Research Papers

Providing meaningful and actionable explanations to end-users is a fundamental prerequisite for implementing explainable intelligent systems in the real world. Explainability is a situated interaction between a user and the AI system rather than being static design principles. The content of explanations is context-dependent and must be defined by evidence about the user and its context. This paper seeks to operationalize this concept by proposing a formative architecture that defines the explanation space from a user-inspired perspective. The architecture comprises five intertwined components to outline explanation requirements for a task: (1) the end-users mental models, (2) the end-users cognitive process, (3) the user interface, (4) the human-explainer agent, and the (5) agent process. We first define each component of the architecture. Then we present the Abstracted Explanation Space, a modeling tool that aggregates the architecture's components to support designers in systematically aligning explanations with the end-users work practices, needs, and goals. It guides the specifications of what needs to be explained (content - end-users mental model), why this explanation is necessary (context - end-users cognitive process), to delimit how to explain it (format - human-explainer agent and user interface), and when should the explanations be given. We then exemplify the tool's use in an ongoing case study in the aircraft maintenance domain. Finally, we discuss possible contributions of the tool, known limitations/areas for improvement, and future work to be done.

Visual Question Answering (VQA) models have achieved significant success in recent times. Despite the success of VQA models, they are mostly black-box models providing no reasoning about the predicted answer, thus raising questions for their applicability in safety-critical such as autonomous systems and cyber-security. Current state of the art fail to better complex questions and thus are unable to exploit compositionality. To minimize the black-box effect of these models and also to make them better exploit compositionality, we propose a Dynamic Neural Network (DMN), which can understand a particular question and then dynamically assemble various relatively shallow deep learning modules from a pool of modules to form a network. We incorporate compositional temporal attention to these deep learning based modules to increase compositionality exploitation. This results in achieving better understanding of complex questions and also provides reasoning as to why the module predicts a particular answer. Experimental analysis on the two benchmark datasets, VQA2.0 and CLEVR, depicts that our model outperforms the previous approaches for Visual Question Answering task as well as provides better reasoning, thus making it reliable for mission critical applications like safety and security.

Given that there are a variety of stakeholders involved in, and affected by, decisions from machine learning (ML) models, it is important to consider that different stakeholders have different transparency needs. Previous work found that the majority of deployed transparency mechanisms primarily serve technical stakeholders. In our work, we want to investigate how well transparency mechanisms might work in practice for a more diverse set of stakeholders by conducting a large-scale, mixed-methods user study across a range of organizations, within a particular industry such as health care, criminal justice, or content moderation. In this paper, we outline the setup for our study.

Causal Analysis of Agent Behavior for AI Safety

Grégoire Déletang, Jordi Grau-Moya, Miljan Martic +6

As machine learning systems become more powerful they also become increasingly unpredictable and opaque. Yet, finding human-understandable explanations of how they work is essential for their safe deployment. This technical report illustrates a methodology for investigating the causal mechanisms that drive the behaviour of artificial agents. Six use cases are covered, each addressing a typical question an analyst might ask about an agent. In particular, we show that each question cannot be addressed by pure observation alone, but instead requires conducting experiments with systematically chosen manipulations so as to generate the correct causal evidence.

Rapid non-verbal communication of task-based stimuli is a challenge in human-machine teaming, particularly in closed-loop interactions such as driving. To achieve this, we must understand the representations of information for both the human and machine, and determine a basis for bridging these representations. Techniques of explainable artificial intelligence (XAI) such as layer-wise relevance propagation (LRP) provide visual heatmap explanations for high-dimensional machine learning techniques such as deep neural networks. On the side of human cognition, visual attention is driven by the bottom-up and top-down processing of sensory input related to the current task. Since both XAI and human cognition should focus on task-related stimuli, there may be overlaps between their representations of visual attention, potentially providing a means of nonverbal communication between the human and machine. In this work, we examine the correlations between LRP heatmap explanations of a neural network trained to predict driving behavior and eye gaze heatmaps of human drivers. The analysis is used to determine the feasibility of using such a technique for enhancing driving performance. We find that LRP heatmaps show increasing levels of similarity with eye gaze according to the task specificity of the neural network. We then propose how these findings may assist humans by visually directing attention towards relevant areas. To our knowledge, our work provides the first known analysis of LRP and eye gaze for driving tasks.

Text entailment, the task of determining whether a piece of text logically follows from another piece of text, is a key component in NLP, providing input for many semantic applications such as question answering, text summarization, information extraction, and machine translation, among others. Entailment scenarios can range from a simple syntactic variation to more complex semantic relationships between pieces of text, but most approaches try a one-size-fits-all solution that usually favors some scenario to the detriment of another. Furthermore, for entailments requiring world knowledge, most systems still work as a "black box", providing a yes/no answer that does not explain the underlying reasoning process. In this work, we introduce XTE - Explainable Text Entailment - a novel composite approach for recognizing text entailment which analyzes the entailment pair to decide whether it must be resolved syntactically or semantically. Also, if a semantic matching is involved, we make the answer interpretable, using external knowledge bases composed of structured lexical definitions to generate natural language justifications that explain the semantic relationship holding between the pieces of text. Besides outperforming well-established entailment algorithms, our composite approach gives an important step towards Explainable AI, allowing the inference model interpretation, making the semantic reasoning process explicit and understandable.

