acceptodds
Withdrawn from ICLR 2027

A Good Initialization is All You Need for Faithful Visual Attribution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Faithful visual attribution, which aims to identify which image regions genuinely support a model's prediction, is a central challenge in interpretability research. Search-based perturbation methods define the insertion–deletion faithfulness Pareto frontier by masking regions and measuring score changes, yet they must construct a complete ordering of all regions, incurring substantial cost when only a compact top-\(k\) evidence mask is needed. We study this mask-first problem from two angles. An analysis of the behavioral gap between greedy and phase-window based search yields CoPAIR, a coarse pairwise initializer that already improves search. A deeper rethinking of the problem leads to TRACE (Top-\(k\) Region Attribution via Cross-Entropy), which directly optimizes a binary evidence mask via cross-entropy at a cost comparable to gradient-based methods. Both can serve as compact attribution masks or warm-start search methods when a complete ranking is required. Across ImageNet classification with CLIP ViT-L/14, CLIP RN101, and ResNet-101, our initialized search methods establish a new state-of-the-art frontier for faithful full-ordering attribution under inclusive forward-call accounting. On POPE and RePOPE with Qwen2.5-VL-3B-Instruct and LLaVA-v1.5-7B, TRACE+Greedy gives the strongest search-based MLLM attribution results. Direct TRACE masks further achieve single-point RePOPE repair rates of \(94.44%\) and \(96.00%\), showing that compact evidence masks can be actionable attribution outputs, not merely prefixes of full rankings.

Primary area:
interpretability and explainable AI
Keywords:
Visual Attribution, Black-box Model Interpretability, XAI.

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_3gjuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies visual attribution from a mask-first perspective: instead of constructing a complete ranking of image regions, it first searches for a compact set of regions that gives a high target response. The authors introduce CoPAIR, which searches over coarse spatial groups and their pairs, and TRACE, which samples fixed-size region subsets and updates its sampling distribution using high-scoring candidates. The resulting mask can be used directly or passed to Greedy or PhaseWin to construct a full ordering. The experiments cover three classification backbones and two MLLMs. Initialization often improves insertion performance while reducing the total number of selection-time forward calls. TRACE also achieves high single-mask repair rates on RePOPE, offering a useful trade-off between repair success and search cost.

Critical strengths

The initialization strategy produces useful gains with inclusive cost accounting. On CLIP ViT-L/14 Correct, TRACE+Greedy improves insertion AUC from 0.8431 to 0.8679 while reducing selection cost from 4.08k to 3.69k forward calls. Similar improvements appear in several other settings. I appreciate that the reported cost includes initialization rather than treating it as free. This makes the practical benefit more convincing.

The direct-mask results offer a meaningful efficiency trade-off. TRACE repairs 255/270 Qwen cases and 288/300 LLaVA cases using one final mask. Table 17 reports roughly 164–165 selection calls, compared with thousands for full search. Full-curve search achieves higher repair rates, so TRACE is not the strongest method in absolute repair success, but retaining most of that success at a much smaller selection budget is a worthwhile result.

Critical weaknesses

The experiments do not fully isolate the benefit of adaptation. Table 8 tests random initializers, but not best-of-160 uniform mask search. The one-round ablation in Table 9 also uses fewer samples, so budget and adaptation are confounded. A matched-budget comparison and repeated TRACE runs would strengthen the evidence. The broader faithfulness claim also needs qualification: on CLIP ViT-L/14 Correct, TRACE+Greedy improves insertion but worsens deletion AUC from 0.1173 to 0.1342.

The recovery analysis relies on favorable assumptions about the update. Equations (38) and (41) assume basin-mass amplification and balance without deriving either from Equation (17). Equation (22) also multiplies round-wise miss probabilities despite their dependence on earlier random samples; this needs proper conditioning or bounds that hold across histories. These issues look fixable, but the current analysis does not establish that TRACE’s actual update improves recovery probability.

Questions for authors

With the same budget of 160 candidate evaluations and the same mask size, how does TRACE compare with uniform random search that returns its best sampled mask? Please use the same ordering and release procedure for the initialization comparison, and report variability across TRACE seeds. This would directly address whether adaptation is responsible for the gain.

Can you derive the amplification and balance conditions in Equations (38) and (41) from Equation (17), or explicitly limit the result to a conditional analysis? How should Equation (22) be reformulated to account for the adaptive sampling history?

For RePOPE repair, what exact answer token is optimized, and how is it selected before mask search? Is the corrected answer supplied to the optimizer? I am not assuming that it is, but this distinction is necessary to interpret the result as either target-conditioned repair or a correction procedure that can operate without the correct answer.

Could you clarify how the main search backends balance sufficiency and necessity, and whether the classification deletion losses mainly come from the sufficiency-only initializer? A comparison using the combined objective already defined in Equation (2) would help establish whether this is an adjustable trade-off.

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

This paper investigates how initialization can reduce the cost of finding compact image regions that support model responses. It proposes CoPAIR, which searches over coarse spatial groups and their pairs, and TRACE, which adapts a Gumbel-top-k sampling distribution using high-scoring masks. TRACE's fixed-size masks can be used directly or to initialize Greedy or PhaseWin.

Critical strengths
  1. The paper addresses a relevant bottleneck in perturbation-based attribution: the cost of searching for informative regions. CoPAIR and TRACE can be integrated with Greedy and PhaseWin, giving the proposal a clear experimental use case.

  2. The evaluation spans three classifiers and two multimodal models, with reported costs that include initialization. Several settings show higher insertion scores with fewer total forward calls than uninitialized Greedy.

