pith. sign in

arxiv: 2606.18008 · v1 · pith:ZOBJ7HJTnew · submitted 2026-06-16 · 💻 cs.CV

PhaseWin: An Efficient Search Algorithm for Faithful Visual Attribution

Zihan Gu , Ruoyu Chen , Junchi Zhang , Li Liu , Xiaochun Cao , Hua Zhang This is my paper

Pith reviewed 2026-06-27 00:55 UTC · model grok-4.3

classification 💻 cs.CV
keywords visual attributionfaithful explanationgreedy searchmodel interpretabilityimage regionslinear complexityvision-language models
0
0 comments X

The pith

PhaseWin reorganizes greedy region selection for visual attribution into a phased window procedure that drops model evaluations from quadratic to linear while keeping near-greedy faithfulness.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper frames faithful visual attribution as an ordered subset search over image regions: the goal is to insert regions so the model response recovers as early as possible. Standard greedy search solves this but requires a quadratic number of model calls because every step re-scores the entire remaining set. PhaseWin replaces that with alternating global screening, adaptive pruning, and localized window refinement. Under monotone evidence accumulation and feature-level structural assumptions the procedure is proved to keep controllable linear cost together with faithfulness guarantees close to the greedy baseline. Experiments on classification, detection, grounding, and captioning tasks confirm the predicted drop to O(n) forward passes while matching or exceeding other attribution methods in faithfulness.

Core claim

PhaseWin attains controllable linear evaluation complexity together with near-greedy faithfulness guarantees by reorganizing greedy region selection into a phased window-search procedure that alternates global candidate screening, adaptive pruning, and localized window refinement.

What carries the argument

Phased window-search procedure that preserves the essential region-ranking behavior of greedy search without re-evaluating the full candidate set at every step.

If this is right

  • Attribution for high-resolution images or dense region partitions becomes practical at scale.
  • The same phased procedure can be applied to any ordered subset-search task whose scoring function satisfies the monotone accumulation property.
  • Model auditing pipelines can now afford exhaustive faithfulness checks on many more inputs without increasing compute budget.
  • Downstream tasks that rely on attribution heatmaps, such as debugging or regulatory reporting, inherit the reduced cost.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the structural assumptions hold for transformer-based vision models, PhaseWin could be inserted into existing explanation libraries with minimal code change.
  • The linear-complexity regime may allow attribution to be recomputed on-the-fly during interactive model inspection sessions.
  • Extensions to video or 3-D data would follow by treating spatio-temporal patches as the region set.

Load-bearing premise

The analysis assumes monotone evidence-accumulation conditions together with feature-level structural assumptions on the model responses.

What would settle it

An experiment that measures the number of forward passes required on images with n regions and checks whether faithfulness metrics fall below the greedy baseline once n exceeds a few hundred.

Figures

Figures reproduced from arXiv: 2606.18008 by Hua Zhang, Junchi Zhang, Li Liu, Ruoyu Chen, Xiaochun Cao, Zihan Gu.

