71446 research outputs found
Sort by
AIlice in Numberland: Comparing numerical understanding in language models and humans through multilingual reasoning puzzles
No full textNumbers in language constitute an extraordinary human cultural innovation. When counting, languages around the world use diverse mathematical strategies to construct and combine their numbers. People learn how to use these systems of numbers despite this diversity. But while large language models (LLMs) appear to independently excel at linguistic and mathematical tasks, they are unable to solve linguistic-mathematical puzzles about systems of numbers in different languages, which humans can learn to solve successfully. This thesis presents a detailed investigation into why this task is difficult for language models.
We design a series of experiments that untangle the linguistic and mathematical aspects of numbers in language, probing at how individual parameters of numeral construction and combination affect model performance. Our experiments establish the novel finding that while individual mathematical features do not hinder the solving ability of current large language models, LLMs are unable to infer the compositional structure of numerals in these problems like humans can. LLMs cannot consistently solve such problems unless the mathematical operations in the problems are explicitly marked using known symbols (+, ×, etc.). Humans are able to use their understanding of numbers in language to make inferences about the implicit compositional structure of numerals — language models seem to lack this notion of numeral structure. We conclude that flexible, adaptive cross-domain use of language appears to remain a challenge for current language models.Computer Scienc
From Portfolio to Precision: Examining the Transition to Multi-Part Bids in European Electricity Markets
No full textThe transition to a decarbonized electricity system in Germany requires fundamental changes in market design to efficiently integrate renewable energy sources (RES) while maintaining grid stability and cost efficiency. This thesis examines the potential adoption of multi-part bids (MPBs) at the unit level and explores how they could improve market efficiency, congestion management, and investment signals compared to Germany’s current portfolio-based bidding system. Through a comparative case study of Italy, which has successfully implemented unit-based MPBs and locational price elements, this research assesses the feasibility of similar reforms in the German electricity market.
Germany’s existing bidding structure, based on a single national price zone and ex-post redispatch, has led to high congestion costs, inefficient price signals, and suboptimal investment decisions. In contrast, Italy’s multi-part bidding framework, integrated with zonal pricing and real-time dispatch mechanisms, enables more transparent cost allocation, reduced redispatch needs, and stronger locational incentives for generation and flexibility assets. The case study highlights key structural differences, including grid topology, regulatory frameworks, and stakeholder adaptation, which influence the applicability of Italy’s model to Germany.
To facilitate a transition, this thesis outlines a phased implementation roadmap tailored to Germany’s market conditions. In the short term, pilot programs in congestion-prone regions could test the feasibility of MPBs, while imbalance pricing reforms could
improve cost allocation. In the medium term, increasing intraday market liquidity and introducing dynamic bidding zones would enhance flexibility. In the long term, a gradual shift toward locational price signals and real-time market integration could improve market efficiency while balancing stakeholder concerns.
This research contributes to the literature on electricity market design and congestion management by analyzing how specific characteristics of national market structures—such as the level of generation centralization, grid topology, and existing bidding formats—influence the effectiveness and feasibility of bidding reforms. The findings suggest that implementing multi-part bids in Germany could significantly reduce redispatch costs, improve market transparency, and support more efficient investment signals. However, the success of such a transition depends on regulatory adjustments, IT infrastructure upgrades, and stakeholder coordination. Future research should incorporate quantitative modeling, political feasibility assessments, and consumer impact analyses to further refine implementation strategies.
By learning from Italy’s experience, Germany can modernize its market design, ensuring a more cost-effective, flexible, and resilient electricity system aligned with European decarbonization goals.Extension Studie
A Taxonomy of Climate Change and Health Solutions Pathways: Development Of A Strategic Policy Framework
No full textClimate change is increasingly being recognized as the single most significant threat facing global health. Urgency is now growing globally to address the health impacts of this threat. However, despite this momentum, existing frameworks fail to integrate climate change and health (CC&H) pathways with solutions and targets. This applied research project aimed to bridge this gap and create a solutions-focused policy tool to support global health practice.
