{"id":"55f23b80-8f14-4846-af60-f4a099b76f94","arxiv_id":"2502.03014","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A technical report introducing xai_evals, a Python package that wraps existing explainability and metric libraries without adding new methods or validated results.","lead":"The authors present xai_evals, a wrapper library that integrates existing explainability tools and evaluation metrics. The report claims the library helps benchmark explanation quality, but provides no experimental validation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim depends on xai_evals being installable and benchmarked, but Section 3.3 still says 'once the package is published' and Section 4 contains no experiments; a clean install and reproduction of Tables 2-3 would settle whether the claim lands.","rationale":"The reader's weakest assumption—that xai_evals exists in installable form and reliably computes the listed metrics—is exactly the load-bearing condition for the paper's central claim. I read the manuscript as a software-contribution report rather than a methods paper; such a report stands or falls on whether the software is available and the demonstrations are reproducible. The text itself undermines availability in Section 3.3, and the promised benchmarking section is absent: Section 4 is titled 'The Importance of Explainable AI (XAI) and Regulatory Compliance,' not an experiments section, yet Section 5 asserts benchmark results. This is an internally verifiable omission, not a disagreement with community consensus. The acknowledgments also state that image metrics are computed by Quantus and attribution methods are drawn from Captum, tf-explain, LIME, and SHAP; wrapping existing libraries can be a legitimate contribution, but it makes code-level verification more important, not less. I give no credit for machine-checked proofs or released code because none are visible in the manuscript; the PyPI URL is the only artifact and is not demonstrated to resolve. The proposed test—clean install plus reproduction of Tables 2 and 3 plus independent metric reimplementation—is a single decisive check: if it fails on availability, no further analysis is needed; if it passes, the paper's core claims are materially supported even though the writing is loose. For these reasons I agree with the reader's rejection and recommend no change to the verdict.","tokens_in":11293,"tokens_out":5228,"duration_ms":44713,"concrete_test":"Query https://pypi.org/pypi/xai-evals/json (or run `pip install xai-evals` in a clean virtual environment) to confirm a release exists on or before the submission date. If it installs, clone the linked repository or inspect the installed package and run the exact Section 3.6.1 tabular SHAP example and the Section 3.6.3 `calculate_metrics()` call; compare the attribution list and metric row against Tables 2 and 3. In parallel, independently implement the Section 3.5.1 formulas for faithfulness, sensitivity, comprehensiveness, sufficiency, monotonicity, complexity, and sparseness on an IRIS RandomForest classifier with SHAP and check whether the package reproduces the published numbers within rounding tolerance. A missing package, a failed install, or materially different metric values would invalidate the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central assertion is that xai_evals is a comprehensive, open-sourced package for generating, benchmarking, and evaluating explanations (Abstract and Section 1). Every downstream claim—integration of SHAP, LIME, Grad-CAM, Integrated Gradients; computation of faithfulness, sensitivity, comprehensiveness, robustness; and the conclusion that 'benchmarking experiments showed that xai_evals is effective' (Section 5)—is contingent on the package existing in installable form and running as described. The body text gives direct evidence against this: Section 3.3.1 titles the pip route 'After Package is Published' and says 'once the package is published,' and the printed command 'pip install xa i_ ev al s' is not executable; no repository URL, commit hash, or version number is supplied. The abstract's PyPI link is unverified. Separately, the paper promises experimental validation: Section 1 says Section 4 'presents experimental results demonstrating the effectiveness of the package,' but Section 4 is a regulatory-compliance discussion with no benchmark data, tables, or protocols. The conclusion's effectiveness claim is therefore unsupported by anything in the manuscript. Even the metric definitions need code-level corroboration: tabular Faithfulness is unnormalized while image Faithfulness Correlation is normalized, and Monotonicity as defined compares signs of consecutive features rather than sign agreement with output changes. If the package is absent or uninstallable, the central claim fails; if the metrics are only wrappers with mismatched outputs, the claimed validation is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents xai_evals, a Python package claimed to generate, benchmark, and evaluate post-hoc local explanations for tabular and image data. It describes supported explainers (SHAP, LIME, Grad-CAM, Integrated Gradients, and the authors' DlBacktrace), a set of tabular and image metrics, and two usage examples with small output tables. The paper