University of Illinois Urbana-Champaign
Illinois Digital Environment for Access to Learning and Scholarship RepositoryNot a member yet
123813 research outputs found
Sort by
News Sources
No full textThis chapter discusses a variety of news source formats and types. This chapter also examines how to evaluate news sources to successfully direct patrons to reliable information that will meet their needs. It also provides opportunities for library professionals to reflect on their own interaction with news sources and consider how that may impact what sources they recommend
Urban Heat Mitigation Effects and Greenery Justice in Washington, D.C.
No full textThe deterioration of urban heat exposure risks is highlighting the benefits of urban greenery for creating healthy and resilient cities. Using Washington, D.C. as a case, this study demonstrates the urban temperature mitigation effect and effective ranges of GVI and NDVI at street-scale, examines greenery justice in residential zones and affordable housing surroundings, the differentiated cooling effects of different genera
Life Inside the POC: Field Realities of UN Personnel in South Sudan During the Rainy Season
No full textThis field account documents the daily challenges faced by United Nations personnel operating in a remote and insecure area of South Sudan. The testimony captures logistical, environmental, and security constraints encountered within and around a Protection of Civilians (POC) site during the rainy season. Interviewees describe extreme isolation, unpredictable weather, limited transport reliability, and the constant threat posed by armed combatants breaching perimeter fences. Additional hardships include inadequate infrastructure, food scarcity, and poor living conditions—exacerbated by mud, flooding, and minimal access to sanitation. The narrative offers an unfiltered portrayal of humanitarian fieldwork under duress, highlighting the intersection of operational resilience, safety risk, and human endurance in conflict environments
Certifying robustness in inference and learning problems
Get PDFThere is a rich literature of algorithms for inference, prediction, and decision-making problems when the underlying distributions governing the data are known and well-modeled. The research from the past few decades has provided powerful learning algorithms when such distributions cannot be easily modeled, for instance with high dimensional data. As a result, such data driven methods are becoming ubiquitous in a wide variety of real-life applications, including safety-critical ones such as self-driving and medical diagnosis. In order for the reliable and safe deployment of such algorithms in practice, there exist several imminent questions to be answered. In this dissertation, a few topics in robustness of inference and learning methods are studied. One of the key issues with data-driven methods in practice is unexpected changes that can occur at inference time potentially affecting the performance of these methods, for instance, deviations in the data generating distributions, irrelevant or unrecognizable inputs, and adversarial attacks from unknown sources. The underlying theme connecting the topics studied in this dissertation is the development of learning algorithms robust to such unexpected, potentially harmful, deviations. Broadly, three problems in robust inference and learning are discussed in this dissertation - out-of-distribution detection for machine learning models, detection robust to distribution shifts, and multi-player multi-armed bandits robust to adversarial attacks. Principled approaches for these problems with theoretical guarantees are derived using tools from statistics, information theory and optimization, that are practical, resilient and efficiently implementable.Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-10-19 without embargo termsThe student, Akshayaa Magesh, accepted the attached license on 2025-02-04 at 04:25.The student, Akshayaa Magesh, submitted this Dissertation for approval on 2025-02-04 at 04:43.This Dissertation was approved for publication on 2025-02-07 at 12:49.DSpace SAF Submission Ingestion Package generated from Vireo submission #21630 on 2025-10-19 at 18:08:4
Examining large language models for safety and robustness through the lens of social science
Get PDFSubmission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-10-19 without embargo termsThe student, Sullam Jeoung, accepted the attached license on 2025-03-05 at 10:51.The student, Sullam Jeoung, submitted this Dissertation for approval on 2025-03-05 at 10:55.This Dissertation was approved for publication on 2025-03-07 at 10:36.DSpace SAF Submission Ingestion Package generated from Vireo submission #21662 on 2025-10-19 at 18:09:09Large language models have demonstrated remarkable capabilities, often achieving human-like performance levels and significantly impacting our daily lives. However, these models can perpetuate and amplify harmful stereotypes and biases associated with socio-demographic representations, potentially generating discriminatory content that adversely affects individuals and communities. Given these risks and their broader societal implications, ensuring the safety and robustness of these models through the identification and mitigation of harmful stereotypes has become imperative. This dissertation presents comprehensive methodologies to address these challenges by integrating insights from social science, psychology, and cognitive studies with methods from natural language processing. First, we present a framework to assess human-like stereotypical patterns in large language models (LLMs), drawing upon established psychological theories of how individuals develop stereotypes toward various social groups. This theoretically-grounded approach provides construct validity in defining and measuring stereotypes. The framework incorporates three key dimensions: warmth-competence analysis, keyword- reasoning patterns, and emotional-behavioral responses. Through clustering analysis, keyword extraction, and reasoning pattern evaluation of LLM responses, we examine how these models align with or deviate from documented human behavioral patterns. Our findings reveal that LLMs demonstrate nuanced perceptions of social groups, consistent with psychological research highlighting the multifaceted nature of stereotypes. Notably, the models’ reasoning patterns, particularly regarding groups’ economic status, demonstrate a nuanced awareness of societal disparities. Second, we propose methods to examine causal sensitivity of language models on socio-demographic attributes. This