Figure 1
Figure 1. Figure 1: Overview of PhaseWin for efficient high-faithfulness visual attribution. From left to right, an input image is partitioned into candidate regions, and a target response is obtained from a vision or vision–language model. Conventional greedy subset search repeatedly rescores all remaining candidates and therefore incurs quadratic evaluation cost. PhaseWin replaces exhaustive global rescoring with phased sub… view at source ↗
Figure 2
Figure 2. Figure 2: PhaseWin workflow. The algorithm alternates between (i) selecting a high-confidence anchor, (ii) pruning candidates via fixed-ratio thresholds, and (iii) performing windowed local refinement with dynamic supervision. Algorithm 1: PhaseWin: Phase-Window Accelerated Search Input: Ground set U, target size k, scoring function F(·), window size ω, window policy ψ Output: Ordered list Π 1 Π ← [ ]; S ← ∅; R ← U;… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on ImageNet classification attribution with CLIP ViT-L/14. Each row shows one target class, and each column compares one method. The overlays visualize the ranked superpixel regions, while the curves report the corresponding insertion and deletion trajectories. PhaseWin produces region orderings visually close to Greedy and achieves comparable insertion/deletion behavior, but require… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of correct object-level attribution cases on MS COCO, RefCOCO, and LVIS v1. Compared with ODAM and RISE, PhaseWin produces sharper and more faithful attributions. It matches or even exceeds Greedy in insertion AUC while requiring only a fraction of the computational budget. TABLE 11 Classification attribution on ImageNet misclassified samples with ResNet-101, using the ground-truth c… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on caption token attribution (Qwen2.5-VL-7B-Instruct). Each column shows one image-caption example, and each row compares one attribution method. The highlighted tokens define the attribution target, and the overlays show the ranked image regions supporting those tokens. PhaseWin closely follows Greedy in token-relevant visual evidence while using far fewer model forward evaluations … view at source ↗
Figure 6
Figure 6. Figure 6: Trade-off between speed and precision [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Insertion response curves under greedy subset search. Ground [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Linear-complexity verification under a fixed window configuration. [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Case study of an extreme Greedy-advantage sample on CLIP [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Case study of an extreme Greedy-advantage sample on CLIP [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Case study of an extreme Greedy-advantage sample on CLIP [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional classification attribution results on CLIP-RN101. From [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗
Figure 16
Figure 16. Figure 16: Additional classification attribution results on ResNet-101. From [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Additional classification attribution results on ResNet-101. From [PITH_FULL_IMAGE:figures/full_fig_p025_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Additional classification attribution results on CLIP ViT-L/14. [PITH_FULL_IMAGE:figures/full_fig_p025_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Additional classification attribution results on CLIP ViT-L/14. [PITH_FULL_IMAGE:figures/full_fig_p025_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Additional classification attribution results on CLIP ViT-L/14. [PITH_FULL_IMAGE:figures/full_fig_p026_20.png] view at source ↗
Figure 27
Figure 27. Figure 27: Additional image-captioning attribution results on Qwen2.5-VL [PITH_FULL_IMAGE:figures/full_fig_p026_27.png] view at source ↗
Figure 22
Figure 22. Figure 22: Additional failure-case visualizations on Grounding DINO for MS [PITH_FULL_IMAGE:figures/full_fig_p026_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Additional failure-case visualizations on Grounding DINO for [PITH_FULL_IMAGE:figures/full_fig_p026_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Additional failure-case visualizations on Grounding DINO for MS [PITH_FULL_IMAGE:figures/full_fig_p026_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Additional failure-case visualizations on Grounding DINO for [PITH_FULL_IMAGE:figures/full_fig_p026_25.png] view at source ↗
read the original abstract

Visual attribution is a fundamental tool for interpreting modern vision and vision-language models, particularly when their decisions must be inspected, diagnosed, or audited. Its goal is to explain how a model's decision depends on local regions of the visual input, typically by assigning an importance ordering over candidate image regions. Given an image partitioned into $n$ regions, faithful attribution can be cast as an ordered subset-search problem, in which progressively inserting the selected regions should recover the target model response as early as possible. Exhaustive search over region subsets incurs exponential cost, while the widely used greedy search still requires a quadratic number of model evaluations, because every selection step rescores all remaining candidates. We propose PhaseWin, an efficient subset-search algorithm for faithful visual attribution. PhaseWin reorganizes greedy region selection into a phased window-search procedure: rather than re-evaluating the full candidate set at every step, it alternates between global candidate screening, adaptive pruning, and localized window refinement, while preserving the essential region-ranking behavior of greedy search. We analyze PhaseWin under monotone evidence-accumulation conditions and show that, under feature-level structural assumptions, it attains controllable linear evaluation complexity together with near-greedy faithfulness guarantees. Extensive experiments on image classification, object detection, visual grounding, and image captioning show that, among all compared attribution methods, PhaseWin reaches high faithfulness with the fewest forward passes, empirically realizing the predicted reduction from $O(n^2)$ to $O(n)$. The code is available at https://github.com/Qihuai27/phasewin-va.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper introduces PhaseWin, a phased window-search algorithm that reorganizes greedy region selection for faithful visual attribution. Under monotone evidence-accumulation conditions and feature-level structural assumptions, the analysis claims controllable linear evaluation complexity together with near-greedy faithfulness guarantees. Experiments across image classification, object detection, visual grounding, and image captioning report that PhaseWin achieves high faithfulness with the fewest forward passes, empirically realizing an O(n) reduction from the quadratic cost of standard greedy search. Code is released.

