Research Papers

We consider the task of enforcing individual fairness in gradient boosting. Gradient boosting is a popular method for machine learning from tabular data, which arise often in applications where algorithmic fairness is a concern. At a high level, our approach is a functional gradient descent on a (distributionally) robust loss function that encodes our intuition of algorithmic fairness for the ML task at hand. Unlike prior approaches to individual fairness that only work with smooth ML models, our approach also works with non-smooth models such as decision trees. We show that our algorithm converges globally and generalizes. We also demonstrate the efficacy of our algorithm on three ML problems susceptible to algorithmic bias.

A Scalable Two Stage Approach to Computing Optimal Decision Sets

Alexey Ignatiev, Edward Lam, Peter J. Stuckey +1

Machine learning (ML) is ubiquitous in modern life. Since it is being deployed in technologies that affect our privacy and safety, it is often crucial to understand the reasoning behind its decisions, warranting the need for explainable AI. Rule-based models, such as decision trees, decision lists, and decision sets, are conventionally deemed to be the most interpretable. Recent work uses propositional satisfiability (SAT) solving (and its optimization variants) to generate minimum-size decision sets. Motivated by limited practical scalability of these earlier methods, this paper proposes a novel approach to learn minimum-size decision sets by enumerating individual rules of the target decision set independently of each other, and then solving a set cover problem to select a subset of rules. The approach makes use of modern maximum satisfiability and integer linear programming technologies. Experiments on a wide range of publicly available datasets demonstrate the advantage of the new approach over the state of the art in SAT-based decision set learning.

There appears to be a common agreement that ethical concerns are of high importance when it comes to systems equipped with some sort of Artificial Intelligence (AI). Demands for ethical AI are declared from all directions. As a response, in recent years, public bodies, governments, and universities have rushed in to provide a set of principles to be considered when AI based systems are designed and used. We have learned, however, that high-level principles do not turn easily into actionable advice for practitioners. Hence, also companies are publishing their own ethical guidelines to guide their AI development. This paper argues that AI software is still software and needs to be approached from the software development perspective. The software engineering paradigm has introduced maturity model thinking, which provides a roadmap for companies to improve their performance from the selected viewpoints known as the key capabilities. We want to voice out a call for action for the development of a maturity model for AI software. We wish to discuss whether the focus should be on AI ethics or, more broadly, the quality of an AI system, called a maturity model for the development of AI systems.

Transdisciplinary AI Observatory -- Retrospective Analyses and Future-Oriented Contradistinctions

Nadisha-Marie Aliman, Leon Kester, Roman Yampolskiy

In the last years, AI safety gained international recognition in the light of heterogeneous safety-critical and ethical issues that risk overshadowing the broad beneficial impacts of AI. In this context, the implementation of AI observatory endeavors represents one key research direction. This paper motivates the need for an inherently transdisciplinary AI observatory approach integrating diverse retrospective and counterfactual views. We delineate aims and limitations while providing hands-on-advice utilizing concrete practical examples. Distinguishing between unintentionally and intentionally triggered AI risks with diverse socio-psycho-technological impacts, we exemplify a retrospective descriptive analysis followed by a retrospective counterfactual risk analysis. Building on these AI observatory tools, we present near-term transdisciplinary guidelines for AI safety. As further contribution, we discuss differentiated and tailored long-term directions through the lens of two disparate modern AI safety paradigms. For simplicity, we refer to these two different paradigms with the terms artificial stupidity (AS) and eternal creativity (EC) respectively. While both AS and EC acknowledge the need for a hybrid cognitive-affective approach to AI safety and overlap with regard to many short-term considerations, they differ fundamentally in the nature of multiple envisaged long-term solution patterns. By compiling relevant underlying contradistinctions, we aim to provide future-oriented incentives for constructive dialectics in practical and theoretical AI safety research.