Critical weaknesses
  1. The evidence for TRACE's added methodological value over straightforward sampling for initialization is incomplete. Table 8 tests random initial masks and Table 9 varies the sample budget, but neither provides a random-search control with the same candidate budget. Without this comparison, the reported gains cannot be attributed specifically to adaptive updates.

  2. The multi-round probability calculation in Equation (22) is not justified as written. For adaptive sampling, each round's failure probability must be conditioned on failure in all preceding rounds. Multiplying unconditional round-wise probabilities does not establish the claimed overall recovery probability.

  3. The conditional recovery bounds do not establish a guarantee for the implemented TRACE update. Although the paper states the basin-amplification and balanced-block-weight assumptions in Equations (38) and (41), it does not show that Equation (17) satisfies them. A separate argument is needed to connect these assumptions to the actual update; otherwise, the recovery claims must remain explicitly conditional.

  4. The faithfulness evidence is mixed because insertion gains can accompany worse deletion performance. In Table 1, TRACE+G improves Correct Ins from Greedy's 0.8431 to 0.8679, while Del rises from 0.1173 to 0.1342, where lower is better. This trade-off limits how strongly the insertion gains can support claims of improved overall faithfulness.

  5. The repair protocol is not specified well enough to assess the practical scope of the reported success rates. The paper states that its repair protocol follows EAGLE, but does not fully specify the inherited conditions. The target prompt, masking setup, and success criterion need to be explicit before these rates can be interpreted as evidence for practical repair. Whether the evaluation assumes known error cases and correction targets also needs explicit clarification.

  6. The claim of low run-to-run variance is not substantiated by the reported seed settings. Table 7 specifies one TRACE seed, whereas the three seeds in Table 8 belong to the random-mask baseline. Without repeated TRACE runs or uncertainty estimates, the reliability of small performance differences remains unresolved.

Questions for authors
  1. Could you compare TRACE with the best of 160 uniformly sampled fixed-size masks, keeping the mask size, scoring, caching, and subsequent search identical? How much of the gain remains when the candidate budgets are matched? Please also report variation across multiple TRACE seeds on representative settings to assess the reliability of small gains.

  2. Could you correct Equation (22) using failure probabilities conditioned on failure in all preceding rounds? What properties of Equation (17) ensure the assumptions in Equations (38) and (41)? Please explain whether the corrected recovery claim holds for the implemented update or remains conditional.

  3. How do you reconcile the insertion gain in Table 1 with the worse deletion score? Which claim about faithfulness remains supported when both metrics are considered?

  4. For TRACE Direct repair, please specify the target answer, prompt, baseline image, and success criterion. Does the intervention follow Equation (1), and which conditions are inherited from EAGLE? Please clarify whether the evaluation assumes known error cases and correction targets.

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

The paper proposes a new method to find the top k regions relevant for a prediction, which can serve as a visual attribution method. The focus is on search-based perturbation algorithms, which can be very expensive. I.e., the goal is to find the right order of regions aligned with increasing target probability when adding regions in the found order. The paper discusses two algorithms, both of which can be used as an effective and efficient initialization of a search-based algorithm, while one of them (TRACE) can also be used to directly generate a visual explanation with the top-k regions. They lead to better faithfulness metrics while using less computation, thanks to the improved initialization compared to the original method. They have been validated with the deletion/insertion protocol on various datasets and for various backbones.

Critical strengths

The paper evaluates multiple backbones, including CNN, ViT, and MLLMs, showing the method's general applicability across backbones, which is useful.

Critical weaknesses

The paper builds on the assumption that the insertion/deletion protocol allows for evaluating true faithfulness. However, the provided experiments do not really allow for a thorough conclusion:

  • Insertion/deletion protocols are only a proxy measurement of faithfulness and are known to have many limitations ([a,b,c,d]), which affect the reliability of the ranking. Specifically, they do not only measure whether the attribution is faithfull but heavily depend on how the model reacts to the interventions, which lead to OOD data. These limitations should be mentioned. Using the protocol by itself is only a limited test of faithfulness. Better would be to run evaluations with multiple baselines for the background image, adding a random-ranking baseline as a sanity check and also adding other evaluations [d,e]. If the reported ranking is consistent among them that would give much stronger evidence about the claimed improved faithfullness.
  • It is surprising that gradient is performing better than integrated gradient (in this paper called IG2). Is there an explanation for that? To better understand these results, further information about how integrated gradient has been applied is needed, e.g., how many evaluation calls are used, which baseline image is used? How have they been chosen? Is the baseline image the same as used for the insertion/deletion protocol? Related to the previous comment, conclusive conclusions from a deletion/insertion protocol about faithfulness are very fragile, as ranking depends on many other choices than just the attribution method itself. Therefore, the presented evaluations do not fully provide convincing evidence for the made claims.

[a] Nieradzik et al., Reliable Evaluation of Attribution Maps in CNNs: A Perturbation-Based Approach, IJCV 2024

[b] Hesse et al., Benchmarking the Attribution Quality of Vision Models, NeurIPS 2024.

[c] Hooker et al., A Benchmark for Interpretability Methods in Deep Neural Networks, NeurIPS 2019

[d] Rong et al., "A Consistent and Efficient Evaluation Strategy for Attribution Methods.", ICML, 2022

[e] Hesse et al., FunnyBirds: A Synthetic Vision Dataset for a Part-Based Analysis of Explainable AI Methods, ICCV 2023

Minor:

Term TRACE is already introduced in XAI for an evaluation framework of the deletion metric [f].

[f] Wang and Wang, Benchmarking Deletion Metrics with the Principled Explanations, ICML 2024.

Questions for authors
  • Is the ranking consistent across multiple variations of the deletion/insertion protocol (e.g., for different baselines) and other evaluation protocols for faithfulness?

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

Related papers

Discussion (0)

Sign in to comment.