Pith. sign in

REVIEW 3 major objections 6 minor 55 references

Who Are You Explaining To? A Multi-Agent System for Audience-Aware XAI Narratives

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A Framer–Narrator–Reviewer pipeline turns one fixed SHAP explanation into distinct clinician, patient, and data-scientist narratives, and independent LLM judges assigned every generated narrative to its intended audience with 100%…

desk verdict A well-engineered audience-aware XAI narrative system with a solid intra-narrative evaluation, but the headline preference claim rests on unvalidated LLM judges and needs human evidence or softer claims. read the letter →

arxiv 2608.11033 v1 pith:OXDWKDIY submitted 2026-08-11 cs.MA

Francesco Musicco , Danilo Danese , Giuseppe Fasano , Angela Lombardi , Alberto Carlo Maria Mancino , Tommaso Di Noia This is my paper · ORCID
classification cs.MA
keywords XAISHAPmulti-agentLLMaudience-awareexplanationnarrativegenerationmedicalAILLM-as-a-judgeCard
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

XstrAI is a multi-agent framework that treats a SHAP explanation as fixed evidence and controls how that evidence is communicated to different readers. The paper claims that separating evidence from communication, through an immutable ExplanationCard and three specialized LLM agents, produces narratives that are faithfully grounded in the attribution values while being recognizably written for clinicians, patients, or data scientists. If true, this would make AI explanations safer and more useful in medical settings, where the same prediction must reach people with very different expertise and risk of misunderstanding. The paper reports that XstrAI outperforms single-prompt and generic baselines for clinician and patient audiences in LLM-judged rankings, and that its patient narratives deliberately promote actionable features over strict SHAP rank order.

What carries the argument

The ExplanationCard is the load-bearing artifact: a machine-readable, immutable representation of the prediction case that serves as the single source of truth for all audience pathways. The pipeline then uses three specialized LLM agents per audience: the Framer produces a structured FramingTrace plan, the Narrator converts the card plus trace into natural-language prose, and the Reviewer scores the candidate narrative across five dimensions and either accepts it or routes a bounded revision. This decoupling ensures that differences between narratives come from communicative choices, not from changes in the underlying evidence.

What would settle it

A preregistered human study with at least twenty participants per audience, comparing XstrAI narratives against the strongest baseline for each audience (such as B2-qwen for Clinician and B2-qwen:DS for Data Scientist) on trust, usefulness, and communicative adequacy, with all surface audience markers stripped from the texts; if human preferences do not follow the LLM-judge ranking, the central claim would be falsified.

Watch

Extended reading notes

Core claim

XstrAI's central claim is that audience-aware explanation communication can be engineered as a controlled transformation pipeline rather than left to a single LLM prompt. Each prediction case is encoded in an ExplanationCard that fixes the prediction, SHAP values, feature semantics, and generation constraints identically across all audience pathways. A Framer agent plans how the evidence should be organized and framed for a given stakeholder, a Narrator realizes that plan as prose, and a Reviewer validates the output on five dimensions: card grounding, SHAP faithfulness, epistemic framing, communication safety, and audience fit. When the Reviewer flags an inconsistency, a bounded revision loop routes the fix back to the Framer or Narrator. In the paper's evaluation, all three independent LLM judge families assigned XstrAI's narratives to their intended audience with 100% accuracy, and XstrAI ranked first among 14 systems for the Clinician audience (Elo +2.60) and Patient audience (Elo +6.27), while remaining competitive for the Data Scientist audience. The authors also show that simple role prompting alone (baseline B2) is insufficient to produce semantically tailored patient narratives, and that patient-oriented narratives systematically promote actionable features.

Load-bearing premise

The headline preference results assume that what LLM judges rank highly is what real clinicians, patients, and data scientists would actually prefer, and the human survey, with only five participants per audience and six cases with no baseline narratives, is too small to validate that assumption.