is based on a controlled experimental framework that uses name frequency analysis from U.S. Census data and systematic evaluation of model predictions through causal graphs. Our findings show that less frequent first names lead to divergent model predictions, highlighting the need for careful demographic consideration in dataset design to ensure fair and consistent model performance across different name representations. Third, we investigate how LLMs exhibit and inflate political stereotypes through the lens of cognitive biases and representative heuristics. We analyze LLMs’ responses using two key theoretical frameworks: ’kernel of truth’ (whether stereotypes reflect empirical realities) and ’representative heuristics’ (whether models overemphasize representative attributes of target groups), comparing model outputs with actual human responses across various political topics. Our findings show that while LLMs can accurately mimic certain political positions, they tend to exaggerate these positions compared to empirical human responses, suggesting a vulnerability to stereotypical thinking similar to human cognitive biases. This implies the need for careful consideration of cognitive bias frameworks in developing and deploying language models, particularly in politically sensitive contexts, and demonstrates the potential effectiveness of prompt-based mitigation strategies in reducing stereotypical responses. Overall, this dissertation enhances the understanding and safety of LLMs by proposing a framework to assess human-like stereotypes, methods to evaluate causal sensitivity based on socio-demographic attributes, and an analysis of political stereotypes through cognitive bias frameworks. By integrating insights from psychology and social sciences with computational methods this research makes a contribution to the ongoing discourse on ethical AI deployment, highlighting the necessity of understanding and addressing biases in language models to promote fairness and reduce discriminatory outcomes
Crafting safe human-centric agents with risk intelligence
Get PDFSubmission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-10-19 without embargo termsThe student, Chenkai Sun, accepted the attached license on 2025-04-17 at 17:14.The student, Chenkai Sun, submitted this Dissertation for approval on 2025-04-17 at 17:19.This Dissertation was approved for publication on 2025-04-18 at 10:30.DSpace SAF Submission Ingestion Package generated from Vireo submission #21745 on 2025-10-19 at 18:09:18Safety has become a critical concern in digital communication environments, where harmful content can propagate rapidly and compromise societal well-being. As artificial intelligence increasingly serves as an automated content creation and distribution mechanism, the risks of adverse algorithmic impacts are amplified when these systems lack awareness of how their outputs influence audiences. This dissertation addresses this challenge by developing an integrated framework for communication risk management that enables systems to assess, anticipate, and mitigate potential negative consequences. The framework comprises four interconnected research contributions. First, we establish the foundations for risk assessment through a novel task formulation and dataset that captures how identical messages affect diverse user personas differently. This approach transcends traditional content moderation by evaluating information safety and appropriateness for different population groups, creating a measurement capability essential for risk-aware AI systems. Furthermore, we address the challenge of evaluating risk for users with minimal digital footprints, often referred to as "lurkers". By leveraging social graphs constructed through large language models, this work presents a solution for more accurately predicting opinions from such users, expanding the applicability of personalized agents in risk management. Another cornerstone of this dissertation is the development of efficient strategies for language model personalization in risk assessment. Through hierarchical and collaborative data refinement techniques, our Persona-DB approach achieves high-accuracy personalization with substantially reduced data retrieval requirements. This work bridges the gap between risk management and practical deployment by addressing the computational costs of personalization. Lastly, we extend risk assessment beyond immediate impacts to encompass temporal dynamics and cascading effects. By employing language models as social simulators, our framework projects how content might influence populations over time, enabling the anticipation of long-term consequences that traditional safety approaches neglect. This capability not only enhances risk assessment but also provides a mechanism for aligning content-generating AI with broader safety considerations
360° and 2D video analytics in network and energy constrained environments
Get PDFSubmission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-10-19 without embargo termsThe student, Benjamin Civjan, accepted the attached license on 2025-04-14 at 11:14.The student, Benjamin Civjan, submitted this Thesis for approval on 2025-04-14 at 12:19.This Thesis was approved for publication on 2025-04-15 at 06:13.DSpace SAF Submission Ingestion Package generated from Vireo submission #21763 on 2025-10-19 at 18:09:22Real-time video analytics enables rapid, automated content understanding, significantly reducing the time required to search through captured footage. However, many real-world scenarios that could benefit from real-time analytics face constraints that remain underexplored. This thesis focuses on one such use case: firefighter training. Firefighters rely on video for decision-making, post-mission feedback, and developing new training scenarios. However, deploying cameras in outdoor training environments presents two major challenges: limited network connectivity due to the distance from Wi-Fi infrastructure and restricted energy availability as cameras operate on battery power. To investigate network limitations, we conducted field tests at the Illinois Fire Service Institute (IFSI) to analyze connectivity from various locations on the grounds. Additionally, we developed a streaming framework supporting multiple video codecs (MJPEG, WebP, Tiled MJPEG, and H.264) and systematically evaluated their impact on bandwidth usage and 360° video streaming performance under real-world conditions at IFSI. To address the problem of energy-efficient video processing we developed a system, EcoLens, that dynamically optimizes processing configurations to minimize energy consumption of the camera while preserving essential video features for deep learning inference. We first conducted an extensive offline evaluation of various configurations comprising of device CPU frequency, frame filtering features, difference thresholds, and video bitrates, to establish apriori knowledge of their impact on energy consumption and inference accuracy. Leveraging this insight, we introduced an online system that employs multi-objective Bayesian optimization to intelligently explore and adapt configurations in real time. Our approach continuously refines processing settings to meet target inference accuracy with minimal edge device energy expenditure. Experimental results demonstrated the system’s effectiveness in reducing video processing energy use while maintaining high analytical performance, offering a practical solution for smart devices and edge computing applications