Through a literature review and semi-structured interviews with global philanthropic organizations, the research identified key challenges and opportunities. Findings revealed that CC&H pathways can be well articulated but that a comprehensive current understanding of the burden of disease associated with CC&H pathways remains elusive. In addition, a lack of formal CC&H strategies amongst philanthropic organizations exists. Many organizations are grappling with how to define and evaluate their CC&H work while overlooking existing global development agenda goals and targets. Overall, a persistent state of confusion prevails.
Investment into the conditions for CC&H solutions to emerge independently of direct philanthropic support was identified as a major conclusion of this project. Philanthropy must invest both directly in CC&H solutions AND indirectly in the conditions that will enable CC&H innovation to occur. Investments must focus on building an understanding of the current burden of disease from climate change and identifying universally agreed-upon CC&H global goals and objectives, including the use of existing global development agenda goals and targets.
To address the state of confusion, a results-oriented novel taxonomy was developed. The Taxonomy for Climate Change and Health Solutions Pathways categorizes seven key CC&H pathways. The seven pathways are extreme temperature, food insecurity, mental well-being (encompassing forced displacement), poor air quality, water insecurity, pathogens and vectors, and health systems. Each pathway includes a description of climate change drivers, the impact of drivers on individual pathophysiology or social systems that influence health, health outcomes associated with each pathway, relevant adaptation and mitigation solutions, and Sustainable Development Goals (SDGs) targets.
As a tool, the taxonomy facilitates communication, enhances understanding of the impact of climate change on health, and can help foster collaboration amongst organizations and stakeholders. In turn, the tool can enable evidence-informed decision-making and transformative policymaking. It is believed to be the first tool to integrate CC&H pathways, solutions and metrics, thereby filling a critical gap in the literature and policy-making practice.Public Healt
Housing Price Prediction with Computer Vision and Image Features
No full textArtificial Intelligence (AI) is transforming nearly every facet of human life, and the housing market is no exception. This thesis explores AI to enhance housing price predictions, developing computer vision and machine learning techniques to achieve greater accuracy. Through a series of experiments, our research demonstrates that integrating Multiple Listing Services (MLS) data with property image features and image-based scores improves the performance of housing price prediction models.Extension Studie
Integration of single-cell multimodal data to define the chromatin landscape of rheumatoid arthritis
No full textRheumatoid arthritis (RA) is a prototypical tissue-mediated autoimmune disease and a major healthcare burden. It is characterized by synovial tissue inflammation. While there have been multiple efforts to identify tissue-resident pathogenic cell populations in the RA synovium using single-cell RNA-seq, how these cell populations are epigenetically regulated remains largely unexplored. Open chromatin, measured via single-nucleus assay for transpose-accessibility chromatin using sequencing (snATAC-seq), can reveal the underpinnings of transcriptional regulation across heterogeneous cell states. As with any single-cell technology, the computational integration of multiple snATAC-seq samples into one consistent dataset is a challenge that relies on correcting batch effects while maintaining biological phenotypes. Multimodal datasets, that measure both chromatin accessibility and gene expression, have given us the opportunity to both link transcriptional regulation to its output gene expression as well as measure the performance of snATAC-seq integration against a gold standard snRNA-seq integration.
Utilizing 30 RA synovial tissue samples assayed using unimodal snATAC-seq and multimodal snATAC-seq + snRNA-seq, I developed a computational pipeline to amalgamate single-nucleus open chromatin datasets into 24 cohesive biological cell classes. I then identified putative transcription factors per class, such as STAT3 within a sublining fibroblast class. By integrating with an RA tissue transcriptional atlas, I proposed that these chromatin classes represented ‘superstates’ corresponding to multiple transcriptional cell states. I demonstrated the utility of this RA tissue chromatin atlas through the associations between disease phenotypes and chromatin class abundance as well as the nomination of classes mediating the effects of putatively causal RA genetic variants. Furthermore, I used these RA multimodal datasets and 2 published COVID-19 datasets to benchmark 57 additional computational pipelines encompassing 5 feature types, 7 integration methods, and 1 batch correction method to define effective strategies for multi-sample snATAC-seq integration. Using a command-line tool I developed and 2 novel multimodal metrics, I determined that SnapATAC2 using ATAC-specific features (peaks, cCRE, tiles) with Harmony correction performed best.
