REVIEW 5 major objections 5 minor 1 cited by
PnPXAI: A Universal XAI Framework Providing Automatic Explanations Across Diverse Modalities and Models
T0 review · 5 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read PnPXAI claims to auto-detect a model's architecture and return tuned explanations with no user configuration.
desk verdict PnPXAI is a useful integration with real engineering and a good liver-tumor ground-truth example, but relying on an undefined fairness-paper metric (ABPC) for 'optimal' explanations is a load-bearing gap. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central mechanism is the recommender's mapping table, which declares which explanation methods are applicable to each combination of data modality (vision, language, structured data, time series) and architecture family (linear, convolutional, recurrent, transformer, decision tree). The detector produces the architecture side of the key by tracing a symbolic graph of the user's model, including residual connections. The evaluator scores candidate explanations with three quantified properties (correctness, continuity, compactness), and the optimizer grid-searches hyperparameters to maximize the default correctness score. The claim of universality rests on this table: any model that can be classified into a row yields a recommended method list automatically.
What would settle it
Take a model with known ground-truth attributions, run every explanation method the framework does and does not recommend, and check whether the recommended set contains the best-scoring method under the framework's own metric; a case where an excluded method beats all recommended ones would falsify the recommender. A second check: if tuning hyperparameters to maximize the internal score decreases agreement with ground-truth masks, the optimizer's objective is not tracking faithfulness.
Extended reading notes
Core claim
The paper's central claim is that a single framework can automatically do what currently requires an XAI specialist: identify the structure of an arbitrary neural network, decide which explanation methods can run on it, and optimize each method's hyperparameters on the user's own data so the resulting attribution maps are as faithful as possible. The claim is realized as AutoExplanation, which takes a model and dataloader, runs a detector that builds a symbolic graph of layer types and residual connections, consults a recommender that intersects the detected data modality with the detected architecture family, and hands the shortlisted explainers to an evaluator and grid-search optimizer. The default optimization objective is a correctness score called ABPC, chosen because it measures faithfulness without needing ground-truth masks. Validity is argued through use cases where optimized heatmaps align with liver-tumor segmentation masks and known clinical biomarkers, and through a survey in which participants rated the automatic recommendation as the most important feature.
Load-bearing premise
The framework's promise of 'optimal' explanations rests on an unvalidated assumption: that its hand-made table of which explainers work on which models is correct, and that its default score really measures explanation faithfulness.
Editorial extensions
If this is right
- A user with no XAI background can obtain attribution maps for vision, language, tabular, and time-series models with a few lines of code.
- Model-specific explainers (gradient-based, CAM-based, relevance propagation, attention-based) become usable alongside model-agnostic ones, widening the pool of candidate explanations.
- Because applicability is defined by the mapping table, adding a new explanation method only requires registering its supported modalities and architectures in that table.
- Hyperparameter choices that visibly alter attribution quality, such as the superpixel mask for perturbation-based methods, are tuned automatically rather than left to user intuition.
- Quantitative evaluation is integrated into the pipeline, so each recommended explanation comes with scores for correctness, continuity, and compactness rather than an unjudged heatmap.
Reading between the lines
- The mapping-table design makes the framework a natural benchmark harness: if the table were extended with new explainers, the same pipeline could compare any candidate method against the recommended set on the same models and data.
- The authors' stated future direction, applying the pipeline to generative language models, is a direct test of the universality claim because attention-only and decoder architectures would need new rows in the table.
- A community-maintained mapping table could drift from the actual applicability of explainers as new architectures appear, so the framework's usefulness will depend on regular empirical revalidation of each row.
- The user survey measures satisfaction, not explanation quality, so the strongest evidence for the framework's core claim comes from the medical and financial ground-truth case studies; extending those to more domains would strengthen the universal claim.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents PnPXAI, an open-source explainable AI (XAI) framework that claims to be universal across data modalities and neural network architectures. The framework comprises five modules: a detector that builds a symbolic graph of a user-provided model, a recommender that filters applicable explanation methods using a hand-authored mapping table, an explainer pool, an evaluator that scores explanations with quantitative metrics, and a hyperparameter optimizer that tunes explanation methods via grid search. The central claim is that a few lines of code produce 'optimal explanations' without user expertise. Validation consists of a 31-participant user survey, a liver tumor detection use case with CT ground-truth masks, an acute kidney injury (AKI) detection use case with known biomarkers on MIMIC-III, and a bank account fraud detection interface.
