Pith. sign in

REVIEW 3 major objections 5 minor 115 references

This paper argues that explainable AI research has inverted its natural order — building methods before defining what explanations are for — and must pivot to foundations: clear definitions, falsifiable properties, task-grounded evaluation,

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 10:41 UTC pith:Z2D4JRPQ

load-bearing objection A well-argued position paper with a useful new empirical survey, but the survey's headline numbers are computed over a broader population than the critique's target, and the post-hoc-only breakdown is missing. the 3 major comments →

arxiv 2607.14123 v1 pith:Z2D4JRPQ submitted 2026-06-18 cs.LG cs.AI

Position: Explainability Research Must Prioritize Foundations over Ad-hoc Methods

Michal Moshkovitz , Suraj Srinivas , Lesia Semenova , Nave Frost , Cyrus Rashtchian , Valentyn Boreiko , Shichang Zhang , Himabindu Lakkaraju
show 2 more authors
Cynthia Rudin Jennifer Wortman Vaughan
This is my paper
classification cs.LG cs.AI
keywords explainable AIXAIinterpretabilityfoundationsevaluationfaithfulnesshuman-centered AIresearch agenda
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that explainable AI (XAI) has stalled not from a lack of methods but from missing foundations: no shared definition of what an explanation is, claims that are rarely testable, evaluation metrics not tied to any real task, and no mechanism for explanations to feed back into model decisions. It supports this with an LLM-assisted review of 617 papers from three major machine learning conferences, finding that 77% propose novel methods while only 11% give formal definitions and 11% run user studies, plus a survey of 34 practitioners who mostly choose methods by ease and popularity and report not knowing how to check explanation quality. If the diagnosis is right, the field should shift resources from inventing new explanation techniques toward building conceptual frameworks, evaluation standards, and human-in-the-loop pipelines. The paper closes with a five-point checklist for making explainability papers define their goals, state falsifiable claims, demonstrate faithful behavior, name a concrete application, and evaluate in context with users.

Core claim

The central claim is that the explainability community has inverted the natural research order — prioritizing methods development over the objectives and evaluation criteria that would let results accumulate — and that this inversion, not any shortage of techniques, explains why explanations rarely influence real-world workflows. The paper's evidence: across 617 recent papers at leading ML conferences, only 11% contain any formal definition of an explainability goal, only 11% involve user studies, and 77% propose novel methods; among 34 surveyed practitioners, method choice is driven by ease of implementation and popularity, 44% do not know how to verify explanation quality, and reported ben

What carries the argument

The argument is carried by an empirical diagnostic plus a conceptual frame. The diagnostic is an LLM-based literature survey that codes 617 papers from ICML, NeurIPS, and ICLR on 20 binary questions about what the papers claim to do (e.g., propose a novel method, contain a formal definition, run a user study, test faithfulness), validated on 25 hand-labelled pairs with 88% agreement; the conceptual frame is a four-challenge structure — definitions, properties, evaluations, applications — in which explanations are treated not as standalone outputs but as components of human-in-the-loop systems. Central to the frame is 'task-grounded objective evaluation,' where automated metrics are published

Load-bearing premise

The load-bearing premise is that the two measurements — an LLM's coding of 617 papers' claims, validated on only 25 hand-labelled pairs, and a self-selected group of 34 practitioners — accurately capture what the field does and how practitioners fare; if the LLM systematically misreads the papers or the respondents are unrepresentative, the headline statistics and the gap narrative built on them shift.