This work demonstrates the value of chromatin accessibility studies to discern disease-specific transcriptional regulation. It also highlights how multiple modalities can complement each other. The benchmarking study shows how one modality can be used to assess another. The superstate hypothesis illustrates how both modalities can be integrated together to learn about the relationships between the underlying biological processes. The RA applications show how both modalities can give mechanistic insight into RA pathology. The pipelines I developed here, both for snATAC-seq and multimodal analysis as well as the benchmarking methodology, are broadly applicable to many diseases and will hopefully inspire more investigations into disease-specific chromatin accessibility.Biomedical Informatic
Lifting As We Climb: Community wealth-building as collaborative solutioning, a journey of transformation for systemic impact in a social change organization
No full textIn an era where policymaking has become increasingly gridlocked, social change organizations, like nonprofits and school systems, must take on a more dynamic role—not just as advocates, but as architects of tangible solutions. The mid-20th-century policies that once undergirded widespread economic mobility and social progress emerged from a political climate willing to confront systemic challenges head-on. Today, however, the tools of large-scale governmental intervention have weakened, leaving many of society’s most persistent and perplexing problems—wealth inequality, educational disparities, housing instability—unresolved or even worsening. In this vacuum, collaborative solutioning becomes an imperative: the process of actively designing and implementing bold, innovative strategies to reshape societal possibilities from the ground up through public-private partnerships that are community-driven, unlocking the resources necessary to fix problems quickly. Rather than waiting for legislation that may never materialize, organizations committed to social change must reimagine what is possible, mobilize strategic partnerships, unlock resources, and activate community-driven change that demonstrate alternative futures in real time.
This capstone illustrates the journey of my 10-month residency at the Chicago Urban League (CUL), examining how a prestigious, social change organization can transform its strategic direction and operating model to drive systemic impact. Through collaborating with a core team to conduct a strategic planning process and leading an internal CUL center to demonstrate a case study for reimagining what is possible, I explore how leading organizational change and transformation efforts require adaptive leadership, innovative approaches, and change management. By adapting CUL’s mission from solely economic empowerment to include an aspirational vision for building generational wealth, my strategic project developed a framework for wealth-building as a form of collaborative solutioning, designing and implementing a bold vision of economic prosperity for all through public-private partnerships.
My strategic project and learning experiences at CUL have significant implications for my own practice of exercising leadership, CUL as an organization, and the nonprofit and education sectors as components of the broader social change ecosystem, stressing the need for collaborative solutioning as a liberatory framework to co-create and implement our bold visions of futures where everyone can thrive. This capstone offers those interested in a more just world, where everyone has access to the financial security they need to live fully and be well, a possible pathway towards that transformation.Educatio
Architectures of Relation: A Rhizomatic Pedagogy for Heartland Brazil
No full textThis thesis is a proposition for a new turn in the architectural pedagogy of Brazil, with a focus on addressing the curricula found in schools in the country’s heartland. The prevailing educational model is the continuation of a colonial project, reflecting its xenocentric premise — the devaluation of local culture in favor of foreign values. As a result, the dominant current pedagogical frameworks are misaligned with the specific demands and disciplinary realities of Brazil, especially its mid-sized cities, its heartlands.
This project introduces the concept of Architectures of Relation, developed after Édouard Glissant’s epistemology. Glissant defines Relation as a break from models of alterity, offering a way to understand cultural identity as “not made up of things that are foreign, but of shared knowledge” (Glissant, 1997). While the Western worldview tends to impose transparency, Relation embraces opacities: the untranslatable and culturally specific.
