the work is good, it seems to be missing a purpose. What problem is the proposed decomposition solving? The presentation is generally good, but it does not seem to lead to a coherent conclusion. The Authors state that "DECAF assigns observable, endpoint-relative roles to paired responses" (L162), but the purpose of this assignment is missing. Most of the experiments show properties in well-controlled settings (synthetic data, subsampled datasets), but seem to miss out on any practical application or goal.

Questions for authors
  1. What problem is DECAF solving? What are its practical applications?

  2. [Sec. 6.1, in the part related to endpoint-null sensitivity] Since the Authors only retains the samples with $d < 0.02 = \epsilon $, shouldn't Fragility F and ordinary response magnitude be identical? Why do they produce completely different results?

Reviewer_vCLFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper analyzes the magnitude response of prediction changes while using perturbation methods. The authors introduces a framework DECAF(Decomposition of Evidence, Contradiction, And Fragility) that decompose the magnitude Abs = E+C+F where E corresponds to Evidence, C for contradiction and F for fragility. They demonstrate that the three components independently track measured behavior across controlled vision and tabular settings .

Critical strengths
  1. The main idea of the paper to decompose beyond magnitude is interesting and novel.

  2. The decomposition to E,C, F is intuitve and the authors provide theoretical justifications for the same.

  3. Strong validation of the main idea through empirical analysis, particularly section 6.2 is convincing.

Critical weaknesses
  1. The paper is difficult to follow, especially the descriptions in Sections 6 and 7, which are very dense. The paper's readability could improve substantially if Sections 6 and 7 present the experimental setup, evaluation protocol, and key findings more clearly and concisely.

  2. The fragility evaluation in Section 6.1 is described as being restricted to endpoint-null pairs satisfying |d| < 0.02. As per Section 3, DECAF sets evidence and contradiction to zero, implying that the ordinary absolute response Abs equals F. Thus, if Abs and F are computed on the same endpoint-null samples using the same trajectory aggregation, they should yield identical rankings and therefore identical AUROC values. However, Section 6.1 reports AUROC 0.878 for F but 0.433 for Abs. Please clarify why the observed AUROCs differ; otherwise, the result in Section 6.1 seems to contradict the definition of F in Section 3.

  3. The criteria |d| < 0.02 in Section 6.1 seems arbitrary. The authors should justify their choice of |d| < 0.02.

  4. In Section 6.3, the authors show that for the same endpoints, changing only the reveal path can shift the magnitude and decomposition a lot. Thus, DECAF is path-dependent, and the results in this section demonstrate that different reveal paths can produce substantially different values. The authors should add some guidelines for choosing a reveal path.

  5. DECAF relies on discrete reveal stages to approximate E,C, F. However, the paper does not report sensitivity of E, C, F to finer reveal stage counts. Additionally, the authors should justify the stage counts used in the experiments.

  6. In Section 7, Table 2 caption and lines 430:431, the authors should clarify why KernelSHAP is stated to have evaluation-specific information while the same is unavailable for the general baseline and DECAF.

  7. In Table 4, the authors report modest gains for DECAF on FunnyBirds, but on ImageNet-1K (for DECAF-3 and DECAF-5) its gains are negative. Since the confidence intervals don't include 0, hence these observations are unlikely to be caused by sampling variability. Therefore, the authors should clarify the negative gains for DECAF-3 and DECAF-5 for ImageNet-1K.

Questions for authors

Please address the weaknesses

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

Related papers

Discussion (0)

Sign in to comment.