Significance. If the central claims were fully supported, PnPXAI would be a useful contribution: it lowers the barrier for non-expert practitioners to apply and tune explanation methods, and the modular design with public code, API documentation, and a demo is a genuine practical asset. The liver tumor and AKI use cases provide positive evidence that hyperparameter optimization improves ground-truth relevance accuracy and that recommended explainers identify expected biomarkers. However, the strongest claim—that PnPXAI produces 'optimal explanations'—currently rests on an undefined default objective (ABPC) and on a hand-authored recommender mapping that is not validated. The absence of comparisons against existing frameworks such as Captum, OmniXAI, AutoXAI, and OpenXAI also makes the 'universal' claim unsubstantiated. These issues are fixable in a revision, but they are load-bearing for the paper's main contribution.
major comments (5)
- [Evaluator / Liver Tumor Detection] The paper states in the Liver Tumor Detection use case that 'ABPC (Han et al. 2023) is chosen as the default objective with the evaluation property of correctness as it measures whether the explanation is faithful to the model's decision without requiring the ground truth attribution mask,' but ABPC is never defined anywhere in the manuscript. The cited reference is a demographic parity paper in which ABPC is a distribution-level disparity metric, not an explanation faithfulness metric. Because the optimizer maximizes this undefined objective for all recommended explainers, the claim that PnPXAI produces 'optimal explanations' is unsupported. The authors must define ABPC for attribution maps, state precisely how it is computed from the model and explanation, and validate it as a faithfulness measure (for example, by reporting its correlation with the ground-truth relevance mass and rank accuracy used in Figure 2).
- [Hyperparameter Optimizer / Figure 2] The quantitative validation of hyperparameter optimization reports relevance mass accuracy and rank accuracy improvements for most explainers, but it is based on a single model (ResNet50) on a single sliced-CT dataset, with no error bars, no repeated runs, and no statistical significance tests. Furthermore, the figure does not show the ABPC values for default versus optimized settings, so the reader cannot verify that the ABPC objective being optimized is actually aligned with the ground-truth fidelity metrics. Please provide per-seed or per-slice statistics, significance tests, and an explicit ABPC-versus-ground-truth correlation, or otherwise restrict the conclusion to the reported cases.
- [Recommender / Table 2] Table 2 is the sole basis for the recommender module, yet there is no empirical or theoretical justification that the mapping of explanation methods to modalities and architectures is correct or complete. For example, the table lists AttentionRollout and TransformerAttribution only for transformers, but it does not explain why these methods are not applicable to recurrent or convolutional architectures. Since the recommender's output is a core claimed advantage over other frameworks, a wrong or incomplete mapping would mislead non-expert users. Please provide a validation protocol (e.g., sanity checks on representative models of each architecture type, or references establishing applicability) and ideally a comparison of recommended versus non-recommended explainers on a common benchmark.
- [Acute Kidney Injury Detection / Figure 3] The text claims that 'the line graph in Figure 3 depicts better identification of the most important features by LIME, Integrated Gradients, which aligns with the highest scores in ABPC metric,' but no ABPC scores or ABPC-based rankings are displayed in Figure 3, and no quantitative measure of alignment (correlation, rank overlap, or similar) is provided. Without the ABPC values and an alignment statistic, the claimed coherence between the evaluator and known biomarkers is an assertion rather than a demonstrated result. Please report the actual ABPC values for the explainers and the alignment measure used.