Editorial extensions

If this is right

  • The same SHAP evidence can support markedly different narratives without losing attribution fidelity, so explanation communication can be standardized as an evidence-preserving transformation.
  • Simply adding a role description to a single LLM prompt is not enough to produce audience-appropriate narratives, especially for patients; a separate validation and revision stage appears necessary.
  • Patient-oriented narratives that promote actionable features over strict SHAP rank order may improve patient understanding and follow-up behavior, at the cost of strict feature-order fidelity.
  • A bounded revision loop with at most two iterations is sufficient to converge on accepted narratives across the evaluated cases.
  • Multi-round refinement is most valuable when there is no single dominant style for an audience, as evidenced by the Clinician results where inter-judge agreement was lowest yet XstrAI's margin was largest.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 100% re-identification accuracy may partly reflect LLM judges' sensitivity to surface register markers; a human study with narratives stripped of headings and stylistic cues would test whether the audience distinction survives.
  • The reviewer's five dimensions (grounding, SHAP faithfulness, epistemic framing, communication safety, audience fit) could be reused as a structured rubric for automatic evaluation of XAI narratives beyond this pipeline.
  • If the actionable-feature promotion effect is real, a direct testable extension is whether patients remember and act on promoted features more than features that merely have high SHAP values.
  • The fixed-evidence design should generalize to other explanation techniques and domains, but the paper only demonstrates it on tabular clinical predictions with SHAP; that generalization remains an untested prediction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces XstrAI, a multi-agent framework for generating audience-specific narrative explanations from fixed SHAP evidence. Each prediction case is encoded as an immutable ExplanationCard, and three agents (Framer, Narrator, Reviewer) plan, realize, and validate the narrative under a bounded revision loop. The framework is instantiated on diabetes and stroke prediction tasks and compared against 11 baseline configurations, including a re-implementation of Explingo and incremental ablations B1–B4. Evaluation is split into an intra-narrative regime (form, content, semantics) and an extra-narrative regime (LLM-as-a-judge re-identification and competitive ranking, plus a small human survey). The central claims are that XstrAI produces narratives that are consistently assigned to their intended audience, ranks first by Elo among 14 systems for Clinician and Patient audiences, and maintains competitive performance for Data Scientist, while preserving fidelity to the underlying SHAP evidence. The paper's stated finding that the framework's patient-oriented narratives intentionally suppress raw feature values and promote actionable features is presented as a design feature rather than a fidelity loss.

Significance. The paper addresses a timely and practically important problem: translating the same local explanation evidence into narratives for audiences with very different expertise, risks, and goals. The architectural decomposition into evidence (ExplanationCard), planning (Framer), realization (Narrator), and validation (Reviewer) is transparent and inspectable, and the release of the implementation supports reproducibility. The evaluation is unusually thorough in scope: three judge families from different model families than the generators, Bradley-Terry/Elo aggregation with Friedman and Nemenyi tests, bootstrap confidence intervals, reference-based semantic accuracy, and a separate human survey. There is no evident circularity: the generation pipeline is not fitted to the evaluation metrics, and the judge models are not from the same families as the generators. If the headline preference results survive closer scrutiny, the framework and its evaluation protocol would be a useful contribution to narrative XAI.