What would settle it

Hand-code a larger random sample of the 617 papers (or a fresh sample from the same venues) with two independent human annotators and compare their answers to the LLM's on the key binary questions; if agreement on 'contains a formal definition' or 'runs a user study' falls below, say, 80% on a sample of 100 papers, the survey statistics are not stable. Alternatively, run a prospective controlled comparison in which one group of practitioners uses a foundations-first pipeline (explicit objectives, task-grounded metrics, feedback loop) and another uses current ad-hoc tools; if the foundations-fi

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the field adopts the paper's checklist, future XAI papers would need to state a precise, purpose-grounded definition of the explanation goal, make falsifiable claims, test faithfulness against that goal, name a concrete application, and evaluate in context with users — criteria that would filter out many current method papers.
  • Review processes would need to treat problem formulation, evaluation design, and human-centered integration as substantive contributions comparable to new methods, changing what gets published and funded.
  • A 'task-grounded objective evaluation' standard would reframe metric design: automated measures such as stability or sparsity would be treated as hypotheses whose relevance must be empirically validated against downstream task performance, rather than assumed universally desirable.
  • Explanations would be designed as interfaces for feedback loops, forcing formal work on what object is modified (parameters vs. constraints), who is authorized to modify it, and how feedback propagates through the system.
  • The field could develop benchmarks and comparison protocols for XAI analogous to those for predictive accuracy, enabling cumulative measurement of progress rather than fragmented case studies.

Where Pith is reading between the lines

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

  • If the diagnosis is correct, a measurable prediction follows: the distribution of claims in the 2025–2026 rounds of the same venues should shift toward definitions, user studies, and task-grounded evaluation if the call is heeded; a repeat of the LLM survey in two years would test that directly.
  • The framework suggests a concrete research program: for any explanation class (e.g., concept-based methods), first fix the definition and falsifiable properties, then build benchmarks; this could dissolve longstanding disputes like whether attention is an explanation by making the target property explicit.
  • The paper's human-centric framing implies that XAI progress should be measured by downstream task performance (debugging speed, decision accuracy, audit effectiveness) rather than by explanation-to-model correspondence alone; this could change how XAI papers are compared and ranked.
  • A testable extension of 'task-grounded objective evaluation' would be to validate existing faithfulness metrics by measuring whether improvements in those metrics predict improvements in human task performance across a battery of tasks; metrics that fail this correlation would be retired.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This position paper argues that explainable AI (XAI) research has, to its detriment, prioritized the development of ad-hoc post-hoc explanation methods over foundational questions: definitions of explanation, formally specified properties, task-grounded evaluation, and pipelines for explanation-driven action and feedback. The authors support this diagnosis with two empirical instruments: an LLM-driven survey of 617 papers from NeurIPS, ICML, and ICLR (with additional CVPR/ACL material) using 20 binary questions, validated on 25 randomly sampled paper-question pairs with 88% agreement; and a survey of 34 XAI practitioners (from 43 total respondents) about their goals, method choices, evaluation practices, and difficulties. From these they derive four challenges (definitions, properties, evaluations, applications), a five-point checklist for future XAI papers, and a set of open problems. The paper is written as a normative position piece, and it explicitly distinguishes post-hoc explainability from interpretable-by-construction models.

Significance. If the paper's central claim is correct, it would justify a reallocation of XAI research effort and publication norms toward problem formulation, falsifiable evaluation, and human-in-the-loop integration rather than incremental method proposal. The paper usefully combines an explicit normative position with a large-scale, reproducible-in-principle survey of the current literature and a practitioner survey. Its strengths include a clear statement of the four foundational challenges, a concrete checklist, transparency about the LLM survey's scope (it records claims, not scientific merit), open acknowledgement of practitioner-survey selection bias, and detailed appendices with full question tables and demographic breakdowns. The empirical support is, however, thinner than the abstract's rhetoric suggests: the literature survey's target population does not match the population used for the headline statistics, and the validation of the LLM instrument is too small to support precise percentages. These issues are fixable and do not invalidate the position, but they should be addressed before the paper is used as an empirical foundation for a field-wide reorientation.

