Research Papers

Reasoning Visual Language Model for Chest X-Ray Analysis

Andriy Myronenko, Dong Yang, Baris Turkbey +10

Vision-language models (VLMs) have shown strong promise for medical image analysis, but most remain opaque, offering predictions without the transparent, stepwise reasoning clinicians rely on. We present a framework that brings chain-of-thought (CoT) reasoning to chest X-ray interpretation. Inspired by reasoning-first training paradigms, our approach is designed to learn how experts reason, not just what they conclude, by aligning intermediate steps with observable image evidence and radiology workflow. Beyond accuracy, the explicit reasoning traces support clinical auditability: they reveal why a conclusion was reached, which alternatives were considered, and where uncertainty remains, enabling quality assurance, error analysis, and safer human-AI collaboration. Our model couples high-fidelity visual encoding with a two-stage training recipe: a reasoning-style supervised fine-tuning (SFT) followed by reinforcement learning (RL) that uses verifiable rewards over a list of X-ray abnormalities. The model outputs reasoning that mirrors radiologists systematic thought process, uncertainty, and differential diagnosis. In out-of-distribution evaluation, the approach achieves competitive multi-label classification while improving interpretability. In a reader study with expert radiologists, full reasoning traces increased confidence, supported error auditing, and reduced time to finalize reports. We release code and the model NV-Reason-CXR-3B to support community progress toward trustworthy, explainable AI in chest radiography and other medical imaging tasks where reasoning quality is as critical as prediction quality.

Generative AI (GenAI) is changing the nature of knowledge work, particularly for Product Managers (PMs) in software development teams. While much software engineering research has focused on developers' interactions with GenAI, there is less understanding of how the work of PMs is evolving due to GenAI. To address this gap, we conducted a mixed-methods study at Microsoft, a large, multinational software company: surveying 885 PMs, analyzing telemetry data for a subset of PMs (N=731), and interviewing a subset of 15 PMs. We contribute: (1) PMs' current GenAI adoption rates, uses cases, and perceived benefits and barriers and; (2) a framework capturing how PMs assess which tasks to delegate to GenAI; (3) PMs adaptation practices for integrating GenAI into their roles and perceptions of how their role is evolving. We end by discussing implications on the broader GenAI workflow adoption process and software development roles.

Minimal Data, Maximum Clarity: A Heuristic for Explaining Optimization

Amirali Rayegan, Tim Menzies

Efficient, interpretable optimization is a critical but underexplored challenge in software engineering, where practitioners routinely face vast configuration spaces and costly, error-prone labeling processes. This paper introduces EZR, a novel and modular framework for multi-objective optimization that unifies active sampling, learning, and explanation within a single, lightweight pipeline. Departing from conventional wisdom, our Maximum Clarity Heuristic demonstrates that using less (but more informative) data can yield optimization models that are both effective and deeply understandable. EZR employs an active learning strategy based on Naive Bayes sampling to efficiently identify high-quality configurations with a fraction of the labels required by fully supervised approaches. It then distills optimization logic into concise decision trees, offering transparent, actionable explanations for both global and local decision-making. Extensive experiments across 60 real-world datasets establish that EZR reliably achieves over 90% of the best-known optimization performance in most cases, while providing clear, cohort-based rationales that surpass standard attribution-based explainable AI (XAI) methods (LIME, SHAP, BreakDown) in clarity and utility. These results endorse "less but better"; it is both possible and often preferable to use fewer (but more informative) examples to generate label-efficient optimization and explanations in software systems. To support transparency and reproducibility, all code and experimental materials are publicly available at https://github.com/amiiralii/Minimal-Data-Maximum-Clarity.

Bias Amplification in Stable Diffusion's Representation of Stigma Through Skin Tones and Their Homogeneity

Kyra Wilson, Sourojit Ghosh, Aylin Caliskan

Text-to-image generators (T2Is) are liable to produce images that perpetuate social stereotypes, especially in regards to race or skin tone. We use a comprehensive set of 93 stigmatized identities to determine that three versions of Stable Diffusion (v1.5, v2.1, and XL) systematically associate stigmatized identities with certain skin tones in generated images. We find that SD XL produces skin tones that are 13.53% darker and 23.76% less red (both of which indicate higher likelihood of societal discrimination) than previous models and perpetuate societal stereotypes associating people of color with stigmatized identities. SD XL also shows approximately 30% less variability in skin tones when compared to previous models and 18.89-56.06% compared to human face datasets. Measuring variability through metrics which directly correspond to human perception suggest a similar pattern, where SD XL shows the least amount of variability in skin tones of people with stigmatized identities and depicts most (60.29%) stigmatized identities as being less diverse than non-stigmatized identities. Finally, SD shows more homogenization of skin tones of racial and ethnic identities compared to other stigmatized or non-stigmatized identities, reinforcing incorrect equivalence of biologically-determined skin tone and socially-constructed racial and ethnic identity. Because SD XL is the largest and most complex model and users prefer its generations compared to other models examined in this study, these findings have implications for the dynamics of bias amplification in T2Is, increasing representational harms and challenges generating diverse images depicting people with stigmatized identities.