major comments (3)
  1. [§5.2.3, Table 7; §4.3.2] The headline claim that XstrAI is 'preferred over all baselines on Clinician and Patient audiences' is supported only by the LLM-as-a-judge Competitive Ranking Analysis. The human evaluation in §5.2.3 uses 5 participants per audience, 6 cases, and no baseline narratives; it measures absolute Likert satisfaction and cannot validate the relative Elo ordering. This is an external-validity gap, but it is load-bearing because the abstract states the preference claim without qualification. I recommend either adding a human head-to-head comparison between XstrAI and the top two baselines per audience, or explicitly restricting the preference claim to the LLM-judge setting in the abstract and conclusions.
  2. [Table 6; §5.2.2] The inter-judge agreement for the Clinician audience is far below the 0.80 reliability threshold: mean Kendall τ values are 0.23–0.33 and ordinal Krippendorff α is 0.34 and 0.45 for diabetes and stroke. Yet §5.2.2 highlights the XstrAI:C lead as its 'widest margin' and interprets the low agreement as 'legitimate diversity'. With this level of disagreement, the pooled Elo scores may average over incompatible preference orderings, and the bootstrap confidence intervals in Figure 3 may understate the true uncertainty. Please report per-judge Elo scores or otherwise analyze whether the Clinician ranking is stable across judge families, and temper the strength of the Clinician claim if it is not.
  3. [Table 3; §5.1.2 and §3.4] The patient-oriented narratives show substantially lower sign agreement than all other configurations: XstrAI:P has SA(D/L) = 0.376/0.698, meaning that under LLM-based extraction roughly 30% of feature directions are inconsistent with the SHAP signs. This is precisely the failure mode the Reviewer is designed to catch (Figure 2 shows a direction inversion being flagged and corrected), and it conflicts with the paper's claims of 'strong empirical alignment' and 'attribution consistency'. The paper should either explain why sign errors persist in the final accepted patient narratives (e.g., intentional suppression or a Reviewer threshold that permits them), or soften the fidelity claims for the patient audience. As it stands, the reader cannot tell whether this is a deliberate trade-off or an unresolved robustness problem.
minor comments (6)
  1. [Tables 2–5] The intra-narrative metrics are reported as point averages over the two datasets without variance estimates, confidence intervals, or significance tests; this makes it difficult to judge whether differences such as the MATTR or Rep-3 advantages of XstrAI are systematic.
  2. [§4.3.1] The semantic reference corpora are described only qualitatively; the paper should report the number of source documents, the number of sentences retained after density-based filtering, and the final per-stakeholder corpus size after downsampling, so that the semantic accuracy values in Table 5 are reproducible.
  3. [§4 (access to code)] The paper mentions both https://github.com/sisinflab/XstrAI and https://anonymous.4open.science/r/XstrAI-5528; the relationship between these two repositories should be clarified, and a canonical link should be provided.
  4. [Figure 3] The x-axis label 'Elo log-strength' is confusing: Elo scores are already log-strength units, but the axis is linear in Elo points; please reword the label to avoid implying a second logarithm.
  5. [§5.2.1] The 100% re-identification accuracy is reported without uncertainty; given 624 trials per judge, a binomial confidence interval would be informative and would help the reader calibrate the strength of the result.
  6. [Table 6 caption] The caption refers to 'the shaded row' falling below the 0.80 reliability cut-off, but the table as rendered has no visible shading; please indicate the row explicitly in the text or mark it with a symbol.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: XstrAI's evaluation is anchored by external judge families, reference corpora, and a human survey, and no claim reduces to its own inputs by construction.

full rationale

The derivation chain is self-contained. XstrAI's generation pipeline is not fitted to the evaluation metrics: the internal Reviewer is a Gemma model with a five-dimension rubric, while the external LLM judges are Claude Sonnet 4.6, GPT-5.5, and Gemini 3.1, and the reference corpora and human survey are external to generation. The re-identification and competitive-ranking tasks measure whether audience-conditioned generation produces distinguishable and preferred outputs, but nothing in the metric construction guarantees those outcomes; the B2 baselines' low semantic accuracy and the patient narratives' low content-agreement scores show the evaluation can and does register failures. The only self-citation (Ref. [30], a prior clinical-ML pipeline by one co-author) appears in a related-work sentence about stakeholder needs and is not load-bearing for any result. The conceptual overlap between the internal Reviewer's 'audience fit' dimension and the extra-narrative audience-appropriateness evaluation is a design convergence, not a circular reduction: the external judges and reference corpora can reject or contradict the internal acceptance, and they do so for several baselines. No equation or metric is defined in terms of the quantity it purports to predict. The paper's own limitation statement about the small human study is an external-validity caveat, not a circular step: the human survey has no baselines and few participants, but that affects generalizability, not the internal logic of the derivation.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim is empirical, so the ledger lists evaluation and design assumptions rather than derivation inputs. There are no fitted parameters in the mathematical sense; the listed free parameters are hand-chosen experimental settings that affect the reported numbers. The axioms concern the validity of SHAP as evidence, representativeness of reference corpora, and validity of LLM judges as proxies for human audiences. No unobserved natural entities are introduced; XstrAI's components are implemented and released.

