rationale generation

AI deployment in high-stakes domains demands explainability to foster trust and accountability, yet prevailing XAI methods often neglect foundational human cognitive mechanisms. This paper introduces the first XAI framework systematically integrating Malle’s five-category model of human explanatory reasoning—namely, knowledge structure, simulation/prediction, covariation, direct recall, and rationalization—by unifying attribution analysis, feature importance, attention visualization, and large language model–generated reasoning chains. The resulting multimodal explanation framework aligns technical outputs with empirically grounded human explanation preferences. Empirical evaluation in real-world credit risk assessment and regulatory compliance tasks demonstrates significant improvements in users’ depth of understanding and perceived trustworthiness of AI decisions. The core contribution lies in pioneering a cognition-informed design paradigm for explainability—shifting XAI from mere *technical interpretability* toward *human comprehensibility*, *acceptability*, and *reliability*.

Existing interpretability research in machine learning predominantly focuses on input-output mapping mechanisms, neglecting the functional role of explanations within concrete application contexts—such as clinical decision support, model debugging, or remedial intervention—leading to potential misuse. Method: This paper proposes a “use-case-driven” paradigm grounded in statistical decision theory, establishing a quantifiable analytical framework for explanation utility: it formally defines the maximum performance gain an explanation can yield in a given task and characterizes its theoretical utility upper bound, while unifying evaluation criteria across diverse application scenarios. Contribution/Results: The framework shifts interpretability research from qualitative description toward a rigorous, analyzable, verifiable, and reproducible scientific paradigm. It significantly enhances the reliability and practical utility of explanation methods in real-world decision-making environments.

Contemporary machine learning models’ complexity poses significant trust, regulatory, and ethical risks, yet existing explainability guidelines lack operational specificity. To address this gap, we conducted a controlled experiment with 124 developers, integrating cognitive process theory and sociological imagination to investigate how policy frameworks influence the design of end-user–oriented explanations for diabetic retinopathy screening models. Results reveal that all participants struggled to generate high-quality, policy-compliant, and empirically verifiable explanations; over 70% failed to accurately anticipate users’ comprehension barriers; and widely adopted technical methods (e.g., SHAP, Anchors) exhibit fundamental misalignment with real-world stakeholder needs. Our core contribution is identifying *developers’ inability to empathize with non-technical stakeholders* as the central mechanism underlying explanation failure—and proposing, for the first time, an empathy-centered educational intervention framework to bridge this gap.

Current explainable reinforcement learning (XRL) lacks objective, actionable metrics for evaluating explanation quality; prevailing approaches rely on subjective human judgments, impeding empirical validation and cross-study comparability. Method: We propose the first objective, behavior-based human evaluation paradigm tailored to debugging and human-AI collaboration tasks. Grounded in observable behavioral outcomes—such as task completion rate, error-correction efficiency, and collaborative response latency—we conduct controlled behavioral tracking experiments in a customized grid-world environment. Contribution/Results: This paradigm overcomes the limitations of subjective assessment by directly linking explanation quality to measurable human performance. Empirical results demonstrate its superior reliability in characterizing explanation efficacy, significantly enhancing reproducibility, cross-method comparability, and epistemic rigor. It establishes the first scientifically grounded, standardized human evaluation benchmark for XRL.

Reinforcement learning (RL) policies suffer from poor interpretability due to their sequential decision-making nature and the loss of qualitative reward attribution during training, hindering non-expert understanding and intervention. To address this, we propose “experiential explanation,” a novel paradigm that jointly trains a policy network and an influence predictor to explicitly model causal relationships between states and rewards, generating counterfactual-based, human-understandable explanations. Our approach is the first to systematically recover reward-source attribution—previously discarded in RL training—enabling users to predict and intervene in agent behavior. In two human-subject evaluations, our method significantly improves user accuracy in predicting agent actions and consistently outperforms existing baselines across five dimensions: understandability, completeness, satisfaction, practicality, and explanatory accuracy.

This study addresses the challenge that existing algorithmic explanations are often poorly understood and misapplied by non-expert users due to semantic ambiguity and insufficient contextual information, leading to a disconnect between explanations and actual decision-making. To bridge this gap, the authors propose an “Explanation Card” framework that augments widely used interpretability methods—such as SHAP and counterfactual explanations—with structured metadata specifying their applicability boundaries, robustness properties, and user-oriented interpretation guidance. By shifting explanatory responsibility from end users to explanation providers, this approach enhances the practical utility and regulatory compliance of model explanations, aligning with the transparency requirements of the EU AI Act. Empirical evaluations demonstrate that Explanation Cards significantly improve users’ comprehension accuracy of complex model explanations and effectively flag scenarios where explanations are unreliable, thereby facilitating trustworthy real-world deployment of algorithmic systems.

Current machine learning evaluation practices predominantly rely on surface-level performance metrics, often neglecting the internal mechanisms of models. This work proposes trustworthy interpretability as a central evaluation paradigm and, for the first time, systematically demonstrates that it satisfies core criteria from the philosophy of science—namely falsifiability, reproducibility, and predictive power. By constructing an evaluation framework that integrates causal analysis with mechanistic probing, the study delineates three functional pathways through which interpretability enables the identification of behavioral origins, detection of latent flaws, and prediction of potential failure modes. This approach advances model assessment beyond performance-oriented benchmarks toward a deeper understanding of underlying mechanisms.

This study addresses the persistent gap in current automated decision-making systems, which often provide explanations without effectively convincing recipients of the legitimacy of their decisions. Distinguishing between “explanation”—clarifying how a decision was generated—and “justification,” which offers normatively acceptable reasons, this work introduces Habermas’s theory of communicative action and Perelman’s new rhetoric into explainable artificial intelligence (XAI) research for the first time. It proposes a recipient-centered framework for analyzing justification, grounded in these theoretical foundations. The framework’s theoretical and practical utility is demonstrated through a case study of university admissions in France, revealing how different forms of explanation variably support justificatory claims. This approach offers a novel paradigm for developing trustworthy automated decision systems that prioritize not only transparency but also normative acceptability from the perspective of affected stakeholders.

This study addresses the persistent challenge that explainable AI (XAI) often fails to effectively support human decision-making due to poor user comprehension. To bridge this gap, the work integrates cognitive modeling with user studies to formally represent—within a computationally tractable framework—the reasoning strategies humans employ when interacting with different XAI methods in structured data tasks. Through formative and summative user experiments, feature attribution analyses, and behavioral alignment evaluations, the resulting cognitive model demonstrates significantly greater accuracy than conventional machine learning surrogates in capturing human forward-simulation decision behavior. Beyond elucidating which XAI mechanisms genuinely aid human judgment, the model offers an empirically grounded foundation for designing more effective XAI systems and serves as a high-fidelity, efficient proxy for human-subject experimentation in XAI research.

This work investigates whether chain-of-thought (CoT) explanations generated by large reasoning models exhibit cross-model generalization—specifically, their ability to induce consistent behaviors across different models. We present the first systematic evaluation of behavioral consistency driven by CoT explanations across models, propose a sentence-level ensemble strategy to enhance such consistency, and analyze its relationship with human preferences and post-training via reinforcement learning. Our experiments demonstrate that CoT explanations generally improve behavioral alignment among diverse models, and this alignment shows a significant positive correlation with both human preference ratings and the effectiveness of reinforcement learning-based post-training. Furthermore, the proposed ensemble method effectively boosts the generalization capability of CoT explanations across models.