does not contain the empirical evaluation promised in the Introduction; Section 4 is a regulatory discussion, and the Conclusion refers to benchmarking experiments that never appear. Section 3.3 states the package can be installed 'once the package is published' and prints a corrupted pip command, so the package itself cannot be obtained or tested from the manuscript.","tokens_in":11614,"tokens_out":6622,"duration_ms":58679,"significance":"If the framework existed and worked as described, it would address a real need in XAI evaluation, and the related-work survey in Section 2.2 is a useful synthesis. The authors are transparent about depending on Quantus, Captum, and tf-explain for large parts of the implementation. The central claim, however, is unsubstantiated: no installation artifact, no benchmark results, and no correct or reproducible metric definitions are provided. Because the paper's contribution is the package and its evaluation, the current submission does not establish that the package exists or that any of its promised functionality is effective. It may be a basis for a future system paper, but as written it is a proposal, not a validated contribution.","major_comments":[{"comment":"Section 1 promises that Section 4 'presents experimental results demonstrating the effectiveness of the package,' but Section 4 is a discussion of GDPR and the EU AI Act with no data, tables, or protocols. Section 5 then states that 'the results of our benchmarking experiments showed that xai_evals is effective,' yet no such experiments appear anywhere in the manuscript. The central claim of demonstrated effectiveness is therefore unsupported.","section":"Section 4 and Section 5"},{"comment":"The only pip command shown is 'pip install xa i_ ev al s', which is not executable, and the surrounding text says installation can be done 'once the package is published.' No repository URL, commit hash, or version is provided. Since the paper's object of study is the package, the absence of any verifiable artifact means none of the integration or metric claims can be checked. The abstract's PyPI link does not remedy this, as no installation or import is demonstrated from it.","section":"Section 3.3.1"},{"comment":"The Monotonicity formula is 1/(n-1) times the sum of indicators of sign(a_i) = sign(a_{i+1}), which tests sign agreement between adjacent feature attributions, not whether attributions move with model output; it does not implement the stated definition. The tabular Faithfulness formula is an unnormalized sum of |f(x)-f(x'_i)| |a_i|, whereas the image Faithfulness Correlation in Section 3.5.2 is normalized by the sum of |a_i|, so equivalent-sounding metrics are not comparable across modalities. Table 3 reports Faithfulness = -0.568218 even though the formula is a nonnegative sum, and reports an 'infidelity' column with no definition in Section 3.5. These issues call into question whether the metric computations are correct.","section":"Section 3.5.1 and Table 3"},{"comment":"The evaluation example sets task='binary-classification' while using the three-class IRIS dataset and the same RandomForestClassifier used for multiclass classification in Section 3.6.1. This example, together with the corrupted whitespace in all code listings, suggests the usage snippets have not been run; combined with the missing package, the illustrative Tables 2 and 3 cannot be reproduced.","section":"Section 3.6.3"}],"minor_comments":[{"comment":"There are numerous typographical and formatting artifacts, including 'F RAMEWORK' and 'POST-H OC' in the title, 'Explainbility' in Section 2.1.1, and code listings with inserted spaces such as 'lo ad _i ri s ()'.","section":"Title and Section 2.1.1"},{"comment":"The 'infidelity' column is reported in Table 3 but no formula or reference for it is given in Section 3.5 or anywhere else.","section":"Table 3"},{"comment":"Figure 1 is referenced as the library overview but has no caption or descriptive text in the body, and Figure 2 is described only by a short caption; neither figure is discussed in enough detail to understand the design or the overlay method.","section":"Figures 1 and 2"},{"comment":"The metrics would benefit from explicit perturbation and masking protocols; formulas such as Comprehensiveness and Sufficiency are underspecified with respect to the choice of S_k, the masking rule, and the construction of x_focused.","section":"Section 3.5.1 and Section 3.5.2"}],"recommendation":"reject","confidential_remarks":"This manuscript reads as a software announcement rather than a research paper. Should the authors resubmit, I would need to see a working package with a pinned version, an installation test, and a real benchmarking section with protocols and datasets, plus correction of the metric definitions. The current submission does not meet the bar for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a software-announcement paper whose central claims are not backed by anything in the manuscript. The idea—a unified wrapper for SHAP, LIME, Grad-CAM, Integrated Gradients, and the authors' own DlBacktrace, with tabular and image evaluation metrics—is fine and could be useful to practitioners. The paper does a decent job situating itself: it cites Quantus, OpenXAI, Captum, tf-explain, FairX, and Inseq, and it is honest in the acknowledgments that image metrics are delegated to Quantus and attribution methods to Captum/tf-explain. The usage examples look plausible and would help a user get started.