major comments (3)
  1. [§2.1, Table 1] The stated target of the critique is ad-hoc post-hoc XAI, yet the headline statistics are computed over the pooled set of 617 papers, which Table 1 Q1/Q2 shows is only 25.4% primarily post-hoc and 50.1% primarily interpretable-by-construction. The 77% novel-method figure (Q4), the 11% formal-definition figure (Q6), and the 11% user-study figure (Q12) are reported for the pooled sample. If the post-hoc subset has a different composition, the central empirical diagnosis is not actually about the population it claims to critique. The authors should either report Q4, Q6, Q12, and ideally all questions for the Q1=True subset separately, or reframe the survey as covering all interpretability/explainability research and adjust the abstract and introduction accordingly. This is a load-bearing issue because the paper's main recommendation targets ad-hoc methods.
  2. [§2.1, validation paragraph] The validation of the LLM-based survey is too thin for the precision with which the results are reported. Only 25 randomly sampled (paper, question) pairs are manually labeled out of roughly 617×20 ≈ 12,340 instances, and only aggregate agreement (88%) is reported. Per-question agreement is unknown, so the specific questions that carry the argument — Q4 (novel method), Q6 (formal definition), Q12 (user study), Q14 (faithfulness) — have no stated error rates. A systematic misreading of even one of these questions could materially change the headline percentages. The authors should report per-question agreement, a confusion matrix for the validation pairs, or confidence intervals, and ideally validate the classification questions that define the target population (Q1, Q2, Q4, Q6) on a larger sample.
  3. [§2.2, Table 4, 'Optimism' bullet] The practitioner survey is presented as one of the two main empirical pillars, but the interpretation of the '76.5% beneficial' result as evidence that XAI is 'beneficial yet ad hoc' is fragile. The paper itself acknowledges selection bias ('practitioners who found explanations useful may have been more inclined to participate'), but the surrounding wording — 'Our survey of researchers and practitioners reveals...' — goes beyond what a self-selected sample of N=34 can support. If the non-responding population found XAI useless, the 'unrealized potential' framing would be substantially weaker. The authors should temper the conclusions drawn from this instrument in the main text, for example by explicitly labeling the survey as a pilot or hypothesis-generating study, and by moving the caveats to the point where the 'Optimism' bullet is introduced.
minor comments (5)
  1. [§2.1 vs Table 1] The text says the survey covers 'NeurIPS, ICML, and ICLR in 2023–2024,' but Table 1 reports columns for ICLR25. Please specify the exact conference cycles included and reconcile the wording.
  2. [Section 2 heading] Typo: 'practioners' should be 'practitioners.'
  3. [References] Several references have spacing/encoding artifacts: 'V ogt' (Sokol & Vogt), 'L¨ofstr¨om', 'Kstner'. Please run a normalization pass.
  4. [§1, last paragraph before 'Alternative Views'] Minor grammar: 'this paper views explainability not as a property of the model alone, but as a fundamentally human-centric' appears to be missing a noun (e.g., 'human-centric endeavor').
  5. [§2.1, finding 4] Q7 indicates 76% of papers claim a concrete downstream impact, yet the paper's opening states explanations rarely influence real-world workflows. Since Q7 captures claims rather than deployed use, please clarify this distinction in the main text so the two statements do not appear contradictory.

Circularity Check

0 steps flagged

No significant circularity: the paper is an argued position with survey evidence; the handful of self-citations are supporting, not load-bearing.

full rationale

This paper is a position paper, not a derivational claim. Its central recommendation—that XAI research should prioritize definitions, falsifiable properties, task-grounded evaluation, and actionability—is argued from survey evidence and prior literature rather than derived from equations or fitted parameters. No step in the argument equates an output to an input by construction. The LLM-driven survey (§2.1) is explicitly framed as extracting claims from papers (with a 25-pair, 88% agreement validation), and the practitioner survey (§2.2) explicitly acknowledges selection bias; these are methodological caveats, not circular reductions. The paper does cite several works by its own authors (Rudin 2019; Dasgupta, Frost & Moshkovitz 2022; Vaughan & Wallach 2021; Han et al. 2022; Bhalla et al. 2024; Liao & Vaughan 2024), but in each case the citation is one of several supporting references or an illustrative example, and the conclusion does not rest on any unique theorem or fitted value from those works. The checklist in §4.2 is grounded in independent examples (Adebayo et al. 2018; Hooker et al. 2019; Ustun et al. 2019; Nauta et al. 2023) and is presented as a recommendation, not as a consequence of the surveys. The pooled statistics in Table 1 mix post-hoc and interpretable-by-construction papers, which is a scope limitation for the empirical claim, but it does not make the claim definitionally true nor reduce the recommendation to its inputs. Overall, there is no circularity in the derivation chain; the only noteworthy feature is a handful of non-load-bearing self-citations, which warrants a low nonzero score.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