As AI agents become more widely deployed, we are likely to see an increasing number of incidents: events involving AI agent use that directly or indirectly cause harm. For example, agents could be prompt-injected to exfiltrate private information or make unauthorized purchases. Structured information about such incidents (e.g., user prompts) can help us understand their causes and prevent future occurrences. However, existing incident reporting processes are not sufficient for understanding agent incidents. In particular, such processes are largely based on publicly available data, which excludes useful, but potentially sensitive, information such as an agent's chain of thought or browser history. To inform the development of new, emerging incident reporting processes, we propose an incident analysis framework for agents. Drawing on systems safety approaches, our framework proposes three types of factors that can cause incidents: system-related (e.g., CBRN training data), contextual (e.g., prompt injections), and cognitive (e.g., misunderstanding a user request). We also identify specific information that could help clarify which factors are relevant to a given incident: activity logs, system documentation and access, and information about the tools an agent uses. We provide recommendations for 1) what information incident reports should include and 2) what information developers and deployers should retain and make available to incident investigators upon request. As we transition to a world with more agents, understanding agent incidents will become increasingly crucial for managing risks.

The integration of Large Language Models (LLMs) into software engineering has revolutionized code generation, enabling unprecedented productivity through promptware and autonomous AI agents. However, this transformation introduces significant risks, including insecure code generation, hallucinated outputs, irreversible actions, and a lack of transparency and accountability. Incidents like the Replit database deletion underscore the urgent need for robust safety and governance mechanisms. This paper comprehensively analyzes the inherent challenges of LLM-assisted code generation, such as vulnerability inheritance, overtrust, misinterpretation, and the absence of standardized validation and rollback protocols. To address these, we propose the SAFE-AI Framework, a holistic approach emphasizing Safety, Auditability, Feedback, and Explainability. The framework integrates guardrails, sandboxing, runtime verification, risk-aware logging, human-in-the-loop systems, and explainable AI techniques to mitigate risks while fostering trust and compliance. We introduce a novel taxonomy of AI behaviors categorizing suggestive, generative, autonomous, and destructive actions to guide risk assessment and oversight. Additionally, we identify open problems, including the lack of standardized benchmarks for code specific hallucinations and autonomy levels, and propose future research directions for hybrid verification, semantic guardrails, and proactive governance tools. Through detailed comparisons of autonomy control, prompt engineering, explainability, and governance frameworks, this paper provides a roadmap for responsible AI integration in software engineering, aligning with emerging regulations like the EU AI Act and Canada's AIDA to ensure safe, transparent, and accountable AI-driven development.

As AI regulations around the world intensify their focus on system safety, contestability has become a mandatory, yet ill-defined, safeguard. In XAI, "contestability" remains an empty promise: no formal definition exists, no algorithm guarantees it, and practitioners lack concrete guidance to satisfy regulatory requirements. Grounded in a systematic literature review, this paper presents the first rigorous formal definition of contestability in explainable AI, directly aligned with stakeholder requirements and regulatory mandates. We introduce a modular framework of by-design and post-hoc mechanisms spanning human-centered interfaces, technical architectures, legal processes, and organizational workflows. To operationalize our framework, we propose the Contestability Assessment Scale, a composite metric built on more than twenty quantitative criteria. Through multiple case studies across diverse application domains, we reveal where state-of-the-art systems fall short and show how our framework drives targeted improvements. By converting contestability from regulatory theory into a practical framework, our work equips practitioners with the tools to embed genuine recourse and accountability into AI systems.

As artificial intelligence (AI) becomes integral to the society, the ability to critically evaluate AI-generated content is increasingly vital. On the context of management education, we examine how academic skills, cognitive traits, and AI scepticism influence students' ability to detect factually incorrect AI-generated responses (hallucinations) in a high-stakes assessment at a UK business school (n=211, Year 2 economics and management students). We find that only 20% successfully identified the hallucination, with strong academic performance, interpretive skills thinking, writing proficiency, and AI scepticism emerging as key predictors. In contrast, rote knowledge application proved less effective, and gender differences in detection ability were observed. Beyond identifying predictors of AI hallucination detection, we tie the theories of epistemic cognition, cognitive bias, and transfer of learning with new empirical evidence by demonstrating how AI literacy could enhance long-term analytical performance in high-stakes settings. We advocate for an innovative and practical framework for AI-integrated assessments, showing that structured feedback mitigates initial disparities in detection ability. These findings provide actionable insights for educators designing AI-aware curricula that foster critical reasoning, epistemic vigilance, and responsible AI engagement in management education. Our study contributes to the broader discussion on the evolution of knowledge evaluation in AI-enhanced learning environments.