- [Introduction / Table 1] The paper's 'universal' claim is supported only by the feature-checklist comparison in Table 1; there is no experimental comparison against Captum, OmniXAI, AutoXAI, Xaitk-Saliency, or OpenXAI on a common task. Since the novelty is framed as overcoming limitations of these frameworks, a direct comparison on at least one task (e.g., the liver tumor or AKI use case) is needed to substantiate the claimed advantages in flexibility, number of supported methods, and recommendation quality. Without such a comparison, the contribution is demonstrated only in absolute terms, not relative to existing tools.
minor comments (5)
- [Table 3] The table header contains a typo: 'A verage' should be 'Average'.
- [Table 1 caption] The caption contains a typo: 'Comparision' should be 'Comparison'.
- [Figure 2 text] The phrase 'As shown in the top center of Figure 2' is ambiguous; the intended panel (hyperparameter optimization trials, relevance mass accuracy, or relevance rank accuracy) should be named explicitly.
- [Code 1] The code snippet would benefit from a brief explanation of the expected data batch format, since the input_extractor, label_extractor, and target_extractor lambdas are the only documentation of the plug-and-play interface.
- [Evaluator section] The paper refers to the Co-12 categories but does not specify which concrete metrics instantiate correctness, continuity, and compactness, nor which of the three properties ABPC is supposed to belong to; a short table of the implemented metrics and their definitions would improve reproducibility.
Circularity Check
No significant circularity found; the framework's validation rests on external ground-truth masks and known medical biomarkers, not on its own optimization objective.
full rationale
The derivation chain in PnPXAI is an engineering pipeline rather than a formal derivation. The detector extracts the model's symbolic graph, the recommender filters explanation methods using a hand-authored mapping table (Table 2), and the optimizer selects hyperparameters by grid search against a chosen evaluator. No step claims to derive a quantity from a definition that already contains that quantity. The central validation is external: liver-tumor attributions are checked against segmentation masks from the LiTS benchmark, and AKI attributions are checked against known clinical biomarkers (eGFR, creatinine, BUN) from Makris and Spanou. The hyperparameter-optimization objective, ABPC, is not the same quantity used to assess success; the paper separately reports relevance mass accuracy and rank accuracy against ground-truth masks, so the improvement claim is not forced by construction. The AKI 'alignment' between ABPC and the share of expected features is a coherence check, not a tautology, because the expected features come from an independent medical source. The paper's self-citations (Kim and Choi 2021; Nam et al. 2020) are ordinary references to prior methods and are not load-bearing in any argument that reduces to them. The main concern is not circularity but evidentiary support: ABPC is cited from a demographic-parity paper (Han et al. 2023) and is never defined or validated as an explanation-faithfulness metric, so the claim that the optimizer finds 'optimal explanations' rests on an unexamined objective. That is a correctness risk, not a circular step, because the framework's conclusions do not reduce to the optimizer's own score. Overall, the paper's core claims are validated against external benchmarks and domain knowledge, and no prediction or derivation is equivalent to its inputs by construction.
Assumptions & free parameters
assumptions (3)
- domain assumption The mapping table (Table 2) correctly determines which explanation methods are applicable to each architecture and modality.
- ad hoc to paper ABPC is a valid faithfulness metric for explanation correctness.
- domain assumption Hyperparameter grid search on the user's dataset with ABPC yields near-optimal explanations.
Cite this review
Pith. "Pith review of PnPXAI: A Universal XAI Framework Providing Automatic Explanations Across Diverse Modalities and Models." pith.science (2026). https://pith.science/paper/S3YJGPWO
@misc{pith2026250510515,
author = {Pith},
title = {Pith review of: PnPXAI: A Universal XAI Framework Providing Automatic Explanations Across Diverse Modalities and Models},
year = {2026},
howpublished = {\url{https://pith.science/paper/S3YJGPWO}},
note = {Machine review of arXiv:2505.10515}
}
read the original abstract
Recently, post hoc explanation methods have emerged to enhance model transparency by attributing model outputs to input features. However, these methods face challenges due to their specificity to certain neural network architectures and data modalities. Existing explainable artificial intelligence (XAI) frameworks have attempted to address these challenges but suffer from several limitations. These include limited flexibility to diverse model architectures and data modalities due to hard-coded implementations, a restricted number of supported XAI methods because of the requirements for layer-specific operations of attribution methods, and sub-optimal recommendations of explanations due to the lack of evaluation and optimization phases. Consequently, these limitations impede the adoption of XAI technology in real-world applications, making it difficult for practitioners to select the optimal explanation method for their domain. To address these limitations, we introduce \textbf{PnPXAI}, a universal XAI framework that supports diverse data modalities and neural network models in a Plug-and-Play (PnP) manner. PnPXAI automatically detects model architectures, recommends applicable explanation methods, and optimizes hyperparameters for optimal explanations. We validate the framework's effectiveness through user surveys and showcase its versatility across various domains, including medicine and finance.