This proposition is informed by a study of the professional context of architecture in Brazil’s mid-sized cities, which are currently the fastest-growing urban centers in the country. To further the studies of pedagogical realities, Escola da Cidade, an independent architecture school in São Paulo, serves as a case study for its pedagogical model tailored to the city it operates within.
Building on the insights from the case study and field analysis, the thesis interprets the parameters of the UNESCO-UIA Charter for Architectural Education under the lens of Glissant’s thought, addressing core architectural realities such as labor, materials, cultural elements, spatial occupation, and climate-responsive design. The final proposed framework is based on pedagogical devices meant to be merged with the core curriculum of schools, aiming to educate architects and urbanists capable of designing cities that are culturally, socially, and ecologically sustainable.Department of Architectur
Association Between Dysfunctional Eating Behaviors and Weight Loss Among Adults in Puerto Rico: A Cross-Sectional Analysis from the PROSPECT Cohort Study
No full textObesity has become one of the most urgent and costly public health crises of the 21st century, with rates climbing at an unprecedented pace across the globe. Since the late 20th century, prevalence has surged due to a complex interplay of environmental, behavioral, and socioeconomic factors. The condition is strongly linked to life-altering non-communicable diseases, including type 2 diabetes, cardiovascular disease, and certain cancers. This contributes to increased illness, premature death, and soaring healthcare costs. By 2035, an estimated 1.53 billion adults worldwide will be classified with obesity.
In the United States, between 2013 and 2016, 66.7% of adults living with obesity and 49% of the adults classified as overweight attempted to lose weight. Unfortunately, research shows that 95% of all dieters fail to achieve weight loss and maintain their results for more than three years. Despite the high level of interest and frequent attempts to lose weight, obesity rates continue to rise. The eating behaviors associated with successful weight loss, particularly dysfunctional behaviors such as emotional or uncontrolled eating, remain unknown among individuals who previously had obesity, especially when success is defined as maintaining weight loss.
It was hypothesized that the participants who experienced weight loss would have a higher association with lower dysfunctional eating scores, such as lower uncontrolled eating and lower emotional eating. To test the hypothesis, this thesis used cross-sectional data (n = 545) from the PROSPECT cohort study of adults aged 30-75 years to investigate the dysfunctional eating behavioral attributes (using a validated three-factor eating questionnaire, TEQ R-18) linked to weight loss. The PROSPECT data included clinic-based anthropometric measurements, along with self-reported data on eating behaviors, lifestyle factors, demographics, and socioeconomic status. All eligible participants were assigned to either a Weight Loss Group based on a self-reported question asking if the participant had intentionally lost at least 10 pounds in the previous 6-months) or No Weight Loss (combining Normal Weight Group (no weight loss and BMI .0 kg/m2) and High Weight Group (no weight loss and BMI ≥25.0 kg/m2)). Logistic models (unadjusted and adjusted for sociodemographic, behavioral, and health factors) were run to determine the likelihood of experiencing dysfunctional eating behaviors (i.e., emotional eating (EE), uncontrolled eating (UE), cognitive restraint (CR), and overall dysfunctional eating (TEQ)) by weight loss status.
The unadjusted results showed that participants who lost a minimum of 10 pounds exhibited a lower association with UE, but a higher association with CR, EE, and overall TEQ scores. In the unadjusted analysis, high UE was associated with lower odds of being in the Weight Loss Group compared to participants with no to moderate UE scores. High EE was associated with higher odds of being in the Weight Loss Group compared to participants with no to moderate EE scores. High CR was associated with significantly higher odds of being in the Weight Loss Group compared to participants with no to moderate CR scores. After adjusting for demographic, socioeconomic, lifestyle, and health-related covariates, the associations for UE and EE were largely attenuated and non-significant. No significant interactions between sex and UE, CR, EE, or TEQ were observed in the adjusted models. The adjusted model did display significant results, indicating that higher CR is associated with higher odds of being in the Weight Loss Group. Furthermore, current alcohol consumption is associated with significantly lower odds of being in the Weight Loss Group. Though significant results were observed, further exploration of eating behaviors related to successful weight loss maintenance needs to be evaluated by performing a longitudinal data analysis on a larger sample size.Extension Studie
Transforming City Operations with StatGPT
No full textCities today face mounting pressure to deliver more effective services using fewer resources. To meet these challenges, performance management systems must evolve. This paper proposes “StatGPT,” a next-generation, technology-enabled approach that builds on the successes of traditional Stat models by injecting them with modern artificial intelligence (AI) and data capabilities. A responsive city measures what matters and then does something about it. Yet, the management systems that determine what should be measured, what data to utilize, and by whom, require modernization. That is especially true now as demands on local government continue to exceed resources, and the legitimacy of elected officials depends on fulfilling residents' expectations. Delivering on daily tasks that matter to residents creates a “responsive cycle,” where trust builds the legitimacy necessary for mayors to rally followers in times of crisis or towards large aspirations.
