Research Papers

Toward Personalized XAI: A Case Study in Intelligent Tutoring Systems

Cristina Conati, Oswald Barral, Vanessa Putnam +1

Our research is a step toward ascertaining the need for personalization, in XAI, and we do so in the context of investigating the value of explanations of AI-driven hints and feedback are useful in Intelligent Tutoring Systems (ITS). We added an explanation functionality for the adaptive hints provided by the Adaptive CSP (ACSP) applet, an interactive simulation that helps students learn an algorithm for constraint satisfaction problems by providing AI-driven hints adapted to their predicted level of learning. We present the design of the explanation functionality and the results of a controlled study to evaluate its impact on students' learning and perception of the ACPS hints. The study includes an analysis of how these outcomes are modulated by several user characteristics such as personality traits and cognitive abilities, to asses if explanations should be personalized to these characteristics. Our results indicate that providing explanations increase students' trust in the ACPS hints, perceived usefulness of the hints, and intention to use them again. In addition, we show that students' access of the explanation and learning gains are modulated by user characteristics, providing insights toward designing personalized Explainable AI (XAI) for ITS.

GRACE: Generating Concise and Informative Contrastive Sample to Explain Neural Network Model's Prediction

Thai Le, Suhang Wang, Dongwon Lee

Despite the recent development in the topic of explainable AI/ML for image and text data, the majority of current solutions are not suitable to explain the prediction of neural network models when the datasets are tabular and their features are in high-dimensional vectorized formats. To mitigate this limitation, therefore, we borrow two notable ideas (i.e., "explanation by intervention" from causality and "explanation are contrastive" from philosophy) and propose a novel solution, named as GRACE, that better explains neural network models' predictions for tabular datasets. In particular, given a model's prediction as label X, GRACE intervenes and generates a minimally-modified contrastive sample to be classified as Y, with an intuitive textual explanation, answering the question of "Why X rather than Y?" We carry out comprehensive experiments using eleven public datasets of different scales and domains (e.g., # of features ranges from 5 to 216) and compare GRACE with competing baselines on different measures: fidelity, conciseness, info-gain, and influence. The user-studies show that our generated explanation is not only more intuitive and easy-to-understand but also facilitates end-users to make as much as 60% more accurate post-explanation decisions than that of Lime.

Under the headline "AI safety", a wide-reaching issue is being discussed, whether in the future some "superhuman artificial intelligence" / "superintelligence" could could pose a threat to humanity. In addition, the late Steven Hawking warned that the rise of robots may be disastrous for mankind. A major concern is that even benevolent superhuman artificial intelligence (AI) may become seriously harmful if its given goals are not exactly aligned with ours, or if we cannot specify precisely its objective function. Metaphorically, this is compared to king Midas in Greek mythology, who expressed the wish that everything he touched should turn to gold, but obviously this wish was not specified precisely enough. In our view, this sounds like requirements problems and the challenge of their precise formulation. (To our best knowledge, this has not been pointed out yet.) As usual in requirements engineering (RE), ambiguity or incompleteness may cause problems. In addition, the overall issue calls for a major RE endeavor, figuring out the wishes and the needs with regard to a superintelligence, which will in our opinion most likely be a very complex software-intensive system based on AI. This may even entail theoretically defining an extended requirements problem.

Towards Explainable Neural-Symbolic Visual Reasoning

Adrien Bennetot, Jean-Luc Laurent, Raja Chatila +1

Many high-performance models suffer from a lack of interpretability. There has been an increasing influx of work on explainable artificial intelligence (XAI) in order to disentangle what is meant and expected by XAI. Nevertheless, there is no general consensus on how to produce and judge explanations. In this paper, we discuss why techniques integrating connectionist and symbolic paradigms are the most efficient solutions to produce explanations for non-technical users and we propose a reasoning model, based on definitions by Doran et al. [2017] (arXiv:1710.00794) to explain a neural network's decision. We use this explanation in order to correct bias in the network's decision rationale. We accompany this model with an example of its potential use, based on the image captioning method in Burns et al. [2018] (arXiv:1803.09797).