free parameters (5)
  • revision_budget = 2 rounds
    Maximum revision iterations per pathway; hand-chosen bound that affects narrative quality and is justified only by reference to multi-agent debate stability.
  • reference_retrieval_k = 10
    Number of nearest neighbors used in the semantic accuracy classifier; no sensitivity analysis is reported.
  • density_zscore_cutoff = +/-2
    Outlier threshold for reference sentence embeddings; hand-chosen, with no sensitivity analysis.
  • minority_undersample_size = 536 (diabetes), 498 (stroke)
    Random undersampling to these sizes addresses class imbalance and affects all reported scores.
  • top_k_features = not stated in text
    The ExplanationCard appears to select top-k features, but k is not specified in the manuscript, leaving a hidden experimental setting.
assumptions (5)
  • domain assumption SHAP values from TreeExplainer on the Random Forest models correctly represent the models' local behavior and are a valid evidence basis for narratives.
    The entire pipeline treats SHAP attributions as fixed ground truth; the paper relies on prior SHAP theory without verifying it on these specific models. Location: Section 4.1.
  • domain assumption LLM judge rankings are a valid proxy for what each human audience considers a good narrative.
    Primary evidence for preference claims; not calibrated against human judgments because the human survey is small and baseline-free. Location: Sections 4.3.2 and 5.2.3.
  • domain assumption Reference corpora from PubMed, IEEE Xplore, and MedlinePlus are representative of clinician, data scientist, and patient language respectively.
    Semantic accuracy depends on these corpora defining the target audience spaces. Location: Section 4.3.1.
  • domain assumption Sentence-BERT cosine nearest-neighbor mixtures measure audience appropriateness.
    The semantic metric classifies a narrative by the audience of its neighbors; this assumes embedding geometry tracks communicative register. Location: Section 4.3.1.
  • standard math Statistical assumptions for Friedman, Nemenyi, and Bradley-Terry models hold for the ranking data.
    Used for significance and Elo computations. Location: Section 4.3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Who Are You Explaining To? A Multi-Agent System for Audience-Aware XAI Narratives." pith.science (2026). https://pith.science/paper/OXDWKDIY

@misc{pith2026260811033,
  author       = {Pith},
  title        = {Pith review of: Who Are You Explaining To? A Multi-Agent System for Audience-Aware XAI Narratives},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OXDWKDIY}},
  note         = {Machine review of arXiv:2608.11033}
}
read the original abstract

Feature-attribution methods such as SHAP provide useful evidence about individual model predictions, but their numerical outputs are rarely sufficient for audiences with different expertise, goals, and risks of misinterpretation. In medical AI, the same local explanation must reach patients, clinicians, and data scientists through markedly different forms of communication, and naive verbalization through large language models (LLMs) is prone to weak grounding, conflation of attribution with causal language, and outputs that are persuasive without being faithful to the underlying model evidence. We introduce XstrAI, an audience-aware multi-agent framework that treats local explanations as fixed evidence and structures how it is communicated to each target reader. Each prediction case is encoded as an immutable structured representation, shared identically across audiences so the underlying evidence remains fixed. Generation is factored into three specialized LLM agents responsible for audience-aware planning, linguistic realization, and validation for grounding, attribution consistency, communicative risk, and audience appropriateness, with a bounded revision loop triggered on detected inconsistencies. We evaluate XstrAI on diabetes and stroke risk prediction against 11 baselines, ranging from direct verbalization to a re-implementation of a state-of-the-art narrator. The evaluation combines an intra-narrative regime measuring fidelity to SHAP evidence with an extra-narrative regime assessing audience appropriateness through reference corpora, multi-family LLM judges, and a survey with target readers. In both evaluations, XstrAI's narratives are consistently assigned to their intended audience by independent judges, and preferred over all baselines on Clinician and Patient audiences, with competitive performance on Data Scientist, where audience-conditioned single-prompt baselines lead.

Figures

Figures reproduced from arXiv: 2608.11033 by the authors.