As a position paper, the content rests on background assumptions about LLM reliability, survey representativeness, and the completeness of the four-challenge taxonomy, rather than on fitted parameters or derived equations. There are no free parameters in the mathematical sense; the small sample sizes (25 validation pairs; 34 practitioners) are design choices bounding every headline percentage.

axioms (5)
  • domain assumption An LLM (Gemini 2.5 Flash) can accurately extract claim-level facts ('proposes method', 'evaluates faithfulness') from research paper abstracts/full texts.
    Invoked in §2.1; validated on only 25 randomly sampled pairs (88% agreement), no error bars on 617-paper aggregates.
  • domain assumption Keyword screening (explainability/interpretable) plus LLM filtering identifies the relevant XAI population across NeurIPS/ICML/ICLR 2023–2024.
    §2.1 two-stage filter; CVPR/ACL robustness check covers only 79+108 papers.
  • domain assumption Self-reported survey answers from 34 XAI practitioners reflect actual usage and evaluation practice.
    §2.2; authors acknowledge selection bias and prior work showing practitioners overstate usefulness (Kaur et al. 2020).
  • ad hoc to paper The four challenges (definitions, properties, evaluations, applications) form a complete-enough decomposition of XAI's foundational failures.
    Introduced in Section 3 as the paper's organizing framework; no independent justification that these four are exhaustive rather than the authors' chosen framing.
  • domain assumption Interpretable-by-design models are preferred when feasible.
    Cites Rudin (2019); underlies the paper's scoping of post-hoc XAI as a fallback rather than primary approach—a contested view in the field.

pith-pipeline@v1.3.0-alltime-deepseek · 21311 in / 9699 out tokens · 90502 ms · 2026-08-02T10:41:16.324027+00:00 · methodology

0 comments
read the original abstract

Despite the proliferation of Explainable AI (XAI) techniques -- from feature attributions to sparse autoencoders -- explanations rarely influence real-world workflows. In practice, they are often generated and discarded without guiding meaningful action. This gap reflects foundational shortcomings: research has not yet established methodologies for integrating explanations into end-to-end, human-in-the-loop systems. This position paper argues that the machine learning community must pivot from ad-hoc XAI methods toward addressing foundational & structural challenges, including unclear problem formulations, underspecified evaluation objectives, and the absence of pipelines for explanation-driven feedback. We support this claim through an analysis of recent ICML, NeurIPS, and ICLR papers and a survey of XAI practitioners, revealing recurring issues that limit cumulative progress. We conclude by outlining a practical checklist designed to shift XAI toward a more human-centered, action-oriented paradigm. By emphasizing foundational clarity over the development of ad-hoc methods, we hope to provide a roadmap for integrating explanations into actionable, feedback-driven AI systems.

Figures

Figures reproduced from arXiv: 2607.14123 by Cynthia Rudin, Cyrus Rashtchian, Himabindu Lakkaraju, Jennifer Wortman Vaughan, Lesia Semenova, Michal Moshkovitz, Nave Frost, Shichang Zhang, Suraj Srinivas, Valentyn Boreiko.