Hercules: A compiler for productive programming of heterogeneous systems
Get PDFSubmission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-10-19 without embargo termsThe student, Russel Arbore, accepted the attached license on 2025-04-15 at 20:55.The student, Russel Arbore, submitted this Thesis for approval on 2025-04-15 at 21:01.This Thesis was approved for publication on 2025-04-16 at 07:41.DSpace SAF Submission Ingestion Package generated from Vireo submission #21788 on 2025-10-19 at 18:09:24Modern computing systems increasingly rely on composing heterogeneous devices to improve performance and efficiency. Programming these systems is often unproductive: algorithm implementations must be coupled to system-specific logic, including device-specific optimizations, partitioning, and inter-device communication and synchronization, which requires developing different programs for different system configurations. We propose the Juno language, which represents general purpose applications in an imperative form that can be transformed into parallel, optimized, system-specific code using an expressive and granular imperative scheduling language. We also introduce the Hercules compiler, which uses a novel intermediate representation to represent general and device-specific parallel code in a manner that is easy to analyze and manipulate using schedules. Our system achieves competitive performance with hand-optimized device-specific code (geomean speedups of 1.25x and 1.48x on the CPU and GPU) and significantly outperforms a prior general purpose heterogeneous programming system (geomean speedups of 9.31x and 16.18x on the CPU and GPU)
Event-based knowledge editing for deterministic knowledge propagation in large language models
Get PDFSubmission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-10-19 without embargo termsThe student, Jiateng Liu, accepted the attached license on 2025-04-17 at 10:21.The student, Jiateng Liu, submitted this Thesis for approval on 2025-04-17 at 10:21.This Thesis was approved for publication on 2025-04-17 at 13:26.DSpace SAF Submission Ingestion Package generated from Vireo submission #21819 on 2025-10-19 at 18:09:31The dynamic nature of real-world information necessitates knowledge editing (KE) in large language models (LLMs). This edited knowledge should propagate and facilitate the deduction of new information based on existing model knowledge. We define the existing related knowledge in a LLM serving as the origination of knowledge propagation as ``deduction anchors''. However, most of current KE approaches only operate on (subject, relation, object) triples. Both theoretically and empirically, we observe that this simplified setting often leads to uncertainty when determining the deduction anchors, causing low confidence in their responses. To mitigate this issue, we propose a novel task of event-based knowledge editing that pairs facts with event descriptions. This task manifests both as a closer simulation of real-world editing scenarios and a more logically sound setting, implicitly defining the deduction anchor and enabling LLMs to propagate knowledge confidently. We curate a new benchmark dataset evedit derived from the CounterFact dataset and validate its superiority in improving model confidence. Moreover, as we observe that the event-based setting is notably challenging for existing approaches, we propose a novel approach Self-Edit that showcases stronger performance, achieving 55.6\% consistency improvement while maintaining the naturalness of generation
Safety, security, and rehabilitation? An analysis of publication censorship practices and previously imprisoned individuals’ experiences in Illinois
Get PDFSubmission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-10-19 without embargo termsThe student, Carileigh Jones, accepted the attached license on 2025-04-21 at 20:32.The student, Carileigh Jones, submitted this Dissertation for approval on 2025-04-21 at 20:38.This Dissertation was approved for publication on 2025-04-25 at 13:18.DSpace SAF Submission Ingestion Package generated from Vireo submission #21860 on 2025-10-19 at 18:09:35Scholarly discourses surrounding publication censorship policies and practices have virtually ignored the voices of the imprisoned. While these discourses help us to understand, and critique, the idea that book censorship promotes rehabilitation, safety, and security, they offer top-down analysis, which take for granted the viewpoints and experiences of those who have been incarcerated. Departing from this line of thinking, my dissertation theorizes the censorship of reading material within U.S. prisons as a practice of carcerality, which not only defines and regulates the limits of what can be known and expressed, but may also be experienced by prisoners as control and surveillance of both the mind and body (Friedman, 2021). Located at the intersections of sociology, critical criminology, and carceral studies, this dissertation analyzes interviews, court case documents, and policy documents to place explanations of the purpose and goals of censorship from the perspective of prison officials, in conversation with the perceptions and experiences of those who have been incarcerated. I argue that centering the voices of the imprisoned when implementing and assessing censorship guidelines is crucial if the purpose of books, and prison reform more generally, is to improve the lives and experiences of the imprisoned. Taking their perspectives into account could 1) enhance policy work related to book censorship practices, 2) promote a level of agency in the lives of the imprisoned, and 3) reshape the ways we understand discipline and rehabilitation within and beyond prison walls