Identifying anomalies in the fuel consumption of the vehicles of a fleet is a crucial aspect for optimizing consumption and reduce costs. However, this information alone is insufficient, since fleet operators need to know the causes behind anomalous fuel consumption. We combine unsupervised anomaly detection techniques, domain knowledge and interpretable Machine Learning models for explaining potential causes of abnormal fuel consumption in terms of feature relevance. The explanations are used for generating recommendations about fuel optimization, that are adjusted according to two different user profiles: fleet managers and fleet operators. Results are evaluated over real-world data from telematics devices connected to diesel and petrol vehicles from different types of industrial fleets. We measure the proposal regarding model performance, and using Explainable AI metrics that compare the explanations in terms of representativeness, fidelity, stability, contrastiveness and consistency with apriori beliefs. The potential fuel reductions that can be achieved is round 35%.

Measuring Social Biases of Crowd Workers using Counterfactual Queries

Bhavya Ghai, Q. Vera Liao, Yunfeng Zhang +1

Social biases based on gender, race, etc. have been shown to pollute machine learning (ML) pipeline predominantly via biased training datasets. Crowdsourcing, a popular cost-effective measure to gather labeled training datasets, is not immune to the inherent social biases of crowd workers. To ensure such social biases aren't passed onto the curated datasets, it's important to know how biased each crowd worker is. In this work, we propose a new method based on counterfactual fairness to quantify the degree of inherent social bias in each crowd worker. This extra information can be leveraged together with individual worker responses to curate a less biased dataset.

Being a complex subject of major importance in AI Safety research, value alignment has been studied from various perspectives in the last years. However, no final consensus on the design of ethical utility functions facilitating AI value alignment has been achieved yet. Given the urgency to identify systematic solutions, we postulate that it might be useful to start with the simple fact that for the utility function of an AI not to violate human ethical intuitions, it trivially has to be a model of these intuitions and reflect their variety $ - $ whereby the most accurate models pertaining to human entities being biological organisms equipped with a brain constructing concepts like moral judgements, are scientific models. Thus, in order to better assess the variety of human morality, we perform a transdisciplinary analysis applying a security mindset to the issue and summarizing variety-relevant background knowledge from neuroscience and psychology. We complement this information by linking it to augmented utilitarianism as a suitable ethical framework. Based on that, we propose first practical guidelines for the design of approximate ethical goal functions that might better capture the variety of human moral judgements. Finally, we conclude and address future possible challenges.

A Psychopathological Approach to Safety Engineering in AI and AGI

Vahid Behzadan, Arslan Munir, Roman V. Yampolskiy

The complexity of dynamics in AI techniques is already approaching that of complex adaptive systems, thus curtailing the feasibility of formal controllability and reachability analysis in the context of AI safety. It follows that the envisioned instances of Artificial General Intelligence (AGI) will also suffer from challenges of complexity. To tackle such issues, we propose the modeling of deleterious behaviors in AI and AGI as psychological disorders, thereby enabling the employment of psychopathological approaches to analysis and control of misbehaviors. Accordingly, we present a discussion on the feasibility of the psychopathological approaches to AI safety, and propose general directions for research on modeling, diagnosis, and treatment of psychological disorders in AGI.

From What to Why: A Multi-Agent System for Evidence-based Chemical Reaction Condition Reasoning

Cheng Yang, Jiaxuan Lu, Haiyuan Wan +2

The chemical reaction recommendation is to select proper reaction condition parameters for chemical reactions, which is pivotal to accelerating chemical science. With the rapid development of large language models (LLMs), there is growing interest in leveraging their reasoning and planning capabilities for reaction condition recommendation. Despite their success, existing methods rarely explain the rationale behind the recommended reaction conditions, limiting their utility in high-stakes scientific workflows. In this work, we propose ChemMAS, a multi-agent system that reframes condition prediction as an evidence-based reasoning task. ChemMAS decomposes the task into mechanistic grounding, multi-channel recall, constraint-aware agentic debate, and rationale aggregation. Each decision is backed by interpretable justifications grounded in chemical knowledge and retrieved precedents. Experiments show that ChemMAS achieves 20-35% gains over domain-specific baselines and outperforms general-purpose LLMs by 10-15% in Top-1 accuracy, while offering falsifiable, human-trustable rationales, which establishes a new paradigm for explainable AI in scientific discovery.