\n\nBut the soft spots are load-bearing. The introduction promises that Section 4 \"presents experimental results demonstrating the effectiveness of the package.\" Section 4 is a regulatory-compliance essay with no benchmark, no tables beyond the illustrative SHAP output, no protocol. The conclusion then says \"benchmarking experiments showed that xai_evals is effective\" without any such experiments appearing. That is a broken internal promise, not a style issue.\n\nThe package also cannot be verified from the text. The pip command is printed as \"pip install xa i_ ev al s\" (spaces inserted), no repository URL or commit hash is given, and the surrounding text says \"once the package is published.\" The PyPI link in the abstract is unverified. As submitted, the core artifact is not available for inspection.\n\nThe metric definitions have their own issues: tabular Faithfulness is unnormalized while image Faithfulness Correlation is normalized; Monotonicity, as written, checks whether adjacent feature attributions have the same sign, which is not the same as checking sign agreement with model-output changes; Sufficiency and Comprehensiveness formulas include an indicator over the top-k set but the averaging scheme is not clear. These might be typos, but they are exactly the kind of thing code-level reproducibility would settle. Integrating DlBacktrace from the authors' own prior work is not itself a problem, but self-citation does not substitute for validation.\n\nNo new algorithm, metric, or dataset is introduced; the contribution is integration. That is fine for a software report, but the paper cannot be evaluated as a research preprint without the promised benchmark and an installable package. If the authors publish the package, fix the install instructions, and add a real Section 4 with model/dataset/protocol details, this could become a useful practitioners' tool. As it stands, I wouldn't spend referee time on it.","headline":"Software-announcement paper whose promised package and experiments are absent; the wrapper idea is sensible, but this draft doesn't support its central claim.","tokens_in":766,"tokens_out":2471,"would_cite":false,"duration_ms":51982,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper presents xai_evals, a Python package aimed at unifying the generation, benchmarking, and evaluation of post-hoc local explanations for machine learning models, across tabular and image data.","keywords":["explainable AI","post-hoc explanations","feature attribution","explanation evaluation","benchmarking","SHAP","LIME","Grad-CAM"],"falsifier":"Install the package from the PyPI URL given in the abstract, run the Section 3.6 IRIS example exactly as written, and check that the computed metrics reproduce Table 3 (for SHAP on the IRIS Random Forest: faithfulness -0.568218, infidelity 0.000048, sensitivity 0.00328, and so on). If installation fails, the package is absent from PyPI, or the values do not match, the paper's central claim is not supported.","tokens_in":11157,"feed_emoji":"🧪","tokens_out":8928,"duration_ms":76311,"temperature":0.7,"pith_summary":"The paper claims that xai_evals, a new Python package, fills a gap in explainable AI by bringing explanation generation and explanation evaluation into one tool. It integrates widely used attribution methods--SHAP, LIME, Grad-CAM, Integrated Gradients, and DlBacktrace--and measures their output with metrics such as faithfulness, sensitivity, comprehensiveness, sufficiency, and robustness. The intended payoff is a standardized way to compare explanation methods across tabular and image models, so that researchers and practitioners can judge not only whether a model is accurate but whether its reasons are trustworthy. If the package works as described, it would give the field a common benchmarking surface for post-hoc local explanations.","feed_headline":"One Python library to unify and grade AI explanations","feed_subtitle":"Wraps SHAP, LIME, Grad-CAM and more, with faithfulness and robustness metrics for tabular and image models.","key_machinery":"The central machinery is the package's paired class structure: illustration classes such as SHAPExplainer, LIMEExplainer, TorchImageExplainer, and DlBacktrace explainers produce attribution maps, while ExplanationMetricsTabular and ExplanationMetricsImage quantify them using perturbation-based formulas--faithfulness compares output change under feature perturbation to attribution magnitude, sensitivity measures attribution change under noise, comprehensiveness and sufficiency mask the top-k features, and image metrics such as MPRT average output changes over perturbed pixels.","core_discovery":"On its own terms, the paper's discovery is that a single Python package, xai_evals, can combine explanation generation and evaluation for both tabular and image data, wrapping established attribution methods and perturbation-based metrics, and that this combination addresses the fragmentation of existing explainability tooling. The package is described as supporting classical machine learning models and deep learning models through a unified