When AI systems are granted the agency to take impactful actions in the real world, there is an inherent risk that these systems behave in ways that are harmful. Typically, humans specify constraints on the AI system to prevent harmful behavior; however, very little work has studied how best to facilitate this difficult constraint specification process. In this paper, we study how to design user interfaces that make this process more effective and accessible, allowing people with a diversity of backgrounds and levels of expertise to contribute to this task. We first present a task design in which workers evaluate the safety of individual state-action pairs, and propose several variants of this task with improved task design and filtering mechanisms. Although this first design is easy to understand, it scales poorly to large state spaces. Therefore, we develop a new user interface that allows workers to write constraint rules without any programming. Despite its simplicity, we show that our rule construction interface retains full expressiveness. We present experiments utilizing crowdworkers to help address an important real-world AI safety problem in the domain of education. Our results indicate that our novel worker filtering and explanation methods outperform baseline approaches, and our rule-based interface allows workers to be much more efficient while improving data quality.

In this study, we propose the leveraging of interpretability for tasks beyond purely the purpose of explainability. In particular, this study puts forward a novel strategy for leveraging gradient-based interpretability in the realm of adversarial examples, where we use insights gained to aid adversarial learning. More specifically, we introduce the concept of spatially constrained one-pixel adversarial perturbations, where we guide the learning of such adversarial perturbations towards more susceptible areas identified via gradient-based interpretability. Experimental results using different benchmark datasets show that such a spatially constrained one-pixel adversarial perturbation strategy can noticeably improve the speed of convergence as well as produce successful attacks that were also visually difficult to perceive, thus illustrating an effective use of interpretability methods for tasks outside of the purpose of purely explainability.

Interaction Design for Explainable AI: Workshop Proceedings

Prashan Madumal, Ronal Singh, Joshua Newn +1

As artificial intelligence (AI) systems become increasingly complex and ubiquitous, these systems will be responsible for making decisions that directly affect individuals and society as a whole. Such decisions will need to be justified due to ethical concerns as well as trust, but achieving this has become difficult due to the `black-box' nature many AI models have adopted. Explainable AI (XAI) can potentially address this problem by explaining its actions, decisions and behaviours of the system to users. However, much research in XAI is done in a vacuum using only the researchers' intuition of what constitutes a `good' explanation while ignoring the interaction and the human aspect. This workshop invites researchers in the HCI community and related fields to have a discourse about human-centred approaches to XAI rooted in interaction and to shed light and spark discussion on interaction design challenges in XAI.

Explainable AI (XAI) methods are usually evaluated for technical properties such as faithfulness and robustness. In high-stakes application domains, experts also need to check model predictions against established knowledge. Different explanation methods often disagree, and this disagreement is hard to see when each method is inspected separately. We propose XALIGN, a visual analytics workflow and interface that shows rule-based explanations and several feature attribution methods (TreeSHAP, KernelSHAP, and LIME) side by side for the same Random Forest model. Agreements and contradictions between the methods become visible signals that users can explore and compare with domain knowledge. We apply XALIGN in two case studies: the German Credit dataset and a meteorological dataset, in which events derived from ICON model output and Cloudnet observations are classified. In both, the views reveal discrepancies between the rule-based explanation and the attribution methods that a single method would not show. In an expert evaluation with three specialists, participants found XALIGN useful for identifying influential features and for avoiding over-trust in predictions. Linking rules to attributions was harder for the participant without an XAI background.

Institutions that deploy large language models in regulated settings often need technical evidence of responsible AI practices before approving a deployment. COMPL-AI provides a technical interpretation of the EU AI Act that organises this evidence under six principle groups and eighteen technical requirements, twelve of which current benchmarks can measure. The benchmark suites that produce this evidence are typically operated from a command line, while the staff who approve deployments are non-engineers. We present VERA, an open-source self-hosted dashboard deployed to facilitate responsible AI processes within a large European bank. A reviewer selects a model and runs public benchmarks against it. VERA scores the twelve requirements that current benchmarks can measure and instruments the energy consumed by each evaluation, adding run-level evidence for environmental impact. We run four served open-weight models through the requirement pipeline, and the dashboard displays a score for each measurable requirement. A within-subject study then compares how participants read these benchmark results: from the raw output files in one condition, and through the VERA dashboard in the other. The dashboard reduced the time needed to perform the responsible AI assessment, reducing the time required from 56 seconds using raw files to 30 seconds using our dashboard, on average. Our dashboard scored highly (4.36 out of 5) on usefulness in the survey conducted within the organization. The dashboard and the study material are released publicly to promote replicability and reproducibility in different organisational contexts.

