Research Papers

Bridging the Gap in XAI-Why Reliable Metrics Matter for Explainability and Compliance

Pratinav Seth, Vinay Kumar Sankarapu

Reliable explainability is not only a technical goal but also a cornerstone of private AI governance. As AI models enter high-stakes sectors, private actors such as auditors, insurers, certification bodies, and procurement agencies require standardized evaluation metrics to assess trustworthiness. However, current XAI evaluation metrics remain fragmented and prone to manipulation, which undermines accountability and compliance. We argue that standardized metrics can function as governance primitives, embedding auditability and accountability within AI systems for effective private oversight. Building upon prior work in XAI benchmarking, we identify key limitations in ensuring faithfulness, tamper resistance, and regulatory alignment. Furthermore, interpretability can directly support model alignment by providing a verifiable means of ensuring behavioral integrity in General Purpose AI (GPAI) systems. This connection between interpretability and alignment positions XAI metrics as both technical and regulatory instruments that help prevent alignment faking, a growing concern among oversight bodies. We propose a Governance by Metrics paradigm that treats explainability evaluation as a central mechanism of private AI governance. Our framework introduces a hierarchical model linking transparency, tamper resistance, scalability, and legal alignment, extending evaluation from model introspection toward systemic accountability. Through conceptual synthesis and alignment with governance standards, we outline a roadmap for integrating explainability metrics into continuous AI assurance pipelines that serve both private oversight and regulatory needs.

Cannot or Should Not? Automatic Analysis of Refusal Composition in IFT/RLHF Datasets and Refusal Behavior of Black-Box LLMs

Alexander von Recum, Christoph Schnabl, Gabor Hollbeck +3

Refusals - instances where large language models (LLMs) decline or fail to fully execute user instructions - are crucial for both AI safety and AI capabilities and the reduction of hallucinations in particular. These behaviors are learned during post-training, especially in instruction fine-tuning (IFT) and reinforcement learning from human feedback (RLHF). However, existing taxonomies and evaluation datasets for refusals are inadequate, often focusing solely on should-not-related (instead of cannot-related) categories, and lacking tools for auditing refusal content in black-box LLM outputs. We present a comprehensive framework for classifying LLM refusals: (a) a taxonomy of 16 refusal categories, (b) a human-annotated dataset of over 8,600 instances from publicly available IFT and RLHF datasets, (c) a synthetic dataset with 8,000 examples for each refusal category, and (d) classifiers trained for refusal classification. Our work enables precise auditing of refusal behaviors in black-box LLMs and automatic analyses of refusal patterns in large IFT and RLHF datasets. This facilitates the strategic adjustment of LLM refusals, contributing to the development of more safe and reliable LLMs.

Human experts are often engaged in the development of machine learning systems to collect and validate data, consult on algorithm development, and evaluate system performance. At the same time, who counts as an 'expert' and what constitutes 'expertise' is not always explicitly defined. In this work, we review 112 academic publications that explicitly reference 'expert' and 'expertise' and that describe the development of machine learning (ML) systems to survey how expertise is characterized and the role experts play. We find that expertise is often undefined and forms of knowledge outside of formal education and professional certification are rarely sought, which has implications for the kinds of knowledge that are recognized and legitimized in ML development. Moreover, we find that expert knowledge tends to be utilized in ways focused on mining textbook knowledge, such as through data annotation. We discuss the ways experts are engaged in ML development in relation to deskilling, the social construction of expertise, and implications for responsible AI development. We point to a need for reflection and specificity in justifications of domain expert engagement, both as a matter of documentation and reproducibility, as well as a matter of broadening the range of recognized expertise.

Do Responsible AI Artifacts Advance Stakeholder Goals? Four Key Barriers Perceived by Legal and Civil Stakeholders

Anna Kawakami, Daricia Wilkinson, Alexandra Chouldechova

The responsible AI (RAI) community has introduced numerous processes and artifacts (e.g., Model Cards, Transparency Notes, Data Cards) to facilitate transparency and support the governance of AI systems. While originally designed to scaffold and document AI development processes in technology companies, these artifacts are becoming central components of regulatory compliance under recent regulations such as the EU AI Act. Much prior work has explored the design of new RAI artifacts or their use by practitioners within technology companies. However, as RAI artifacts begin to play key roles in enabling external oversight, it becomes critical to understand how stakeholders--particularly those situated outside of technology companies who govern and audit industry AI deployments--perceive the efficacy of RAI artifacts. In this study, we conduct semi-structured interviews and design activities with 19 government, legal, and civil society stakeholders who inform policy and advocacy around responsible AI efforts. While participants believe that RAI artifacts are a valuable contribution to the broader AI governance ecosystem, many are concerned about their potential unintended, longer-term impacts on actors outside of technology companies (e.g., downstream end-users, policymakers, civil society stakeholders). We organize these beliefs into four barriers that help explain how RAI artifacts may (inadvertently) reconfigure power relations across civil society, government, and industry, impeding civil society and legal stakeholders' ability to protect downstream end-users from potential AI harms. Participants envision how structural changes, along with changes in how RAI artifacts are designed, used, and governed, could help redirect the role of artifacts to support more collaborative and proactive external oversight of AI systems. We discuss research and policy implications for RAI artifacts.