Figure 1
Figure 1. Figure 1: We lay out four interdependent challenges centered on definitions, properties, evaluations, and applications. Stakeholders’ needs and constraints shape how each challenge is instantiated, guiding the design, evaluation, and deployment of explanations. Because these challenges are deeply interconnected, progress on one challenge requires and supports progress in the others. structural issues — specifically,… view at source ↗
Figure 2
Figure 2. Figure 2: A five-point checklist for evaluating explainability papers. For each criterion, green checkmarks indicate desirable properties, while red crosses illustrate common pitfalls that undermine rigor or clarity. 4. A Call to Action: Foundations-driven XAI 4.1. What Becomes Possible with Foundational Work Addressing the foundational challenges identified in Sec￾tion 3 opens up new classes of research questions t… view at source ↗

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

115 extracted references · 1 canonical work pages

  1. [1]

    Transactions on Machine Learning Research , year=

    Holistic Evaluation of Language Models , author=. Transactions on Machine Learning Research , year=

  2. [2]

    Adaptive Testing and Debugging of NLP Models

    Ribeiro, Marco Tulio and Lundberg, Scott. Adaptive Testing and Debugging of NLP Models. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022

    2022

  3. [3]

    Dingli Yu and Simran Kaur and Arushi Gupta and Jonah Brown-Cohen and Anirudh Goyal and Sanjeev Arora , booktitle=

  4. [4]

    PLOS Digital Health , year=

    Generating Realistic Counterfactuals for Retinal Fundus and OCT Images using Diffusion Models , author=. PLOS Digital Health , year=

  5. [5]

    Advances in Neural Information Processing Systems , year=

    Diffusion Visual Counterfactual Explanations , author=. Advances in Neural Information Processing Systems , year=

  6. [6]

    International Conference on Machine Learning , year=

    An Interpretable N-gram Perplexity Threat Model for Large Language Model Jailbreaks , author=. International Conference on Machine Learning , year=

  7. [7]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Sanity checks for saliency metrics , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  8. [8]

    International Conference on Learning Representations , year=

    Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting , author=. International Conference on Learning Representations , year=

  9. [9]

    International Conference on Computer Vision , year=

    Adaptive Testing of Computer Vision Models , author=. International Conference on Computer Vision , year=

  10. [10]

    Identification of Systematic Errors of Image Classifiers on Rare Subgroups , author =

  11. [11]

    International Conference on Computer Vision , year=

    Spurious Features Everywhere -- Large-Scale Detection of Harmful Spurious Features in ImageNet , author=. International Conference on Computer Vision , year=

  12. [12]

    International Conference on Artificial Intelligence and Statistics , year =

    Generating Interpretable Counterfactual Explanations By Implicit Minimisation of Epistemic and Aleatoric Uncertainties , author =. International Conference on Artificial Intelligence and Statistics , year =

  13. [13]

    Hines and John P

    Sahil Verma and Varich Boonsanong and Minh Hoang and Keegan E. Hines and John P. Dickerson and Chirag Shah , title =. arXiv preprint, arXiv:2010.10596 , year =

    Pith/arXiv arXiv 2010

  14. [14]

    German Conference on Pattern Recognition , year=

    Sparse Visual Counterfactual Explanations in Image Space , author=. German Conference on Pattern Recognition , year=

  15. [15]

    Advances in Neural Information Processing Systems , year =

    Deep Structural Causal Models for Tractable Counterfactual Inference , author =. Advances in Neural Information Processing Systems , year =

  16. [16]

    Medical Image Computing and Computer Assisted Intervention , year=

    Visual Explanations for the Detection of Diabetic Retinopathy from Retinal Fundus Images , author=. Medical Image Computing and Computer Assisted Intervention , year=

  17. [17]

    2020 , booktitle =

    The Hidden Assumptions behind Counterfactual Explanations and Principal Reasons , author =. 2020 , booktitle =

    2020

  18. [18]

    Advances in Neural Information Processing Systems , year=

    Explanations based on the Missing: Towards Contrastive Explanations with Pertinent Negatives , author=. Advances in Neural Information Processing Systems , year=

