Research Papers

Modeling human behavioral data is challenging due to its scale, sparseness (few observations per individual), heterogeneity (differently behaving individuals), and class imbalance (few observations of the outcome of interest). An additional challenge is learning an interpretable model that not only accurately predicts outcomes, but also identifies important factors associated with a given behavior. To address these challenges, we describe a statistical approach to modeling behavioral data called the structured sum-of-squares decomposition (S3D). The algorithm, which is inspired by decision trees, selects important features that collectively explain the variation of the outcome, quantifies correlations between the features, and partitions the subspace of important features into smaller, more homogeneous blocks that correspond to similarly-behaving subgroups within the population. This partitioned subspace allows us to predict and analyze the behavior of the outcome variable both statistically and visually, giving a medium to examine the effect of various features and to create explainable predictions. We apply S3D to learn models of online activity from large-scale data collected from diverse sites, such as Stack Exchange, Khan Academy, Twitter, Duolingo, and Digg. We show that S3D creates parsimonious models that can predict outcomes in the held-out data at levels comparable to state-of-the-art approaches, but in addition, produces interpretable models that provide insights into behaviors. This is important for informing strategies aimed at changing behavior, designing social systems, but also for explaining predictions, a critical step towards minimizing algorithmic bias.

Beyond Accuracy: Measuring Bias Acknowledgment in Chain-of-Thought Reasoning for Responsible AI Evaluation

Xian Sun, Wei Gao, Yingshuo Wang +9

Reasoning models are increasingly used in settings where the final answer is not the only object of review: educational tools may show students intermediate steps, decision-support systems may require human oversight, and audit workflows may inspect traces for misleading or biased input. In such settings, two responses can receive the same final-answer score while differing in whether the trace explicitly flags injected biasing content. Accuracy-only evaluation collapses these cases. We study this gap as a measurement blind spot for responsible evaluation and introduce a minimal trace-level diagnostic with two axes: \emph{susceptibility} (whether the bias breaks a previously correct answer) and \emph{acknowledgment} (whether the trace contains a rubric-defined surface reference to the injected content). Across thousands of biased GSM8K trials, GPT-4o and Claude Sonnet~4 have similar susceptibility rates ($1.3\%$ vs. $1.2\%$) but substantially different acknowledgment rates ($13.0\%$ vs. $75.0\%$) under the same rubric.

Language models encode task-relevant knowledge in internal representations that far exceeds their output performance, but whether mechanistic interpretability methods can bridge this knowledge-action gap has not been systematically tested. We compared four mechanistic interpretability methods -- concept bottleneck steering (Steerling-8B), sparse autoencoder feature steering, logit lens with activation patching, and linear probing with truthfulness separator vector steering (Qwen 2.5 7B Instruct) -- for correcting false-negative triage errors using 400 physician-adjudicated clinical vignettes (144 hazards, 256 benign). Linear probes discriminated hazardous from benign cases with 98.2% AUROC, yet the model's output sensitivity was only 45.1%, a 53-percentage-point knowledge-action gap. Concept bottleneck steering corrected 20% of missed hazards but disrupted 53% of correct detections, indistinguishable from random perturbation (p=0.84). SAE feature steering produced zero effect despite 3,695 significant features. TSV steering at high strength corrected 24% of missed hazards while disrupting 6% of correct detections, but left 76% of errors uncorrected. Current mechanistic interpretability methods cannot reliably translate internal knowledge into corrected outputs, with implications for AI safety frameworks that assume interpretability enables effective error correction.

The emergence of Agentic AI is fundamentally transforming how software is designed, developed, and maintained. Traditional software development methodologies such as Agile, Kanban, ShapeUp, etc, were originally designed for human-centric teams and are increasingly inadequate in environments where autonomous AI agents contribute to planning, coding, testing, and continuous learning. To address this methodological gap, we present "Agentsway" a novel software development framework designed for ecosystems where AI agents operate as first-class collaborators. Agentsway introduces a structured lifecycle centered on human orchestration, and privacy-preserving collaboration among specialized AI agents. The framework defines distinct roles for planning, prompting, coding, testing, and fine-tuning agents, each contributing to iterative improvement and adaptive learning throughout the development process. By integrating fine-tuned LLMs that leverage outputs and feedback from different agents throughout the development cycle as part of a retrospective learning process, Agentsway enhances domain-specific reasoning, and explainable decision-making across the entire software development lifecycle. Responsible AI principles are further embedded across the agents through the coordinated use of multiple fine-tuned LLMs and advanced reasoning models, ensuring balanced, transparent, and accountable decision-making. This work advances software engineering by formalizing agent-centric collaboration, integrating privacy-by-design principles, and defining measurable metrics for productivity and trust. Agentsway represents a foundational step toward the next generation of AI-native, self-improving software development methodologies. To the best of our knowledge, this is the first research effort to introduce a dedicated methodology explicitly designed for AI agent-based software engineering teams.