The rapid advancement of AI technologies yields numerous future impacts on individuals and society. Policymakers are tasked to react quickly and establish policies that mitigate those impacts. However, anticipating the effectiveness of policies is a difficult task, as some impacts might only be observable in the future and respective policies might not be applicable to the future development of AI. In this work we develop a method for using large language models (LLMs) to evaluate the efficacy of a given piece of policy at mitigating specified negative impacts. We do so by using GPT-4 to generate scenarios both pre- and post-introduction of policy and translating these vivid stories into metrics based on human perceptions of impacts. We leverage an already established taxonomy of impacts of generative AI in the media environment to generate a set of scenario pairs both mitigated and non-mitigated by the transparency policy in Article 50 of the EU AI Act. We then run a user study (n=234) to evaluate these scenarios across four risk-assessment dimensions: severity, plausibility, magnitude, and specificity to vulnerable populations. We find that this transparency legislation is perceived to be effective at mitigating harms in areas such as labor and well-being, but largely ineffective in areas such as social cohesion and security. Through this case study we demonstrate the efficacy of our method as a tool to iterate on the effectiveness of policy for mitigating various negative impacts. We expect this method to be useful to researchers or other stakeholders who want to brainstorm the potential utility of different pieces of policy or other mitigation strategies.

With the rise of complex cyber devices Cyber Forensics (CF) is facing many new challenges. For example, there are dozens of systems running on smartphones, each with more than millions of downloadable applications. Sifting through this large amount of data and making sense requires new techniques, such as from the field of Artificial Intelligence (AI). To apply these techniques successfully in CF, we need to justify and explain the results to the stakeholders of CF, such as forensic analysts and members of the court, for them to make an informed decision. If we want to apply AI successfully in CF, there is a need to develop trust in AI systems. Some other factors in accepting the use of AI in CF are to make AI authentic, interpretable, understandable, and interactive. This way, AI systems will be more acceptable to the public and ensure alignment with legal standards. An explainable AI (XAI) system can play this role in CF, and we call such a system XAI-CF. XAI-CF is indispensable and is still in its infancy. In this paper, we explore and make a case for the significance and advantages of XAI-CF. We strongly emphasize the need to build a successful and practical XAI-CF system and discuss some of the main requirements and prerequisites of such a system. We present a formal definition of the terms CF and XAI-CF and a comprehensive literature review of previous works that apply and utilize XAI to build and increase trust in CF. We discuss some challenges facing XAI-CF. We also provide some concrete solutions to these challenges. We identify key insights and future research directions for building XAI applications for CF. This paper is an effort to explore and familiarize the readers with the role of XAI applications in CF, and we believe that our work provides a promising basis for future researchers interested in XAI-CF.

Language-Guided World Models: A Model-Based Approach to AI Control

Alex Zhang, Khanh Nguyen, Jens Tuyls +2

This paper introduces the concept of Language-Guided World Models (LWMs) -- probabilistic models that can simulate environments by reading texts. Agents equipped with these models provide humans with more extensive and efficient control, allowing them to simultaneously alter agent behaviors in multiple tasks via natural verbal communication. In this work, we take initial steps in developing robust LWMs that can generalize to compositionally novel language descriptions. We design a challenging world modeling benchmark based on the game of MESSENGER (Hanjie et al., 2021), featuring evaluation settings that require varying degrees of compositional generalization. Our experiments reveal the lack of generalizability of the state-of-the-art Transformer model, as it offers marginal improvements in simulation quality over a no-text baseline. We devise a more robust model by fusing the Transformer with the EMMA attention mechanism (Hanjie et al., 2021). Our model substantially outperforms the Transformer and approaches the performance of a model with an oracle semantic parsing and grounding capability. To demonstrate the practicality of this model in improving AI safety and transparency, we simulate a scenario in which the model enables an agent to present plans to a human before execution, and to revise plans based on their language feedback.