  19. [19]

    arXiv preprint arXiv:2202.01602 , year=

    The disagreement problem in explainable machine learning: A practitioner's perspective , author=. arXiv preprint arXiv:2202.01602 , year=

    Pith/arXiv arXiv

  20. [20]

    2025 , month =

    Dario Amodei , title =. 2025 , month =

    2025

  21. [21]

    Transformer Circuits Thread , year=

    Towards Monosemanticity: Decomposing Language Models with Sparse Autoencoders , author=. Transformer Circuits Thread , year=

  22. [22]

    Manning and Christopher Potts , journal=

    Zhengxuan Wu and Aryaman Arora and Atticus Geiger and Zheng Wang and Jing Huang and Dan Jurafsky and Christopher D. Manning and Christopher Potts , journal=. Ax

  23. [23]

    Transformer Circuits Thread , year=

    A Mathematical Framework for Transformer Circuits , author=. Transformer Circuits Thread , year=

  24. [24]

    Advances in Neural Information Processing Systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in Neural Information Processing Systems , volume=

  25. [25]

    ACM Conference on Fairness, Accountability, and Transparency , year=

    Explainable Machine Learning in Deployment , author=. ACM Conference on Fairness, Accountability, and Transparency , year=

  26. [26]

    arXiv preprint arXiv:2501.16496 , year=

    Open problems in mechanistic interpretability , author=. arXiv preprint arXiv:2501.16496 , year=

    Pith/arXiv arXiv

  27. [27]

    2023 , month =

    Steven Casper , title =. 2023 , month =

    2023

  28. [28]

    2025 , month =

    Smith, Lewis and Rajamanoharan, Senthooran and Conmy, Arthur and McDougall, Callum and Lieberum, Tom and Kramár, János and Shah, Rohin and Nanda, Neel , title =. 2025 , month =

    2025

  29. [29]

    Distill , year =

    Olah, Chris and Cammarata, Nick and Ludwig, Ludwig and Schubert, Ludwig and others , title =. Distill , year =

  30. [30]

    Advances in Neural Information Processing Systems , volume=

    A unified approach to interpreting model predictions , author=. Advances in Neural Information Processing Systems , volume=

  31. [31]

    International Conference on Machine Learning , year=

    Understanding black-box predictions via influence functions , author=. International Conference on Machine Learning , year=

  32. [32]

    Harvard Journal of Law & Technology , year =

    Sandra Wachter and Brent Mittelstadt and Chris Russell , title =. Harvard Journal of Law & Technology , year =

  33. [33]

    arXiv preprint arXiv:2411.04430 , year=

    Towards unifying interpretability and control: Evaluation via intervention , author=. arXiv preprint arXiv:2411.04430 , year=

    Pith/arXiv arXiv

  34. [34]

    Faithfulness Tests for Natural Language Explanations

    Atanasova, Pepa and Camburu, Oana-Maria and Lioma, Christina and Lukasiewicz, Thomas and Simonsen, Jakob Grue and Augenstein, Isabelle. Faithfulness Tests for Natural Language Explanations. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2023

    2023

  35. [35]

    arXiv preprint arXiv:2307.13702 , year=

    Measuring Faithfulness in Chain-of-Thought Reasoning , author=. arXiv preprint arXiv:2307.13702 , year=

    Pith/arXiv arXiv

  36. [36]

    Advances in Neural Information Processing Systems , volume=

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting , author=. Advances in Neural Information Processing Systems , volume=

  37. [37]

    2022 , journal=

    Toy Models of Superposition , author=. 2022 , journal=

    2022

  38. [38]

    Which Explanation Should

    Han, Tessa and Srinivas, Suraj and Lakkaraju, Himabindu , booktitle=. Which Explanation Should

  39. [39]

    Proceedings of NAACL-HLT , pages=

    Attention is not Explanation , author=. Proceedings of NAACL-HLT , pages=

  40. [40]

    Attention is not not Explanation , author=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , year=