Significance. If the stated assumptions hold for the evaluated models, the work supplies a theoretically grounded, scalable alternative to quadratic-cost greedy attribution while preserving ranking behavior. The combination of complexity analysis, near-greedy guarantees, and broad empirical coverage on four vision tasks would constitute a useful contribution to model-interpretation tooling.

major comments (1)
  1. [Analysis section] Analysis section: the linear-complexity and near-greedy faithfulness claims are derived under the monotone evidence-accumulation conditions and feature-level structural assumptions. The manuscript provides no verification (e.g., monotonicity plots or counter-example checks) that these conditions are satisfied by the ResNet, detection, grounding, or captioning models used in the experiments. Because the theoretical guarantees are invoked to explain the observed O(n) scaling, this verification step is load-bearing for the central claim that the empirical results realize the predicted theoretical behavior.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive comment on the analysis section. We address the point directly below.

read point-by-point responses

  1. Referee: Analysis section: the linear-complexity and near-greedy faithfulness claims are derived under the monotone evidence-accumulation conditions and feature-level structural assumptions. The manuscript provides no verification (e.g., monotonicity plots or counter-example checks) that these conditions are satisfied by the ResNet, detection, grounding, or captioning models used in the experiments. Because the theoretical guarantees are invoked to explain the observed O(n) scaling, this verification step is load-bearing for the central claim that the empirical results realize the predicted theoretical behavior.

    Authors: We agree that the manuscript does not contain explicit verification (such as monotonicity plots or counter-example checks) that the monotone evidence-accumulation conditions and feature-level structural assumptions hold for the ResNet, detection, grounding, and captioning models used in the experiments. The theoretical claims are stated conditionally on these assumptions, while the reported O(n) scaling is an empirical observation. To strengthen the link between the analysis and the experimental results, the revised manuscript will include an appendix with monotonicity verification and any relevant counter-example analysis for the evaluated models. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation self-contained under explicit assumptions

full rationale

The paper presents PhaseWin as a reorganization of greedy subset search into phased window screening and refinement. It states theoretical guarantees explicitly under the monotone evidence-accumulation conditions and feature-level structural assumptions, without deriving those assumptions from the target faithfulness metric or from any fitted parameter. No equations reduce a claimed prediction to a fitted input by construction, and no load-bearing step relies on a self-citation chain whose content is itself unverified within the paper. Empirical results on ResNet, detection, grounding, and captioning tasks are reported separately as validation of the O(n) complexity, independent of the analytic bounds. The derivation chain therefore remains non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim depends on two domain assumptions for its theoretical guarantees; no free parameters or invented entities are mentioned in the abstract.

axioms (2)
  • domain assumption monotone evidence-accumulation conditions
    Invoked to prove controllable linear complexity and near-greedy faithfulness.
  • domain assumption feature-level structural assumptions
    Required for the windowed search to preserve the essential region-ranking behavior.

pith-pipeline@v0.9.1-grok · 5823 in / 1306 out tokens · 32415 ms · 2026-06-27T00:55:24.116848+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

66 extracted references · 9 canonical work pages · 6 internal anchors

  1. [1]

    Attribution explanations for deep neural networks: A theoretical perspective,

    H. Deng, H. Pei, Q. Zhang, and M. Du, “Attribution explanations for deep neural networks: A theoretical perspective,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  2. [2]

    Explainable AI (XAI): Core ideas, techniques, and solutions,

    R. Dwivedi, D. Dave, H. Naik, S. Singhal, R. Omer, P. Patel, B. Qian, Z. Wen, T. Shah, G. Morganet al., “Explainable AI (XAI): Core ideas, techniques, and solutions,”ACM Comput. Surv., vol. 55, no. 9, pp. 1–33, 2023

    2023

  3. [3]

    A survey of feature attribution techniques in explainable ai: Taxonomy, analysis and comparison,

    M. A. Nazir, E. Evangelista, S. M. S. Bukhari, and R. Sharma, “A survey of feature attribution techniques in explainable ai: Taxonomy, analysis and comparison,”Annals of Mathematics and Computer Science, vol. 28, pp. 115–126, 2025

    2025

  4. [4]

    A review and comparative study on probabilistic object detection in autonomous driv- ing,

    D. Feng, A. Harakeh, S. L. Waslander, and K. Dietmayer, “A review and comparative study on probabilistic object detection in autonomous driv- ing,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 8, pp. 9961–9980, 2021

    2021

  5. [5]