As artificial intelligence (AI) tools become widely adopted, large language models (LLMs) are increasingly involved on both sides of decision-making processes, ranging from hiring to content moderation. This dual adoption raises a critical question: do LLMs systematically favor content that resembles their own outputs? Prior research in computer science has identified self-preference bias -- the tendency of LLMs to favor their own generated content -- but its real-world implications have not been empirically evaluated. We focus on the hiring context, where job applicants often rely on LLMs to refine resumes, while employers deploy them to screen those same resumes. Using a large-scale controlled resume correspondence experiment, we find that LLMs consistently prefer resumes generated by themselves over those written by humans or produced by alternative models, even when content quality is controlled. The bias against human-written resumes is particularly substantial, with self-preference bias ranging from 67% to 82% across major commercial and open-source models. To assess labor market impact, we simulate realistic hiring pipelines across 24 occupations. These simulations show that candidates using the same LLM as the evaluator are 23% to 60% more likely to be shortlisted than equally qualified applicants submitting human-written resumes, with the largest disadvantages observed in business-related fields such as sales and accounting. We further demonstrate that this bias can be reduced by more than 50% through simple interventions targeting LLMs' self-recognition capabilities. These findings highlight an emerging but previously overlooked risk in AI-assisted decision making and call for expanded frameworks of AI fairness that address not only demographic-based disparities, but also biases in AI-AI interactions.

Recent advancements in large language model (LLM) agents have significantly accelerated scientific discovery automation, yet concurrently raised critical ethical and safety concerns. To systematically address these challenges, we introduce \textbf{SafeScientist}, an innovative AI scientist framework explicitly designed to enhance safety and ethical responsibility in AI-driven scientific exploration. SafeScientist proactively refuses ethically inappropriate or high-risk tasks and rigorously emphasizes safety throughout the research process. To achieve comprehensive safety oversight, we integrate multiple defensive mechanisms, including prompt monitoring, agent-collaboration monitoring, tool-use monitoring, and an ethical reviewer component. Complementing SafeScientist, we propose \textbf{SciSafetyBench}, a novel benchmark specifically designed to evaluate AI safety in scientific contexts, comprising 240 high-risk scientific tasks across 6 domains, alongside 30 specially designed scientific tools and 120 tool-related risk tasks. Extensive experiments demonstrate that SafeScientist significantly improves safety performance by 35\% compared to traditional AI scientist frameworks, without compromising scientific output quality. Additionally, we rigorously validate the robustness of our safety pipeline against diverse adversarial attack methods, further confirming the effectiveness of our integrated approach. The code and data will be available at https://github.com/ulab-uiuc/SafeScientist. \textcolor{red}{Warning: this paper contains example data that may be offensive or harmful.}

Trends in AI Supercomputers

Konstantin F. Pilz, James Sanders, Robi Rahman +1

Frontier AI development relies on powerful AI supercomputers, yet analysis of these systems is limited. We create a dataset of 500 AI supercomputers from 2019 to 2025 and analyze key trends in performance, power needs, hardware cost, ownership, and global distribution. We find that the computational performance of AI supercomputers has doubled every nine months, while hardware acquisition cost and power needs both doubled every year. The leading system in March 2025, xAI's Colossus, used 200,000 AI chips, had a hardware cost of \$7B, and required 300 MW of power, as much as 250,000 households. As AI supercomputers evolved from tools for science to industrial machines, companies rapidly expanded their share of total AI supercomputer performance, while the share of governments and academia diminished. Globally, the United States accounts for about 75% of total performance in our dataset, with China in second place at 15%. If the observed trends continue, the leading AI supercomputer in 2030 will achieve $2\times10^{22}$ 16-bit FLOP/s, use two million AI chips, have a hardware cost of \$200 billion, and require 9 GW of power. Our analysis provides visibility into the AI supercomputer landscape, allowing policymakers to assess key AI trends like resource needs, ownership, and national competitiveness.