    2019

  41. [41]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    Evaluating the visualization of what a deep neural network has learned , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2016 , publisher=

    2016

  42. [42]

    Advances in Neural Information Processing Systems , year =

    Full-Gradient Representation for Neural Network Visualization , author =. Advances in Neural Information Processing Systems , year =

  43. [43]

    The Thirteenth International Conference on Learning Representations , year=

    Scaling and evaluating sparse autoencoders , author=. The Thirteenth International Conference on Learning Representations , year=

  44. [44]

    Nature Machine Intelligence , volume=

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead , author=. Nature Machine Intelligence , volume=. 2019 , publisher=

    2019

  45. [45]

    Opportunities and challenges in explainable artificial intelligence (

    Das, Arun and Rad, Paul , journal=. Opportunities and challenges in explainable artificial intelligence (

  46. [46]

    Explainable

    Dwivedi, Rudresh and Dave, Devam and Naik, Het and Singhal, Smiti and Omer, Rana and Patel, Pankesh and Qian, Bin and Wen, Zhenyu and Shah, Tejal and Morgan, Graham and others , journal=. Explainable. 2023 , publisher=

    2023

  47. [47]

    Ribeiro, Marco Tulio and Singh, Sameer and Guestrin, Carlos , booktitle=. ``

  48. [48]

    arXiv preprint arXiv:1702.08608 , year=

    Towards a rigorous science of interpretable machine learning , author=. arXiv preprint arXiv:1702.08608 , year=

    Pith/arXiv arXiv

  49. [49]

    Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP , pages=

    Mechanistic? , author=. Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP , pages=

  50. [50]

    One explanation does not fit all: A toolkit and taxonomy of

    Arya, Vijay and Bellamy, Rachel KE and Chen, Pin-Yu and Dhurandhar, Amit and Hind, Michael and Hoffman, Samuel C and Houde, Stephanie and Liao, Q Vera and Luss, Ronny and Mojsilovi. One explanation does not fit all: A toolkit and taxonomy of. arXiv preprint arXiv:1909.03012 , year=

    Pith/arXiv arXiv 1909

  51. [51]

    Towards faithful model explanation in

    Lyu, Qing and Apidianaki, Marianna and Callison-Burch, Chris , journal=. Towards faithful model explanation in. 2024 , publisher=

    2024

  52. [52]

    arXiv preprint arXiv:2401.07927 , year=

    Are self-explanations from Large Language Models faithful? , author=. arXiv preprint arXiv:2401.07927 , year=

    Pith/arXiv arXiv

  53. [53]

    Advances in Neural Information Processing Systems , year=

    Sanity checks for saliency maps , author=. Advances in Neural Information Processing Systems , year=

  54. [54]

    arXiv preprint arXiv:2404.13874 , year=

    Valor-eval: Holistic coverage and faithfulness evaluation of large vision-language models , author=. arXiv preprint arXiv:2404.13874 , year=

    Pith/arXiv arXiv

  55. [55]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year=

    A Comparative Study of Faithfulness Metrics for Model Interpretability Methods , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year=

  56. [56]

    Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    On the Faithfulness of Vision Transformer Explanations , author=. Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  57. [57]

    arXiv preprint arXiv:2407.20899 , year=

    Faithful and plausible natural language explanations for image classification: A pipeline approach , author=. arXiv preprint arXiv:2407.20899 , year=

    Pith/arXiv arXiv

  58. [58]

    International Conference on Machine Learning , year=

    Framework for evaluating faithfulness of local explanations , author=. International Conference on Machine Learning , year=

  59. [59]

    Knowledge-Based Systems , volume=

    Faithful Counterfactual Visual Explanations (FCVE) , author=. Knowledge-Based Systems , volume=. 2024 , publisher=

    2024

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Openxai: Towards a transparent evaluation of model explanations , author=. Advances in Neural Information Processing Systems , volume=

  61. [61]

    ArXiv , year=

    Generating Contrastive Explanations with Monotonic Attribute Functions , author=. ArXiv , year=