Embracing Contradiction: Theoretical Inconsistency Will Not Impede the Road of Building Responsible AI Systems

Gordon Dai, Yunze Xiao

This position paper argues that the theoretical inconsistency often observed among Responsible AI (RAI) metrics, such as differing fairness definitions or tradeoffs between accuracy and privacy, should be embraced as a valuable feature rather than a flaw to be eliminated. We contend that navigating these inconsistencies, by treating metrics as divergent objectives, yields three key benefits: (1) Normative Pluralism: Maintaining a full suite of potentially contradictory metrics ensures that the diverse moral stances and stakeholder values inherent in RAI are adequately represented. (2) Epistemological Completeness: The use of multiple, sometimes conflicting, metrics allows for a more comprehensive capture of multifaceted ethical concepts, thereby preserving greater informational fidelity about these concepts than any single, simplified definition. (3) Implicit Regularization: Jointly optimizing for theoretically conflicting objectives discourages overfitting to one specific metric, steering models towards solutions with enhanced generalization and robustness under real-world complexities. In contrast, efforts to enforce theoretical consistency by simplifying or pruning metrics risk narrowing this value diversity, losing conceptual depth, and degrading model performance. We therefore advocate for a shift in RAI theory and practice: from getting trapped in inconsistency to characterizing acceptable inconsistency thresholds and elucidating the mechanisms that permit robust, approximated consistency in practice.

Artificial intelligence (AI) is rapidly transforming education, presenting unprecedented opportunities for personalized learning and streamlined content creation. However, realizing the full potential of AI in educational settings necessitates careful consideration of the quality, cognitive depth, and ethical implications of AI-generated materials. This paper synthesizes insights from four related studies to propose a comprehensive framework for enhancing AI-driven educational tools. We integrate cognitive assessment frameworks (Bloom's Taxonomy and SOLO Taxonomy), linguistic analysis of AI-generated feedback, and ethical design principles to guide the development of effective and responsible AI tools. We outline a structured three-phase approach encompassing cognitive alignment, linguistic feedback integration, and ethical safeguards. The practical application of this framework is demonstrated through its integration into OneClickQuiz, an AI-powered Moodle plugin for quiz generation. This work contributes a comprehensive and actionable guide for educators, researchers, and developers aiming to harness AI's potential while upholding pedagogical and ethical standards in educational content generation.

Insider threats pose a significant challenge to organizational security, often evading traditional rule-based detection systems due to their subtlety and contextual nature. This paper presents an AI-powered Insider Risk Management (IRM) system that integrates behavioral analytics, dynamic risk scoring, and real-time policy enforcement to detect and mitigate insider threats with high accuracy and adaptability. We introduce a hybrid scoring mechanism - transitioning from the static PRISM model to an adaptive AI-based model utilizing an autoencoder neural network trained on expert-annotated user activity data. Through iterative feedback loops and continuous learning, the system reduces false positives by 59% and improves true positive detection rates by 30%, demonstrating substantial gains in detection precision. Additionally, the platform scales efficiently, processing up to 10 million log events daily with sub-300ms query latency, and supports automated enforcement actions for policy violations, reducing manual intervention. The IRM system's deployment resulted in a 47% reduction in incident response times, highlighting its operational impact. Future enhancements include integrating explainable AI, federated learning, graph-based anomaly detection, and alignment with Zero Trust principles to further elevate its adaptability, transparency, and compliance-readiness. This work establishes a scalable and proactive framework for mitigating emerging insider risks in both on-premises and hybrid environments.

Physical attractiveness matters. It has been shown to influence human perception and decision-making, often leading to biased judgments that favor those deemed attractive in what is referred to as the "attractiveness halo effect". While extensively studied in human judgments in a broad set of domains, including hiring, judicial sentencing or credit granting, the role that attractiveness plays in the assessments and decisions made by multimodal large language models (MLLMs) is unknown. To address this gap, we conduct an empirical study with 7 diverse open-source MLLMs evaluated on 91 socially relevant scenarios and a diverse dataset of 924 face images - corresponding to 462 individuals both with and without beauty filters applied to them. Our analysis reveals that attractiveness impacts the decisions made by MLLMs in 86.2% of the scenarios on average, demonstrating substantial bias in model behavior in what we refer to as an attractiveness bias. Similarly to humans, we find empirical evidence of the existence of the attractiveness halo effect in 94.8% of the relevant scenarios: attractive individuals are more likely to be attributed positive traits, such as intelligence or confidence, by MLLMs than unattractive individuals. Furthermore, we uncover gender, age and race biases in a significant portion of the scenarios which are also impacted by attractiveness, particularly in the case of gender, highlighting the intersectional nature of the algorithmic attractiveness bias. Our findings suggest that societal stereotypes and cultural norms intersect with perceptions of attractiveness in MLLMs in a complex manner. Our work emphasizes the need to account for intersectionality in algorithmic bias detection and mitigation efforts and underscores the challenges of addressing biases in modern MLLMs.