Figures
Figures from the paper (1 more)
Forward citations
Cited by 1 Pith paper
-
XAI and Statistical Analysis for Reliable Intrusion Detection in the UAVIDS-2025 Dataset: From Tree to Hybrid and Tabular DNN Ensembles
XGBoost with SHAP and statistical distribution analysis on UAVIDS-2025 identifies density support intersection as the cause of false predictions for Wormhole and Blackhole attacks in UAV intrusion detection.
Reference graph
Works this paper leans on
-
[1]
, " * write output.state after.block = add.period write newline
ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
-
[2]
write newline
" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
-
[3]
Abid, A.; Abdalla, A.; Abid, A.; Khan, D.; Alfozan, A.; and Zou, J. 2019. Gradio: Hassle-Free Sharing and Testing of ML Models in the Wild. arXiv preprint arXiv:1906.02569
arXiv 2019
-
[4]
Abnar, S.; and Zuidema, W. 2020. Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928
arXiv 2020
-
[5]
Adebayo, J.; Gilmer, J.; Muelly, M.; Goodfellow, I.; Hardt, M.; and Kim, B. 2018. Sanity checks for saliency maps. Advances in neural information processing systems, 31
2018
-
[6]
Agarwal, C.; Krishna, S.; Saxena, E.; Pawelczyk, M.; Johnson, N.; Puri, I.; Zitnik, M.; and Lakkaraju, H. 2022. Openxai: Towards a transparent evaluation of model explanations. Advances in neural information processing systems, 35: 15784--15799
work page 2022
-
[7]
J.; Neumann, D.; Samek, W.; M \"u ller, K.-R.; and Lapuschkin, S
Anders, C. J.; Neumann, D.; Samek, W.; M \"u ller, K.-R.; and Lapuschkin, S. 2021. Software for dataset-wide XAI: from local explanations to global insights with Zennit, CoRelAy, and ViRelAy. arXiv preprint arXiv:2106.13200
arXiv 2021
-
[8]
Arras, L.; Osman, A.; and Samek, W. 2022. CLEVR-XAI: A benchmark dataset for the ground truth evaluation of neural network explanations. Information Fusion, 81: 14--40
work page 2022
Show all 30 references
-
[9]
Bach, S.; Binder, A.; Montavon, G.; Klauschen, F.; M \"u ller, K.-R.; and Samek, W. 2015. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one, 10(7): e0130140
2015
-
[10]
R.; Dertkigil, S
Bassi, P. R.; Dertkigil, S. S.; and Cavalli, A. 2024. Improving deep neural network generalization and robustness to background bias via layer-wise relevance propagation optimization. Nature Communications, 15(1): 291
2024
-
[11]
B.; Vorontsov, E.; Ben-Cohen, A.; Kaissis, G.; Szeskin, A.; Jacobs, C.; Mamani, G
Bilic, P.; Christ, P.; Li, H. B.; Vorontsov, E.; Ben-Cohen, A.; Kaissis, G.; Szeskin, A.; Jacobs, C.; Mamani, G. E. H.; Chartrand, G.; et al. 2023. The liver tumor segmentation benchmark (lits). Medical Image Analysis, 84: 102680
2023
-
[12]
Chefer, H.; Gur, S.; and Wolf, L. 2021. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 782--791
2021
-
[13]
Cugny, R.; Aligon, J.; Chevalier, M.; Roman Jimenez, G.; and Teste, O. 2022. Autoxai: A framework to automatically select the most adapted xai solution. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, 315--324
2022
-
[14]
Han, X.; Jiang, Z.; Jin, H.; Liu, Z.; Zou, N.; Wang, Q.; and Hu, X. 2023. Retiring DP: New Distribution-Level Metrics for Demographic Parity. arXiv preprint arXiv:2301.13443
2023 arXiv
-
[15]