Achieving these goals requires a new approach that leverages the power of generative AI (GenAI), the Internet of Things, ubiquitous mobile devices, open data, and advanced analytics. This new approach requires modernizing the Stat systems that cities have widely adopted over the past two decades, dating back to when New York City Police Chief Bill Bratton introduced CompStat to manage crime-fighting efforts and Baltimore Mayor Martin O’Malley applied CitiStat to municipal services. The Stat model combined data with accountability in a way city leaders had not seen before. Precinct commanders in New York, or department heads in Baltimore, would attend a high-level meeting where they would present their latest data on crime or service levels, which would be projected on large screens for everyone in the room to see. Then, senior executives – and sometimes Bratton or O’Malley themselves – would grill managers about the numbers. Why are robberies up on this block? Why are restaurant inspections down this quarter? Often, there was a theatrical element to these meetings. These events proved effective in keeping managers focused on results and surfacing hidden problems. Crime in New York plunged. Baltimore saved millions on employee overtime related to chronic absenteeism. When utilized correctly, Stat is today’s gold standard for performance management. Bob Behn, my colleague at the Harvard Kennedy School and an expert on Stat programs, underscores that using the model effectively requires hands-on leadership, timely data, regular meetings, active follow-through, a constructive culture, constant learning, and flexibility. These principles remain as cities rethink Stat, but now they are merely the table stakes for the broader transformation that leverages technology. This paper proposes a new approach to measuring and utilizing performance data as the gateway to management changes that lead to operational excellenceVersion of Recor
Driving Security? U.S. Auto Industrial Policy in a Changing World
No full textIndustrial policy may have come back into fashion—controversially—in the last few years, but it was arguably never out for the auto sector. Auto manages to have outsized political importance in the U.S., and around the world, thanks to its relatively high paying jobs for workers without college education, unionization, and the political romance around manufacturing jobs. That political pull wins the auto sector repeat special treatment. Today’s U.S. industrial policy is not about picking winners for export potential, unlike 20th century industrial policy, but about economic security. It aims either to build domestic capacity in industries viewed as pivotal (semiconductors), achieve geopolitically secure supply by reducing dependence on China (critical minerals, rare earths), or protect politically important domestic industries and their share in the U.S. market (autos). Today is different. Auto is transforming rapidly in the shift to electric and autonomous vehicles (EVs and AVs)—a change that requires different engineering and supply chains. Automakers around the world continue to emphasize that EVs are the future of the industry, although Toyota is a notable skeptic. The EV shift is upending a long-stable competitive landscape. China’s indigenous EV automakers have become major producers and exporters in a handful of years. The Chinese government is pervasively involved in the growth of its EV industry, while having developed a national monopoly position in global EV supply chains. Adding to security considerations, domestic auto manufacturing offers contingent defense manufacturing capability—as has long been true— while the computerization of vehicles raises new surveillance and cyberattack risks. This article examines today’s U.S. auto industrial policy in international and technological context, reviews the short track record of Biden EV policies, discusses how we should evaluate Trump’s auto industrial policy, and closes with recommendations for where industrial policy would be best focused to support the U.S. auto sector.Version of Recor