  62. [62]

    Advances in Neural Information Processing Systems , volume=

    A benchmark for interpretability methods in deep neural networks , author=. Advances in Neural Information Processing Systems , volume=

  63. [63]

    Advances in Neural Information Processing Systems , volume=

    Do input gradients highlight discriminative features? , author=. Advances in Neural Information Processing Systems , volume=

  64. [64]

    arXiv preprint arXiv:1911.03429 , year=

    ERASER: A benchmark to evaluate rationalized NLP models , author=. arXiv preprint arXiv:1911.03429 , year=

    Pith/arXiv arXiv 1911

  65. [65]

    Towards human-centered explainable

    Rong, Yao and Leemann, Tobias and Nguyen, Thai-Trang and Fiedler, Lisa and Qian, Peizhu and Unhelkar, Vaibhav and Seidel, Tina and Kasneci, Gjergji and Kasneci, Enkelejda , journal=. Towards human-centered explainable. 2023 , publisher=

    2023

  66. [66]

    Evaluation metrics in explainable artificial intelligence (

    Coroama, Loredana and Groza, Adrian , booktitle=. Evaluation metrics in explainable artificial intelligence (. 2022 , organization=

    2022

  67. [67]

    From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating explainable

    Nauta, Meike and Trienes, Jan and Pathak, Shreyasi and Nguyen, Elisa and Peters, Michelle and Schmitt, Yasmin and Schl. From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating explainable. ACM Computing Surveys , volume=. 2023 , publisher=

    2023

  68. [68]

    International Conference on Machine Learning , year=

    Explainable k-means and k-medians clustering , author=. International Conference on Machine Learning , year=

  69. [69]

    arXiv preprint arXiv:2503.01087 , year=

    Rashomon Sets for Prototypical-Part Networks: Editing Interpretable Models in Real-Time , author=. arXiv preprint arXiv:2503.01087 , year=

    Pith/arXiv arXiv

  70. [70]

    Explainable

    The (un) reliability of saliency methods , author=. Explainable. 2019 , publisher=

    2019

  71. [71]

    Slack, Dylan and Hilgard, Sophie and Jia, Emily and Singh, Sameer and Lakkaraju, Himabindu , booktitle=. Fooling

  72. [72]

    What do you see?

    Lin, Yi-Shan and Lee, Wen-Chuan and Celik, Z Berkay , booktitle=. What do you see?

  73. [73]

    What does evaluation of explainable artificial intelligence actually tell us?

    Sokol, Kacper and Vogt, Julia E , booktitle=. What does evaluation of explainable artificial intelligence actually tell us?

  74. [74]

    International Conference on Advanced Information Systems Engineering , pages=

    A meta survey of quality evaluation criteria in explanation methods , author=. International Conference on Advanced Information Systems Engineering , pages=. 2022 , organization=

    2022

  75. [75]

    arXiv preprint arXiv:2511.19636 , year=

    Many Ways to be Right: Rashomon Sets for Concept-Based Neural Networks , author=. arXiv preprint arXiv:2511.19636 , year=

    Pith/arXiv arXiv

  76. [76]

    Position: Amazing Things Come From Having Many Good Models , author=

  77. [77]

    Machines we trust: Getting along with artificial intelligence , year=

    A human-centered agenda for intelligible machine learning , author=. Machines we trust: Getting along with artificial intelligence , year=

  78. [78]

    Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    Explanatory interactive machine learning , author=. Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    2019

  79. [79]

    From explainable to interactive

    Raees, Muhammad and Meijerink, Inge and Lykourentzou, Ioanna and Khan, Vassilis-Javed and Papangelis, Konstantinos , journal=. From explainable to interactive. 2024 , publisher=

    2024

  80. [80]

    Proceedings of the 8th international conference on Intelligent user interfaces , pages=

    Interactive machine learning , author=. Proceedings of the 8th international conference on Intelligent user interfaces , pages=

Showing first 80 references.