Hu, B.; Tunison, P.; RichardWebster, B.; and Hoogs, A. 2023. Xaitk-saliency: An open source explainable ai toolkit for saliency. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 15760--15766
2023
-
[16]
Jesus, S.; Pombal, J.; Alves, D.; Cruz, A.; Saleiro, P.; Ribeiro, R.; Gama, J.; and Bizarro, P. 2022. Turning the tables: Biased, imbalanced, dynamic tabular datasets for ml evaluation. Advances in Neural Information Processing Systems, 35: 33563--33575
2022
-
[17]
E.; Pollard, T
Johnson, A. E.; Pollard, T. J.; Shen, L.; Lehman, L.-w. H.; Feng, M.; Ghassemi, M.; Moody, B.; Szolovits, P.; Anthony Celi, L.; and Mark, R. G. 2016. MIMIC-III, a freely accessible critical care database. Scientific data, 3(1): 1--9
2016
-
[18]
Kim, S.; and Choi, J. 2021. Explaining the decisions of deep policy networks for robotic manipulations. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2663--2669. IEEE
2021
-
[19]
Kokhlikyan, N.; Miglani, V.; Martin, M.; Wang, E.; Alsallakh, B.; Reynolds, J.; Melnikov, A.; Kliushkina, N.; Araya, C.; Yan, S.; et al. 2020. Captum: A unified and generic model interpretability library for pytorch. arXiv preprint arXiv:2009.07896
2020 arXiv
-
[20]
M.; and Lee, S.-I
Lundberg, S. M.; and Lee, S.-I. 2017. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30
2017
-
[21]
Makris, K.; and Spanou, L. 2016. Acute kidney injury: definition, pathophysiology and clinical phenotypes. The clinical biochemist reviews, 37(2): 85
2016
-
[22]
H.; Osterrieder, J.; Hirsa, A.; Kulkarni, O.; and Lin, S
Misheva, B. H.; Osterrieder, J.; Hirsa, A.; Kulkarni, O.; and Lin, S. F. 2021. Explainable AI in credit risk management. arXiv preprint arXiv:2103.00949
2021 arXiv
-
[23]
Nam, W.-J.; Gur, S.; Choi, J.; Wolf, L.; and Lee, S.-W. 2020. Relative attributing propagation: Interpreting the comparative contributions of individual units in deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, 2501--2508
2020
-
[24]
Nauta, M.; Trienes, J.; Pathak, S.; Nguyen, E.; Peters, M.; Schmitt, Y.; Schl \" o tterer, J.; van Keulen, M.; and Seifert, C. 2023. From Anecdotal Evidence to Quantitative Evaluation Methods: A Systematic Review on Evaluating Explainable AI . ACM Comput. Surv. , 55(13s): 295:...
2023
-
[25]
Why should i trust you?
Ribeiro, M. T.; Singh, S.; and Guestrin, C. 2016. "Why should i trust you?" Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 1135--1144
2016
-
[26]
R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; and Batra, D
Selvaraju, R. R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; and Batra, D. 2017. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, 618--626
2017
-
[27]
Smilkov, D.; Thorat, N.; Kim, B.; Vi \'e gas, F.; and Wattenberg, M. 2017. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825
2017 arXiv
-
[28]
Srinivas, S.; and Fleuret, F. 2019. Full-gradient representation for neural network visualization. Advances in neural information processing systems, 32
2019
-
[29]
Sundararajan, M.; Taly, A.; and Yan, Q. 2017. Axiomatic attribution for deep networks. In International conference on machine learning, 3319--3328. PMLR
2017
-
[30]
Yang, W.; Le, H.; Laud, T.; Savarese, S.; and Hoi, S. C. 2022. Omnixai: A library for explainable ai. arXiv preprint arXiv:2206.01612
2022 arXiv
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.