Current LLM safety benchmarks predominantly rely on binary safe or unsafe labels, lacking explanatory depth and suffering from low consistency across automated judges. To address this, we introduce DiagSafe, a diagnostic evaluation framework that moves beyond binary classification by employing prompt-level rubrics. Grounded in globally applicable AI safety policies, each rubric provides explicit criteria along two dimensions: safety, which targets concrete harms with specific exemptions to suppress false positives, and usefulness, which independently evaluates instruction fulfillment. To ensure a robust and realistic evaluation, we construct a bilingual (English-Chinese) benchmark grounded in a comprehensive 84-label safety taxonomy, consisting of two subsets: DiagBase, which covers single- and multi-turn prompts with direct harmful queries and template-based jailbreaks, and DiagAttack, which embeds harmful intents into complex instruction-following tasks to form challenging instruction attacks. Extensive experiments demonstrate that DiagSafe significantly enhances the reliability of LLM-as-a-judge evaluation, achieving 95.34% agreement with human annotations -- substantially higher than the 69%-87% achieved by the baselines -- and high inter-judge consistency (Cohen's $κ$ = 0.864). Furthermore, applying our framework reveals that leading LLMs differ markedly in their unsafe rates, and each model also exhibits unique vulnerabilities along different dimensions. In addition, we observe that models are particularly susceptible in Chinese and multi-turn contexts. Ultimately, DiagSafe provides a reliable and explainable tool for auditing and optimizing LLM alignment.

Large language models increasingly address users in the register of care: they remember personal details, adopt a first-person emotional voice, and sustain relationships across many sessions. This anthropomorphic mode of interaction generates risks that output-level metrics capture poorly, because the same empathic utterance may console one user, entrap a second, and consolidate a delusion in a third. Governance has responded in two ways. Uniform restriction refuses sensitive requests or suppresses empathic phrasing across all outputs; group-targeted protection profiles vulnerable populations and applies stricter or prohibitive rules to them. This article argues that both fail. Uniform restriction mislocates harm in the lexical form of responses rather than in the trajectory of a relationship; group-based protection depends on a list that can never be complete, treats a fluid property as a static attribute of persons, and converts protective attention into stigmatizing labels. Drawing on philosophical work that treats vulnerability as universal, relational, and differentiated, we propose that the unit of ethical governance should be the vulnerability situation rather than the vulnerable group, develop a typology of five situations - physical, emotional, cognitive, autonomy, and relational - and specify seven principles into requirements that guide response calibration without diagnosing or persistently classifying users. The framework relocates AI safety from output control to relational safety.

AI agents are increasingly capable of constructing quantum algorithms and proofs, yet machine-checked proofs alone cannot guarantee that the intended problem is solved or that resource claims match the verified computation.Here we introduce Lean-QuantumAlg, a Lean 4 framework for trustworthy AI-assisted development of quantum algorithms.Its Algorithm-Solves-Problem contract binds correctness, success probability, borrowed-state restoration and quantum and classical resource bounds to a single execution, while separating the guarantees that Lean checks from the problem specifications and execution models that humans review. We demonstrate the framework on nine AI-formalized quantum algorithms, from Grover search to quantum attacks on RSA-2048 and P-256, with less than 1% of each algorithm's imported source code requiring statement-level human review once the shared framework has been audited. By integrating reusable mathematical foundations, certified algorithmic primitives and execution-grounded verification, Lean-QuantumAlg provides a basis for AI agents to construct and certify quantum algorithms, opening a path toward scalable AI-driven discovery of quantum algorithms.