Figure 1
Figure 1. XstrAI architecture. Each case is represented as a structured [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The Reviewer in action. Top: skeleton of the audience-conditioned Reviewer prompt; the same five dimen￾sions are instantiated for Clinician, Patient, and Data Scien￾tist. Bottom: a real Reviewer output that catches a SHAP direc￾tion inversion in a patient narrative and routes the revision back to the Narrator, preserving the existing FramingTrace. were implemented in Python and executed on an Ubuntu 22.04 workstatio… view at source ↗
Figure 3
Figure 3. Competitive ranking analysis: Elo log-strength (zero-centered; 0 = average system) across three audiences, pooled over [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 36 canonical work pages

  1. [1]

    Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Túlio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. InCHI. ACM, 81:1–81:16

    work page 2021

  2. [2]

    Ahsan Bilal, David Ebert, and Beiyu Lin. 2025. Llms for explainable ai: A com- prehensive survey.arXiv preprint arXiv:2504.00125(2025)

    arXiv 2025

  3. [3]

    RALPH ALLAN BRADLEY and MILTON E. TERRY. 1952. RANK ANALYSIS OF INCOMPLETE BLOCK DESIGNS: THE METHOD OF PAIRED COMPARISONS.Biometrika39, 3-4 (12 1952), 324–345. arXiv:https://academic.oup.com/biomet/article-pdf/39/3-4/324/930466/39- 3-4-324.pdf doi:10.1093/biomet/39.3-4.324

    doi:10.1093/biomet/39.3-4.324 1952

  4. [4]

    David Cheng-Han Chiang and Hung-yi Lee. 2023. Can Large Language Models Be an Alternative to Human Evaluations?. InACL (1). Association for Computational Linguistics, 15607–15631

    work page 2023

  5. [5]

    Covington and Joe D

    Michael A. Covington and Joe D. McFall. 2010. Cutting the Gordian Knot: The Moving-Average Type-Token Ratio (MATTR).J. Quant. Linguistics17, 2 (2010), 94–100

    work page 2010

  6. [6]

    Pedro Henrique Luz de Araujo, Paul Röttger, Dirk Hovy, and Benjamin Roth. 2025. Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance. InEMNLP. Association for Computational Linguistics, 26857–26886

    work page 2025

  7. [7]

    Janez Demsar. 2006. Statistical Comparisons of Classifiers over Multiple Data Sets.J. Mach. Learn. Res.7 (2006), 1–30

    2006

  8. [8]

    Tenenbaum, and Igor Mor- datch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mor- datch. 2024. Improving Factuality and Reasoning in Language Models through Multiagent Debate. InICML (Proceedings of Machine Learning Research). PMLR / OpenReview.net, 11733–11763

    work page 2024

Show all 55 references
  1. [9]

    Latifa Dwiyanti, Sergio Ryan Wibisono, and Hidetaka Nambo. 2025. Contextual- SHAP : Enhancing SHAP Explanations Through Contextual Language Genera- tion. InWSSE. ACM, 231–237

    2025

  2. [10]

    Vera Liao, Michael J

    Upol Ehsan, Q. Vera Liao, Michael J. Muller, Mark O. Riedl, and Justin D. Weisz

  3. [11]

    An Yang et al. 2025. Qwen3 Technical Report.arXiv preprint arXiv:2505.09388 (2025)

    2025 arXiv

  4. [12]

    Rudolph Flesch. 1948. A new readability yardstick.Journal of applied psychology 32, 3 (1948), 221

    1948

  5. [13]

    Montgomery Flora, Corey Potvin, Amy McGovern, and Shawn Handler. 2022. Comparing explanation methods for traditional machine learning models part 1: an overview of current methods and quantifying their disagreement.arXiv preprint arXiv:2211.08943(2022)

    2022 arXiv

  6. [14]

    Zihao Fu, Wai Lam, Anthony Man-Cho So, and Bei Shi. 2021. A Theoretical Analysis of the Repetition Problem in Text Generation. InAAAI. AAAI Press, 12848–12856

    2021

  7. [15]

    Google DeepMind. 2026. Gemma 4 Model Card. https://ai.google.dev/gemma/ docs/core/model_card_4

    2026

  8. [16]

    Kilem Li Gwet. 2008. Computing Inter-Rater Reliability and Its Variance in the Presence of High Agreement.Brit. J. Math. Statist. Psych.61, 1 (2008), 29–48. doi:10.1348/000711006X126600

    2008 doi

  9. [17]