API, with separate paths for tabular and image modalities.","pith_inferences":["As an editorial extension: the paper states that benchmarking experiments showed the package is effective, but the text reports no such experiments; verifying effectiveness on a fixed dataset is a natural and necessary next step.","As an editorial extension: because the package builds on existing open-source explanation and metric libraries, its numerical outputs inherit those libraries' version-specific behaviors, so comparing releases may require care.","As an editorial extension: a testable extension of the package's own logic would be to run all supported explainers on one tabular and one image benchmark and publish the metric table, which the present report does not provide."],"forward_implications":["A practitioner could run SHAP, LIME, and Integrated Gradients on the same model and dataset through one API and read their faithfulness, sensitivity, and complexity scores from a single table.","Tabular and image models could be evaluated under the same metric conventions, making cross-modal comparisons of explanation quality possible.","Teams deploying models in regulated settings could use the generated metric values as documentation for why a prediction was made.","The package would lower the barrier to running explanation evaluations, since it removes the need to assemble separate libraries for generation and evaluation."],"supporting_citations":[{"why":"Supplies the SHAP attribution method the package integrates for tabular and deep models.","marker":"Lundberg and Lee [2017]"},{"why":"Supplies the LIME surrogate-model explanation method the package integrates.","marker":"Ribeiro et al. [2016]"},{"why":"Supplies the Grad-CAM attribution method used for image models.","marker":"Selvaraju et al. [2016]"},{"why":"Supplies the Integrated Gradients attribution method.","marker":"Sundararajan et al. [2017]"},{"why":"Supplies the DlBacktrace layer-wise backpropagation method integrated by the package.","marker":"Sankarapu et al. [2024]"},{"why":"Provides the Quantus metric set used for image-model metric computation in the current version.","marker":"Hedström et al. [2023]"},{"why":"Provides Captum attribution methods, including Integrated Gradients, Saliency, DeepLift, and Grad-CAM, for PyTorch models.","marker":"Kokhlikyan et al. [2020]"},{"why":"Provides tf-explain methods such as Grad-CAM and Occlusion Sensitivity for TensorFlow and Keras models.","marker":"Meudec [2021]"},{"why":"Defines the Max Sensitivity and Avg Sensitivity image metrics adopted by the package.","marker":"Yeh et al. [2019]"},{"why":"Defines the MPRT robustness metric adopted by the package.","marker":"Adebayo et al. [2018]"}],"fun_headline_variants":["Grade your AI explanations with xai_evals","One library to generate and score explanation methods","Unified evaluation for SHAP, LIME, Grad-CAM and more","Faithfulness and robustness metrics for any explanation method","Toolkit for benchmarking post-hoc local explanations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the xai_evals package actually exists in installable form and reliably computes the listed metrics; the paper gives no working installation command, repository identifier, or reported experiments to confirm this.","fun_headline_variants_meta":{"raw":{"variants":["Grade your AI explanations with xai_evals","One library to generate and score explanation methods","Unified evaluation for SHAP, LIME, Grad-CAM and more","Faithfulness and robustness metrics for any explanation method","Toolkit for benchmarking post-hoc local explanations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1581,"prompt_tokens":799,"completion_tokens":782,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":415,"completion_tokens_details":{"reasoning_tokens":706}},"tokens_in":415,"tokens_out":782,"duration_ms":7583,"temperature":1.0,"reasoning_tokens":706,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:11:01.393862+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Install the package from the PyPI URL given in the abstract, run the Section 3.6 IRIS example exactly as written, and check that the computed metrics reproduce Table 3 (for SHAP on the IRIS Random Forest: faithfulness -0.568218, infidelity 0.000048, sensitivity 0.00328, and so on). If installation fails, the package is absent from PyPI, or the values do not match, the paper's central claim is not supported.","supporting_citations":[{"cited_title":"why should i trust you?","cited_arxiv_id":null,"evidence_quote":"Supplies the LIME surrogate-model explanation method the package integrates."},{"cited_title":"Axiomatic attribution for deep networks","cited_arxiv_id":null,"evidence_quote":"Supplies the Integrated Gradients attribution method."},{"cited_title":"tf-explain, 2021","cited_arxiv_id":null,"evidence_quote":"Provides tf-explain methods such as Grad-CAM and Occlusion Sensitivity for TensorFlow and Keras models."},{"cited_title":"Inouye, and Pradeep Ravikumar","cited_arxiv_id":null,"evidence_quote":"Defines the Max Sensitivity and Avg Sensitivity image metrics adopted by the package."},{"cited_title":"Goodfellow, Moritz Hardt, and Been Kim","cited_arxiv_id":null,"evidence_quote":"Defines the MPRT robustness metric adopted by the package."}],"review_version":1}