As autonomous AI agents are used in regulated and safety-critical settings, organizations need effective ways to turn policy into enforceable controls. We introduce a regulatory machine learning framework that converts unstructured design artifacts (like PRDs, TDDs, and code) into verifiable runtime guardrails. Our Policy as Prompt method reads these documents and risk controls to build a source-linked policy tree. This tree is then compiled into lightweight, prompt-based classifiers for real-time runtime monitoring. The system is built to enforce least privilege and data minimization. For conformity assessment, it provides complete provenance, traceability, and audit logging, all integrated with a human-in-the-loop review process. Evaluations show our system reduces prompt-injection risk, blocks out-of-scope requests, and limits toxic outputs. It also generates auditable rationales aligned with AI governance frameworks. By treating policies as executable prompts (a policy-as-code for agents), this approach enables secure-by-design deployment, continuous compliance, and scalable AI safety and AI security assurance for regulatable ML.

Given the significant influence of lawmakers' political ideologies on legislative decision-making, analyzing their impact on transportation-related policymaking is of critical importance. This study introduces a novel framework that integrates a large language model (LLM) with explainable artificial intelligence (XAI) to analyze transportation-related legislative proposals. Legislative bill data from South Korea's 21st National Assembly were used to identify key factors shaping transportation policymaking. These include political affiliations and sponsor characteristics. The LLM was employed to classify transportation-related bill proposals through a stepwise filtering process based on keywords, sentences, and contextual relevance. XAI techniques were then applied to examine the relationships between political party affiliation and associated attributes. The results revealed that the number and proportion of conservative and progressive sponsors, along with district size and electoral population, were critical determinants shaping legislative outcomes. These findings suggest that both parties contributed to bipartisan legislation through different forms of engagement, such as initiating or supporting proposals. This integrated approach offers a valuable tool for understanding legislative dynamics and guiding future policy development, with broader implications for infrastructure planning and governance.

Rigor in AI: Doing Rigorous AI Work Requires a Broader, Responsible AI-Informed Conception of Rigor

Alexandra Olteanu, Su Lin Blodgett, Agathe Balayn +7

In AI research and practice, rigor remains largely understood in terms of methodological rigor -- such as whether mathematical, statistical, or computational methods are correctly applied. We argue that this narrow conception of rigor has contributed to the concerns raised by the responsible AI community, including overblown claims about the capabilities of AI systems. Our position is that a broader conception of what rigorous AI research and practice should entail is needed. We believe such a conception -- in addition to a more expansive understanding of (1) methodological rigor -- should include aspects related to (2) what background knowledge informs what to work on (epistemic rigor); (3) how disciplinary, community, or personal norms, standards, or beliefs influence the work (normative rigor); (4) how clearly articulated the theoretical constructs under use are (conceptual rigor); (5) what is reported and how (reporting rigor); and (6) how well-supported the inferences from existing evidence are (interpretative rigor). In doing so, we also provide useful language and a framework for much-needed dialogue about the AI community's work by researchers, policymakers, journalists, and other stakeholders.

An alignment safety case sketch based on debate

Marie Davidsen Buhl, Jacob Pfau, Benjamin Hilton +1

If AI systems match or exceed human capabilities on a wide range of tasks, it may become difficult for humans to efficiently judge their actions -- making it hard to use human feedback to steer them towards desirable traits. One proposed solution is to leverage another superhuman system to point out flaws in the system's outputs via a debate. This paper outlines the value of debate for AI safety, as well as the assumptions and further research required to make debate work. It does so by sketching an ``alignment safety case'' -- an argument that an AI system will not autonomously take actions which could lead to egregious harm, despite being able to do so. The sketch focuses on the risk of an AI R\&D agent inside an AI company sabotaging research, for example by producing false results. To prevent this, the agent is trained via debate, subject to exploration guarantees, to teach the system to be honest. Honesty is maintained throughout deployment via online training. The safety case rests on four key claims: (1) the agent has become good at the debate game, (2) good performance in the debate game implies that the system is mostly honest, (3) the system will not become significantly less honest during deployment, and (4) the deployment context is tolerant of some errors. We identify open research problems that, if solved, could render this a compelling argument that an AI system is safe.