    Sandra G Hart and Lowell E Staveland. 1988. Development of NASA-TLX (Task Load Index): Results of empirical and theoretical research. InAdvances in psy- chology. Vol. 52. Elsevier, 139–183

    1988

  10. [18]

    Yifan He and David Martens. 2026. An Agentic Approach to Generating XAI- Narratives.arXiv preprint arXiv:2603.20003(2026)

    2026

  11. [19]

    Robert R Hoffman, Shane T Mueller, Gary Klein, and Jordan Litman. 2023. Mea- sures for explainable AI: Explanation goodness, user satisfaction, mental models, curiosity, trust, and human-AI performance.Frontiers in Computer Science5 (2023), 1096257

    2023

  12. [20]

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. 2024. MetaGPT: Meta Programming for A Multi-Agent Collaborat...

    2024

  13. [21]

    David R. Hunter. 2003. MM algorithms for generalized Bradley-Terry models. Annals of Statistics32 (2003), 384–406. https://api.semanticscholar.org/CorpusID: 17436309

    2003

  14. [22]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation.ACM Comput. Surv.55, 12 (2023), 248:1–248:38

    2023

  15. [23]

    M. G. KENDALL. 1938. A NEW MEASURE OF RANK CORRELATION.Biometrika 30, 1-2 (06 1938), 81–93. arXiv:https://academic.oup.com/biomet/article-pdf/30/1- 2/81/423380/30-1-2-81.pdf doi:10.1093/biomet/30.1-2.81

    1938 doi

  16. [24]

    Minjung Kim, Saebyeol Kim, Jinwoo Kim, Tae-Jin Song, and Yuyoung Kim

  17. [25]

    klaus krippendorff. 2011. Computing Krippendorff’s Alpha-Reliability. (01 2011)

    2011

  18. [26]

    Satyapriya Krishna, Tessa Han, Alex Gu, Steven Wu, Shahin Jabbari, and Himabindu Lakkaraju. 2024. The Disagreement Problem in Explainable Ma- chine Learning: A Practitioner’s Perspective.Trans. Mach. Learn. Res.2024 (2024)

    2024

  19. [27]

    Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023. CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society. InNeurIPS

    2023

  20. [28]

    Vera Liao and Kush R

    Q. Vera Liao and Kush R. Varshney. 2021. Human-Centered Explainable AI (XAI): From Algorithms to User Experiences.CoRRabs/2110.10790 (2021)

    2021 arXiv

  21. [29]

    Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment. InEMNLP. Association for Computational Linguistics, 2511–2522

    2023

  22. [30]

    Angela Lombardi, Francesca Arezzo, Eugenio Di Sciascio, Carmelo Ardito, Michele Mongelli, Nicola Di Lillo, Fabiana Divina Fascilla, Erica Silvestris, Anila Kardhashi, Carmela Putino, et al. 2023. A human-interpretable machine learning pipeline based on ultrasound to support le...

    2023

  23. [31]

    Luca Longo, Mario Brcic, Federico Cabitza, Jaesik Choi, Roberto Confalonieri, Javier Del Ser, Riccardo Guidotti, Yoichi Hayashi, Francisco Herrera, Andreas Holzinger, Richard Jiang, Hassan Khosravi, Freddy Lécué, Gianclaudio Malgieri, Andrés Páez, Wojciech Samek, Johannes Schn...

    2024

  24. [32]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. 2017. A Unified Approach to Interpreting Model Predictions. InNIPS. 4765–4774

    2017

  25. [33]

    David Martens, James Hinns, Camille Dams, Mark Vergouwen, and Theodoros Evgeniou. 2025. Tell me a story! Narrative-driven XAI with Large Language Models.Decis. Support Syst.191 (2025), 114402

    2025

  26. [34]

    Tim Miller. 2019. Explanation in artificial intelligence: Insights from the social sciences.Artif. Intell.267 (2019), 1–38

    2019

  27. [35]

    Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. 2019. Model Cards for Model Reporting. InFAT. ACM, 220–229

    2019

  28. [36]

    Meike Nauta, Jan Trienes, Shreyasi Pathak, Elisa Nguyen, Michelle Peters, Yasmin Schmitt, Jörg Schlötterer, Maurice van Keulen, and Christin Seifert. 2023. From Anecdotal Evidence to Quantitative Evaluation Methods: A Systematic Review on Evaluating Explainable AI.ACM Comput. ...