To achieve optimal human-system integration in the context of user-AI interaction it is important that users develop a valid representation of how AI works. In most of the everyday interaction with technical systems users construct mental models (i.e., an abstraction of the anticipated mechanisms a system uses to perform a given task). If no explicit explanations are provided by a system (e.g. by a self-explaining AI) or other sources (e.g. an instructor), the mental model is typically formed based on experiences, i.e. the observations of the user during the interaction. The congruence of this mental model and the actual systems functioning is vital, as it is used for assumptions, predictions and consequently for decisions regarding system use. A key question for human-centered AI research is therefore how to validly survey users' mental models. The objective of the present research is to identify suitable elicitation methods for mental model analysis. We evaluated whether mental models are suitable as an empirical research method. Additionally, methods of cognitive tutoring are integrated. We propose an exemplary method to evaluate explainable AI approaches in a human-centered way.

Despite the recent successes of deep learning, such models are still far from some human abilities like learning from few examples, reasoning and explaining decisions. In this paper, we focus on organ annotation in medical images and we introduce a reasoning framework that is based on learning fuzzy relations on a small dataset for generating explanations. Given a catalogue of relations, it efficiently induces the most relevant relations and combines them for building constraints in order to both solve the organ annotation task and generate explanations. We test our approach on a publicly available dataset of medical images where several organs are already segmented. A demonstration of our model is proposed with an example of explained annotations. It was trained on a small training set containing as few as a couple of examples.

Visualizing the decision-making process in deep neural decision forest

Shichao Li, Kwang-Ting Cheng

Deep neural decision forest (NDF) achieved remarkable performance on various vision tasks via combining decision tree and deep representation learning. In this work, we first trace the decision-making process of this model and visualize saliency maps to understand which portion of the input influence it more for both classification and regression problems. We then apply NDF on a multi-task coordinate regression problem and demonstrate the distribution of routing probabilities, which is vital for interpreting NDF yet not shown for regression problems. The pre-trained model and code for visualization will be available at https://github.com/Nicholasli1995/VisualizingNDF

In Random Forests, proximity distances are a metric representation of data into decision space. By observing how changes in input map to the movement of instances in this space we are able to determine the independent contribution of each feature to the decision-making process. For binary feature vectors, this process is fully specified. As these changes in input move particular instances nearer to the in-group or out-group, the independent contribution of each feature can be uncovered. Using this technique, we are able to calculate the contribution of each feature in determining how black-box decisions were made. This allows explication of the decision-making process, audit of the classifier, and post-hoc analysis of errors in classification.

Ensuring trust in AI systems is essential for the safe and ethical integration of machine learning systems into high-stakes domains such as digital health. Key dimensions, including robustness, explainability, fairness, accountability, and privacy, need to be addressed throughout the AI lifecycle, from problem formulation and data collection to model deployment and human interaction. While various contributions address different aspects of trustworthy AI, a focused synthesis on robustness and explainability, especially tailored to the healthcare context, remains limited. This review addresses that need by organizing recent advancements into an accessible framework, highlighting both technical and practical considerations. We present a structured overview of methods, challenges, and solutions, aiming to support researchers and practitioners in developing reliable and explainable AI solutions for digital health. This review article is organized into three main parts. First, we introduce the pillars of trustworthy AI and discuss the technical and ethical challenges, particularly in the context of digital health. Second, we explore application-specific trust considerations across domains such as intensive care, neonatal health, and metabolic health, highlighting how robustness and explainability support trust. Lastly, we present recent advancements in techniques aimed at improving robustness under data scarcity and distributional shifts, as well as explainable AI methods ranging from feature attribution to gradient-based interpretations and counterfactual explanations. This paper is further enriched with detailed discussions of the contributions toward robustness and explainability in digital health, the development of trustworthy AI systems in the era of LLMs, and various evaluation metrics for measuring trust and related parameters such as validity, fidelity, and diversity.

Not All LLM Reasoning is Visible in the Chain-of-Thought

Vatsal Baherwani, Tom Goldstein, Ashwinee Panda

A key question for AI safety is whether a language model expresses all of its reasoning in its output tokens. We demonstrate a concrete failure mode where frontier models exhibit invisible reasoning by leveraging semantically irrelevant filler tokens to improve performance on synthetic reasoning tasks. We evaluate 13 frontier language models across three tasks and find that many models benefit significantly from filler tokens, with accuracy improvements of up to 13 percentage points. The benefit depends on which tokens are used and differs across models. We further show that filler tokens enable Claude Opus 4.5 to satisfy a hidden modular arithmetic constraint without sacrificing accuracy on its primary task, demonstrating that invisible reasoning can serve objectives entirely invisible to CoT monitoring. Reinforcement learning gives Qwen3-235B strong preferences over filler token content, but neither RL nor supervised fine-tuning produces a filler token benefit that persists at test time. Our results indicate that frontier models already perform consequential computation with no interpretable trace in their output tokens.