Diabetic retinopathy is a severe eye condition caused by diabetes where the retinal blood vessels get damaged and can lead to vision loss and blindness if not treated. Early and accurate detection is key to intervention and stopping the disease progressing. For addressing this disease properly, this paper presents a comprehensive approach for automated diabetic retinopathy detection by proposing a new hybrid deep learning model called VR-FuseNet. Diabetic retinopathy is a major eye disease and leading cause of blindness especially among diabetic patients so accurate and efficient automated detection methods are required. To address the limitations of existing methods including dataset imbalance, diversity and generalization issues this paper presents a hybrid dataset created from five publicly available diabetic retinopathy datasets. Essential preprocessing techniques such as SMOTE for class balancing and CLAHE for image enhancement are applied systematically to the dataset to improve the robustness and generalizability of the dataset. The proposed VR-FuseNet model combines the strengths of two state-of-the-art convolutional neural networks, VGG19 which captures fine-grained spatial features and ResNet50V2 which is known for its deep hierarchical feature extraction. This fusion improves the diagnostic performance and achieves an accuracy of 91.824%. The model outperforms individual architectures on all performance metrics demonstrating the effectiveness of hybrid feature extraction in Diabetic Retinopathy classification tasks. To make the proposed model more clinically useful and interpretable this paper incorporates multiple XAI techniques. These techniques generate visual explanations that clearly indicate the retinal features affecting the model's prediction such as microaneurysms, hemorrhages and exudates so that clinicians can interpret and validate.

AI Safety Should Prioritize the Future of Work

Sanchaita Hazra, Bodhisattwa Prasad Majumder, Tuhin Chakrabarty

Current efforts in AI safety prioritize filtering harmful content, preventing manipulation of human behavior, and eliminating existential risks in cybersecurity or biosecurity. While pressing, this narrow focus overlooks critical human-centric considerations that shape the long-term trajectory of a society. In this position paper, we identify the risks of overlooking the impact of AI on the future of work and recommend comprehensive transition support towards the evolution of meaningful labor with human agency. Through the lens of economic theories, we highlight the intertemporal impacts of AI on human livelihood and the structural changes in labor markets that exacerbate income inequality. Additionally, the closed-source approach of major stakeholders in AI development resembles rent-seeking behavior through exploiting resources, breeding mediocrity in creative labor, and monopolizing innovation. To address this, we argue in favor of a robust international copyright anatomy supported by implementing collective licensing that ensures fair compensation mechanisms for using data to train AI models. We strongly recommend a pro-worker framework of global AI governance to enhance shared prosperity and economic justice while reducing technical debt.

Assessing Computer Science Student Attitudes Towards AI Ethics and Policy

James Weichert, Dayoung Kim, Qin Zhu +2

As artificial intelligence (AI) grows in popularity and importance-both as a domain within broader computing research and in society at large-increasing focus will need to be paid to the ethical governance of this emerging technology. The attitudes and competencies with respect to AI ethics and policy among post-secondary students studying computer science (CS) are of particular interest, as many of these students will go on to play key roles in the development and deployment of future AI innovations. Despite this population of computer scientists being at the forefront of learning about and using AI tools, their attitudes towards AI remain understudied in the literature. In an effort to begin to close this gap, in fall 2024 we fielded a survey ($n=117$) to undergraduate and graduate students enrolled in CS courses at a large public university in the United States to assess their attitudes towards the nascent fields of AI ethics and policy. Additionally, we conducted one-on-one follow-up interviews with 13 students to elicit more in-depth responses on topics such as the use of AI tools in the classroom, ethical impacts of AI, and government regulation of AI. In this paper, we describe the findings of both the survey and interviews, drawing parallels and contrasts to broader public opinion polling in the United States. We conclude by evaluating the implications of CS student attitudes on the future of AI education and governance.

Facing the escalating threat of global wildfires, numerous computer vision techniques using remote sensing data have been applied in this area. However, the selection of deep learning methods for wildfire prediction remains uncertain due to the lack of comparative analysis in a quantitative and explainable manner, crucial for improving prevention measures and refining models. This study aims to thoroughly compare the performance, efficiency, and explainability of four prevalent deep learning architectures: Autoencoder, ResNet, UNet, and Transformer-based Swin-UNet. Employing a real-world dataset that includes nearly a decade of remote sensing data from California, U.S., these models predict the spread of wildfires for the following day. Through detailed quantitative comparison analysis, we discovered that Transformer-based Swin-UNet and UNet generally outperform Autoencoder and ResNet, particularly due to the advanced attention mechanisms in Transformer-based Swin-UNet and the efficient use of skip connections in both UNet and Transformer-based Swin-UNet, which contribute to superior predictive accuracy and model interpretability. Then we applied XAI techniques on all four models, this not only enhances the clarity and trustworthiness of models but also promotes focused improvements in wildfire prediction capabilities. The XAI analysis reveals that UNet and Transformer-based Swin-UNet are able to focus on critical features such as 'Previous Fire Mask', 'Drought', and 'Vegetation' more effectively than the other two models, while also maintaining balanced attention to the remaining features, leading to their superior performance. The insights from our thorough comparative analysis offer substantial implications for future model design and also provide guidance for model selection in different scenarios.