    2023

  29. [37]

    Bowman, and Shi Feng

    Arjun Panickssery, Samuel R. Bowman, and Shi Feng. 2024. LLM Evaluators Recognize and Favor Their Own Generations. InNeurIPS

    2024

  30. [38]

    O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S

    Joon Sung Park, Joseph C. O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. InUIST. ACM, 2:1–2:22

    2023

  31. [39]

    Preece, Dan Harborne, Dave Braines, Richard Tomsett, and Supriyo Chakraborty

    Alun D. Preece, Dan Harborne, Dave Braines, Richard Tomsett, and Supriyo Chakraborty. 2018. Stakeholders in Explainable AI.CoRRabs/1810.00184 (2018)

    2018 arXiv

  32. [40]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Confer- ence on Natural Language Processing (EMNLP-IJ...

    2019 doi

  33. [41]

    Qiyang Sun, Alican Akman, and Björn W Schuller. 2025. Explainable artificial intelligence for medical applications: A review.ACM Transactions on Computing for Healthcare6, 2 (2025), 1–31

    2025

  34. [42]

    Xiangru Tang, Anni Zou, Zhuosheng Zhang, Ziming Li, Yilun Zhao, Xingyao Zhang, Arman Cohan, and Mark Gerstein. 2024. MedAgents: Large Language Models as Collaborators for Zero-shot Medical Reasoning. InACL (Findings) (Findings of ACL). Association for Computational Linguistics...

    2024

  35. [43]

    Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. 2023. Lan- guage Models Don’t Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting. InNeurIPS

    2023

  36. [44]

    Vittoria Vineis, Matteo Silvestri, Lorenzo Antonelli, Filippo Betello, and Gabriele Tolomei. 2026. PONTE: Personalized Orchestration for Natural Language Trust- worthy Explanations.CoRRabs/2603.06485 (2026)

    2026

  37. [45]

    Noah Wang, Zhongyuan Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, Man Zhang, Zhaoxiang Zhang, Wanli Ouyang, Ke Xu, Wenhao Huang, Jie Fu, and Junran Peng

  38. [46]

    Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Lingpeng Kong, Qi Liu, Tianyu Liu, and Zhifang Sui. 2024. Large Language Mod- els are not Fair Evaluators. InACL (1). Association for Computational Linguistics, 9440–9450

    2024

  39. [47]

    Xintao Wang, Yunze Xiao, Jen-tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, Jiangjie Chen, Cheng Li, and Yanghua Xiao. 2024. InCharacter: Evaluating Personality Fidelity in Role-Playing Agents Musicco et al. through Psychological Interview...

    2024

  40. [48]

    InACL (Findings) (Findings of ACL)

    RoleLLM: Benchmarking, Eliciting, and Enhancing Role-Playing Abilities of Large Language Models. InACL (Findings) (Findings of ACL). Association for Computational Linguistics, 14743–14777

  41. [49]

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert.arXiv preprint arXiv:1904.09675(2019)

    2019 arXiv

  42. [50]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-Judge with MT- Bench and Chatbot Arena. InNeurIPS

    2023

  43. [51]

    Matej Zecevic, Moritz Willig, Devendra Singh Dhami, and Kristian Kersting

  44. [55]

    Alexandra Zytek, Sara Pidò, Sarah Alnegheimish, Laure Berti-Équille, and Kalyan Veeramachaneni. 2024. Explingo: Explaining AI Predictions using Large Lan- guage Models. InIEEE Big Data. IEEE, 1197–1208

    2024

  45. [2021]

    Expanding Explainability: Towards Social Transparency in AI systems. In CHI. ACM, 82:1–82:19

  46. [2023]

    Causal Parrots: Large Language Models May Talk Causality But Are Not Causal.Trans. Mach. Learn. Res.2023 (2023)

    2023

  47. [2024]

    Do stakeholder needs differ? - Designing stakeholder-tailored Explainable Artificial Intelligence (XAI) interfaces.Int. J. Hum. Comput. Stud.181 (2024), 103160

    2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.