    Going beyond xai: A systematic survey for explanation-guided learning,

    Y . Gao, S. Gu, J. Jiang, S. R. Hong, D. Yu, and L. Zhao, “Going beyond xai: A systematic survey for explanation-guided learning,”ACM Computing Surveys, vol. 56, no. 7, pp. 1–39, 2024

    2024

  6. [6]

    Can Attribution Predict Risk? From Multi-View Attribution to Planning Risk Signals in End-to-End Autonomous Driving

    L. Yang, R. Chen, H. Liu, J. Liang, S. Sun, and X. Cao, “Can attribution predict risk? from multi-view attribution to planning risk signals in end- to-end autonomous driving,”arXiv preprint arXiv:2605.06264, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  7. [7]

    Where Not to Learn: Prior-Aligned Training with Subset-based Attribution Constraints for Reliable Decision-Making

    R. Chen, S. Sun, X. Guo, S. Zhang, K. Liu, S. Liu, Z. Wang, Q. Zhang, H. Zhang, and X. Cao, “Where not to learn: Prior-aligned training with subset-based attribution constraints for reliable decision-making,”arXiv preprint arXiv:2602.07008, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  8. [8]

    Safe: Sensitivity-aware features for out-of-distribution object detection,

    S. Wilson, T. Fischer, F. Dayoub, D. Miller, and N. Sünderhauf, “Safe: Sensitivity-aware features for out-of-distribution object detection,” in ICCV, 2023, pp. 23 565–23 576

    2023

  9. [9]

    Safemobile: Chain-level jailbreak detection and automated evaluation for multimodal mobile agents,

    S. Liang, T. Fang, Z. Liu, A. Liu, Y . Xiao, J. He, E.-C. Chang, and X. Cao, “Safemobile: Chain-level jailbreak detection and automated evaluation for multimodal mobile agents,”arXiv preprint arXiv:2507.00841, 2025

    work page arXiv 2025

  10. [10]

    RISE: Randomized input sampling for explanation of black-box models,

    V . Petsiuk, A. Das, and K. Saenko, “RISE: Randomized input sampling for explanation of black-box models,” inBMVC, 2018, p. 151

    2018

  11. [11]

    Black-box explanation of object detectors via saliency maps,

    V . Petsiuk, R. Jain, V . Manjunatha, V . I. Morariu, A. Mehra, V . Ordonez, and K. Saenko, “Black-box explanation of object detectors via saliency maps,” inCVPR, 2021, pp. 11 443–11 452

    2021

  12. [12]

    Less is More: Efficient Black-box Attribution via Minimal Interpretable Subset Selection

    R. Chen, S. Liang, J. Li, S. Liu, L. Liu, H. Zhang, and X. Cao, “Less is more: Efficient black-box attribution via minimal interpretable subset selection,”arXiv preprint arXiv:2504.00470, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    Spatial sensitive grad-cam: Visual explanations for object detection by incorporating spatial sensitivity,

    T. Yamauchi and M. Ishikawa, “Spatial sensitive grad-cam: Visual explanations for object detection by incorporating spatial sensitivity,” in ICIP, 2022, pp. 256–260

    2022

  14. [14]

    Spatial Sensitive Grad-CAM++: Improved visual expla- nation for object detectors via weighted combination of gradient map,

    T. Yamauchi, “Spatial Sensitive Grad-CAM++: Improved visual expla- nation for object detectors via weighted combination of gradient map,” inCVPR Workshop, 2024, pp. 8164–8168

    2024

  15. [15]

    Less is more: Fewer interpretable region via submodular subset selection,

    R. Chen, H. Zhang, S. Liang, J. Li, and X. Cao, “Less is more: Fewer interpretable region via submodular subset selection,” inICLR, 2024

    2024

  16. [16]

    Interpreting Object-level Foundation Models via Visual Pre- cision Search,

    R. Chen, S. Liang, J. Li, S. Liu, M. Li, Z. Huang, H. Zhang, and X. Cao, “Interpreting Object-level Foundation Models via Visual Pre- cision Search,” inCVPR, 2025

    2025

  17. [17]

    Explain any concept: Seg- ment anything meets concept-based explanation,

    A. Sun, P. Ma, Y . Yuan, and S. Wang, “Explain any concept: Seg- ment anything meets concept-based explanation,” inNeurIPS, 2023, pp. 21 826–21 840

    2023

  18. [18]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” inCVPR, 2009, pp. 248–255

    2009

  19. [19]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inICML, 2021, pp. 8748–8763