Saliency-Bench: A Comprehensive Benchmark for Evaluating Visual Explanations

Yifei Zhang, James Song, Siyi Gu +4

Explainable AI (XAI) has gained significant attention for providing insights into the decision-making processes of deep learning models, particularly for image classification tasks through visual explanations visualized by saliency maps. Despite their success, challenges remain due to the lack of annotated datasets and standardized evaluation pipelines. In this paper, we introduce Saliency-Bench, a novel benchmark suite designed to evaluate visual explanations generated by saliency methods across multiple datasets. We curated, constructed, and annotated eight datasets, each covering diverse tasks such as scene classification, cancer diagnosis, object classification, and action classification, with corresponding ground-truth explanations. The benchmark includes a standardized and unified evaluation pipeline for assessing faithfulness and alignment of the visual explanation, providing a holistic visual explanation performance assessment. We benchmark these eight datasets with widely used saliency methods on different image classifier architectures to evaluate explanation quality. Additionally, we developed an easy-to-use API for automating the evaluation pipeline, from data accessing, and data loading, to result evaluation. The benchmark is available via our website: https://xaidataset.github.io.

Feature Importance versus Feature Influence and What It Signifies for Explainable AI

Kary Främling

When used in the context of decision theory, feature importance expresses how much changing the value of a feature can change the model outcome (or the utility of the outcome), compared to other features. Feature importance should not be confused with the feature influence used by most state-of-the-art post-hoc Explainable AI methods. Contrary to feature importance, feature influence is measured against a reference level or baseline. The Contextual Importance and Utility (CIU) method provides a unified definition of global and local feature importance that is applicable also for post-hoc explanations, where the value utility concept provides instance-level assessment of how favorable or not a feature value is for the outcome. The paper shows how CIU can be applied to both global and local explainability, assesses the fidelity and stability of different methods, and shows how explanations that use contextual importance and contextual utility can provide more expressive and flexible explanations than when using influence only.

No Fair Lunch: A Causal Perspective on Dataset Bias in Machine Learning for Medical Imaging

Charles Jones, Daniel C. Castro, Fabio De Sousa Ribeiro +3

As machine learning methods gain prominence within clinical decision-making, addressing fairness concerns becomes increasingly urgent. Despite considerable work dedicated to detecting and ameliorating algorithmic bias, today's methods are deficient with potentially harmful consequences. Our causal perspective sheds new light on algorithmic bias, highlighting how different sources of dataset bias may appear indistinguishable yet require substantially different mitigation strategies. We introduce three families of causal bias mechanisms stemming from disparities in prevalence, presentation, and annotation. Our causal analysis underscores how current mitigation methods tackle only a narrow and often unrealistic subset of scenarios. We provide a practical three-step framework for reasoning about fairness in medical imaging, supporting the development of safe and equitable AI prediction models.

Explaining and visualizing black-box models through counterfactual paths

Bastian Pfeifer, Mateusz Krzyzinski, Hubert Baniecki +3

Explainable AI (XAI) is an increasingly important area of machine learning research, which aims to make black-box models transparent and interpretable. In this paper, we propose a novel approach to XAI that uses the so-called counterfactual paths generated by conditional permutations of features. The algorithm measures feature importance by identifying sequential permutations of features that most influence changes in model predictions. It is particularly suitable for generating explanations based on counterfactual paths in knowledge graphs incorporating domain knowledge. Counterfactual paths introduce an additional graph dimension to current XAI methods in both explaining and visualizing black-box models. Experiments with synthetic and medical data demonstrate the practical applicability of our approach.

Convolutional Neural Networks (CNNs) are frequently and successfully used in medical prediction tasks. They are often used in combination with transfer learning, leading to improved performance when training data for the task are scarce. The resulting models are highly complex and typically do not provide any insight into their predictive mechanisms, motivating the field of "explainable" artificial intelligence (XAI). However, previous studies have rarely quantitatively evaluated the "explanation performance" of XAI methods against ground-truth data, and transfer learning and its influence on objective measures of explanation performance has not been investigated. Here, we propose a benchmark dataset that allows for quantifying explanation performance in a realistic magnetic resonance imaging (MRI) classification task. We employ this benchmark to understand the influence of transfer learning on the quality of explanations. Experimental results show that popular XAI methods applied to the same underlying model differ vastly in performance, even when considering only correctly classified examples. We further observe that explanation performance strongly depends on the task used for pre-training and the number of CNN layers pre-trained. These results hold after correcting for a substantial correlation between explanation and classification performance.