    2021

  20. [20]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778

    2016

  21. [21]

    Microsoft COCO: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick, “Microsoft COCO: Common objects in context,” inECCV, 2014, pp. 740–755

    2014

  22. [22]

    LVIS: A dataset for large vocabulary instance segmentation,

    A. Gupta, P. Dollar, and R. Girshick, “LVIS: A dataset for large vocabulary instance segmentation,” inCVPR, 2019, pp. 5356–5364

    2019

  23. [23]

    ReferItGame: Referring to objects in photographs of natural scenes,

    S. Kazemzadeh, V . Ordonez, M. Matten, and T. Berg, “ReferItGame: Referring to objects in photographs of natural scenes,” inEMNLP, 2014, pp. 787–798

    2014

  24. [24]

    Grounding DINO: Marrying dino with grounded pre- training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, C. Li, J. Yang, H. Su, J. Zhuet al., “Grounding DINO: Marrying dino with grounded pre- training for open-set object detection,” inECCV, 2024

    2024

  25. [25]

    Florence-2: Advancing a unified representation for a variety of vision tasks,

    B. Xiao, H. Wu, W. Xu, X. Dai, H. Hu, Y . Lu, M. Zeng, C. Liu, and L. Yuan, “Florence-2: Advancing a unified representation for a variety of vision tasks,” inCVPR, 2024, pp. 4818–4829

    2024

  26. [26]

    Qwen2.5-VL Technical Report

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tanget al., “Qwen2.5-vl technical report,”arXiv preprint arXiv:2502.13923, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    Phasewin search framework enable efficient object-level interpretation,

    Z. Gu, R. Chen, J. Zhang, Y . Hu, H. Zhang, and X. Cao, “Phasewin search framework enable efficient object-level interpretation,” inCVPR, 2026

    2026

  28. [28]

    On pixel-wise explanations for non-linear classifier decisions 16 by layer-wise relevance propagation,

    S. Bach, A. Binder, G. Montavon, F. Klauschen, K.-R. Müller, and W. Samek, “On pixel-wise explanations for non-linear classifier decisions 16 by layer-wise relevance propagation,”PloS one, vol. 10, no. 7, p. e0130140, 2015

    2015

  29. [29]

    Gradient methods for submodular maximization,

    H. Hassani, M. Soltanolkotabi, and A. Karbasi, “Gradient methods for submodular maximization,”Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  30. [30]

    Grad-CAM: visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: visual explanations from deep networks via gradient-based localization,”International Journal of Computer Vision, vol. 128, pp. 336–359, 2020

    2020

  31. [31]

    Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks,

    A. Chattopadhay, A. Sarkar, P. Howlader, and V . N. Balasubramanian, “Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks,” inWACV, 2018, pp. 839–847

    2018

  32. [32]

    Score-cam: Score-weighted visual explanations for convolu- tional neural networks,

    H. Wang, Z. Wang, M. Du, F. Yang, Z. Zhang, S. Ding, P. Mardziel, and X. Hu, “Score-cam: Score-weighted visual explanations for convolu- tional neural networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 24–25

    2020

  33. [33]

    Vit-cx: Causal explanation of vision transformers,

    W. Xie, X. Li, C. C. Cao, and N. L. Zhang, “Vit-cx: Causal explanation of vision transformers,” inIJCAI, 2023

    2023

  34. [34]

    Gradient- based visual explanation for transformer-based clip,

    C. Zhao, K. Wang, X. Zeng, R. Zhao, and A. B. Chan, “Gradient- based visual explanation for transformer-based clip,” inICML, 2024, pp. 61 072–61 091

    2024

  35. [35]

    Axiomatic attribution for deep networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic attribution for deep networks,” inICML, 2017, pp. 3319–3328

    2017

  36. [36]

    iGOS++: integrated gradient optimized saliency by bilateral perturbations,

    S. Khorram, T. Lawson, and L. Fuxin, “iGOS++: integrated gradient optimized saliency by bilateral perturbations,” inProceedings of the Conference on Health, Inference, and Learning, 2021, pp. 174–182

    2021

  37. [37]

    Making sense of dependence: Ef- ficient black-box explanations using dependence measure,

    P. Novello, T. Fel, and D. Vigouroux, “Making sense of dependence: Ef- ficient black-box explanations using dependence measure,” inNeurIPS, 2022, pp. 4344–4357

    2022

  38. [38]

    "why should I trust you?

    M. T. Ribeiro, S. Singh, and C. Guestrin, “"why should I trust you?": Ex- plaining the predictions of any classifier,” inACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016

    2016

  39. [39]

    A value for n-person games,

    L. S. Shapley, “A value for n-person games,”Contribution to the Theory of Games, vol. 2, 1953

    1953

  40. [40]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” inNeurIPS, 2017, pp. 4765–4774

    2017

  41. [41]

    Harsanyinet: Computing accurate shapley values in a single forward propagation,

    L. Chen, S. Lou, K. Zhang, J. Huang, and Q. Zhang, “Harsanyinet: Computing accurate shapley values in a single forward propagation,” inICML, 2023

    2023

  42. [42]

    Problems with shapley-value-based explanations as feature importance measures,

    I. E. Kumar, S. Venkatasubramanian, C. Scheidegger, and S. A. Friedler, “Problems with shapley-value-based explanations as feature importance measures,” inICML, 2020

    2020

  43. [43]

    One explanation is not enough: structured attention graphs for image classification,

    V . Shitole, F. Li, M. Kahng, P. Tadepalli, and A. Fern, “One explanation is not enough: structured attention graphs for image classification,” in NeurIPS, 2021, pp. 11 352–11 363

    2021

  44. [44]

    Explain to Fix: A Framework to Interpret and Correct DNN Object Detector Predictions

    D. Gudovskiy, A. Hodgkinson, T. Yamaguchi, Y . Ishii, and S. Tsukizawa, “Explain to fix: A framework to interpret and correct dnn object detector predictions,”arXiv preprint arXiv:1811.08011, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  45. [45]

    Gradient-based instance-specific visual explanations for object specification and object discrimination,

    C. Zhao, J. H. Hsiao, and A. B. Chan, “Gradient-based instance-specific visual explanations for object specification and object discrimination,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  46. [46]

    Diverse explanations for object detectors with nesterov-accelerated igos++

    M. Jiang, S. Khorram, and F. Li, “Diverse explanations for object detectors with nesterov-accelerated igos++.” inBMVC, 2023, pp. 188– 189

    2023

  47. [47]

    Comparing the decision-making mechanisms by transformers and cnns via explanation methods,

    M. Jiang, S. Khorram, and L. Fuxin, “Comparing the decision-making mechanisms by transformers and cnns via explanation methods,” in CVPR, 2024, pp. 9546–9555

    2024

  48. [48]

    Interpreting clip’s image representation via text-based decomposition,

    Y . Gandelsman, A. A. Efros, and J. Steinhardt, “Interpreting clip’s image representation via text-based decomposition,” inICLR, 2024

    2024

  49. [49]

    Explaining Object Detectors via Collective Contribution of Pixels

    T. Yamauchi, H. Kera, and K. Kawamoto, “Explaining object detectors via collective contribution of pixels,”arXiv preprint arXiv:2412.00666, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  50. [50]

    Lvlm- intrepret: An interpretability tool for large vision-language models,

    G. Ben Melech Stan, E. Aflalo, R. Y . Rohekar, A. Bhiwandiwalla, S.-Y . Tseng, M. L. Olson, Y . Gurwicz, C. Wu, N. Duan, and V . Lal, “Lvlm- intrepret: An interpretability tool for large vision-language models,” in IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) Workshops, 2024, pp. 8182–8187

    2024

  51. [51]

    Smooth Grad-CAM++: An enhanced inference level visualization technique for deep convolutional neural network models,

    D. Omeiza, S. Speakman, C. Cintas, and K. Weldermariam, “Smooth Grad-CAM++: An enhanced inference level visualization technique for deep convolutional neural network models,”arXiv preprint arXiv:1908.01224, 2019

    work page arXiv 1908

  52. [52]

    From redundancy to relevance: Enhancing explainability in multimodal large language models,

    X. Zhang, Y . Quan, C. Shen, X. Yuan, S. Yan, L. Xie, W. Wang, C. Gu, H. Tang, and J. Ye, “From redundancy to relevance: Enhancing explainability in multimodal large language models,” inNAACL, 2025

    2025

  53. [53]

    Where do large vision-language models look at when answering questions?

    X. Xing, C.-W. Kuo, L. Fuxin, Y . Niu, F. Chen, M. Li, Y . Wu, L. Wen, and S. Zhu, “Where do large vision-language models look at when answering questions?”arXiv preprint arXiv:2503.13891, 2025

    work page arXiv 2025

  54. [54]

    Token activation map to visually explain multimodal llms,

    Y . Li, H. Wang, X. Ding, H. Wang, and X. Li, “Token activation map to visually explain multimodal llms,” inICCV, 2025

    2025

  55. [55]

    Mllms know where to look: Training-free perception of small visual details with multimodal llms,

    J. Zhang, M. Khayatkhoei, P. Chhikara, and F. Ilievski, “Mllms know where to look: Training-free perception of small visual details with multimodal llms,” inICLR, 2025

    2025

  56. [56]

    Where mllms attend and what they rely on: Explaining autoregressive token generation,

    R. Chen, X. Guo, K. Liu, S. Liang, S. Liu, Q. Zhang, L. Wang, H. Zhang, and X. Cao, “Where mllms attend and what they rely on: Explaining autoregressive token generation,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2026

    2026

  57. [57]

    Accelerated greedy algorithms for maximizing submodular set functions,

    M. Minoux, “Accelerated greedy algorithms for maximizing submodular set functions,”INFOR: Information Systems and Operational Research, vol. 14, no. 3, pp. 247–255, 1978

    1978

  58. [58]

    Cost-effective outbreak detection in networks,

    J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. VanBriesen, and N. Glance, “Cost-effective outbreak detection in networks,” inProceed- ings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2007, pp. 420–429

    2007

  59. [59]

    Lazier than lazy greedy,

    B. Mirzasoleiman, A. Badanidiyuru, A. Karbasi, J. V ondrák, and A. Krause, “Lazier than lazy greedy,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 29, no. 1, 2015

    2015

  60. [60]

    Fast multi-stage submodular maximiza- tion,

    K. Wei, R. Iyer, and J. Bilmes, “Fast multi-stage submodular maximiza- tion,” inInternational conference on machine learning. PMLR, 2014, pp. 1494–1502

    2014

  61. [61]

    The fast algorithm for submod- ular maximization,

    A. Breuer, E. Balkanski, and Y . Singer, “The fast algorithm for submod- ular maximization,” inInternational Conference on Machine Learning. PMLR, 2020, pp. 1134–1143

    2020

  62. [62]

    Deep inside convolutional networks: Visualising image classification models and saliency maps,

    K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep inside convolutional networks: Visualising image classification models and saliency maps,” in ICLR Workshop, 2014

    2014

  63. [63]

    SLIC superpixels compared to state-of-the-art superpixel methods,

    R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. Süsstrunk, “SLIC superpixels compared to state-of-the-art superpixel methods,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 34, no. 11, pp. 2274–2282, 2012

    2012

  64. [64]

    Look at the variance! efficient black-box explanations with sobol- based sensitivity analysis,

    T. Fel, R. Cadene, M. Chalvidal, M. Cord, D. Vigouroux, and T. Serre, “Look at the variance! efficient black-box explanations with sobol- based sensitivity analysis,” inAdvances in Neural Information Processing Systems, vol. 34, 2021, pp. 26 005–26 014

    2021

  65. [65]

    Top- down neural attention by excitation backprop,

    J. Zhang, S. A. Bargal, Z. Lin, J. Brandt, X. Shen, and S. Sclaroff, “Top- down neural attention by excitation backprop,”International Journal of Computer Vision, vol. 126, no. 10, pp. 1084–1102, 2018. APPENDIXA SUBMODULARITY ANDSUPERMODULARITY This appendix clarifies why the score optimized by greedy-based visual attribution should not be regarded as a g...

    2018

  66. [66]

    Therefore λ2G(S) +b 2 ≥F(S)≥c F(T)−η≥c λ1G(T) +b 1 −η

    IfS, T⊆Uandc, η≥0satisfy F(S)≥c F(T)−η,(50) then G(S)≥ cλ1 λ2 G(T) + cb1 −η−b 2 λ2 .(51) Proof.The upper side of Assumption 3 givesF(S)≤λ 2G(S)+ b2, while the lower side givesF(T)≥λ 1G(T) +b 1. Therefore λ2G(S) +b 2 ≥F(S)≥c F(T)−η≥c λ1G(T) +b 1 −η. (52) Movingb 2 to the right-hand side and dividing byλ 2 proves the claim. Proof of Theorem 1.Let the firstk...