9607 research outputs found
Sort by
Trading off performance and human oversight in algorithmic policy: evidence from Danish college admissions
No full textStudent dropout is a significant concern for educational institutions due to its social and economic impact, driving the need for risk prediction systems to identify at-risk students before enrollment. We explore the accuracy of such systems in the context of higher education by predicting degree completion before admission, with potential applications for prioritizing admissions decisions. Using a large-scale dataset from Danish higher education admissions, we demonstrate that advanced sequential AI models offer more precise and fair predictions compared to current practices that rely on either high school grade point averages or human judgment. These models not only improve accuracy but also outperform simpler models, even when the simpler models use protected sociodemographic attributes. Importantly, our predictions reveal how certain student profiles are better matched with specific programs and fields, suggesting potential efficiency and welfare gains in public policy. We estimate that even the use of simple AI models to guide admissions decisions, particularly in response to a newly implemented nationwide policy reducing admissions by 10 percent, could yield significant economic benefits. However, this improvement would come at the cost of reduced human oversight and lower transparency. Our findings underscore both the potential and challenges of incorporating advanced AI into educational policymaking
RomanLens: The Role Of Latent Romanization In Multilinguality In LLMs.
No full textLarge Language Models (LLMs) exhibit strong multilingual performance despite being predominantly trained on English-centric corpora. This raises a fundamental question: How do LLMs achieve such multilingual capabilities? Focusing on languages written in non-Roman scripts, we investigate the role of Romanization—the representation of non-Roman scripts using Roman characters—as a potential bridge in multilingual processing. Using mechanistic interpretability techniques, we analyze next-token generation and find that intermediate layers frequently represent target words in Romanized form before transitioning to native script, a phenomenon we term Latent Romanization. Further, through activation patching experiments, we demonstrate that LLMs encode semantic concepts similarly across native and Romanized scripts, suggesting a shared underlying representation. Additionally, for translation into non-Roman script languages, our findings reveal that when the target language is in Romanized form, its representations emerge earlier in the model’s layers compared to native script. These insights contribute to a deeper understanding of multilingual representation in LLMs and highlight the implicit role of Romanization in facilitating language transfer
BikeNodePlanner: A data-driven decision support tool for bicycle node network planning
No full textA bicycle node network is a wayfinding system targeted at recreational cyclists, consisting of numbered signposts placed alongside already existing infrastructure. Bicycle node networks are becoming increasingly popular as they encourage sustainable tourism and rural cycling, while also being flexible and cost-effective to implement. However, the lack of a formalized methodology and data-driven tools for the planning of such networks is a hindrance to their adaptation on a larger scale. To address this need, we present the BikeNodePlanner: A fully open-source decision support tool, consisting of modular Python scripts to be run in the free and open-source geographic information system QGIS. The BikeNodePlanner allows the user to evaluate and compare bicycle node network plans through a wide range of metrics, such as land use, proximity to points of interest, and elevation across the network. The BikeNodePlanner provides data-driven decision support for bicycle node network planning and can hence be of great use for regional planning, cycling tourism, and the promotion of rural cycling
When Less Human Is More: Introducing Stigma Sensitivity to Anthropomorphism Research in AI
No full textRDFMutate : Mutation-Based Generation of Knowledge Graphs
No full textThis paper introduces RDFMutate, the first mutation-based tool to generate RDF knowledge graphs. RDFMutate enables developers to analyze the robustness of applications that operate on RDF, by generating variants of seed knowledge graphs, which are mutated according to a set of mutation operations and SHACL constraints. In contrast to existing tools to generate synthetic RDF graphs, RDFMutate provides a mutation-based approach that is accessible for both researchers and application developers, by providing a framework to define mutation rules and flexible selection of generated graphs
Exploring the Zero-Shot Known-Item Retrieval Capabilities of LLMs for Casual Leisure Information Needs
No full textThe rapidly increasing popularity of LLM-powered chatbots has led to them being used for a increasing number of different tasks by the general public. One of these tasks is searching for information instead of using a search engine. Previous work has shown that complex search tasks can be problematic for traditional search engines to solve, but little is known about the capability of LLMs on the same task. We compared four LLMs on their capability to answer a specific type of complex search task: known-item requests from casual leisure domains. We constructed a test collection by gathering known-item requests for books, games and movies from online forums along with verified answers by the original requester. We prompted four LLMs multiple times with the same prompt and analyzed the results with respect to accuracy and the degree to which answers were fabricated by the LLM. Our results show that LLMs are not particularly effective in fulfilling these complex casual leisure needs, but there are are big differences between LLMs and across domains.<br/
It's Always Better When We're Together: Physiological Synchrony and Shared Experiences in Social VR
No full textWhen people share experiences in physically co-located environments, such as attending a music experience, their physiological responses align, reflecting a shared social experience through synchrony. In virtual reality (VR), however, it remains unclear whether physiological synchrony similarly reflects the quality of social experiences. Our study examines how different sensory attendance modalities affect physiological synchrony and perceived shared experience during social VR music experiences. In a within-subject design, participants (N = 66) watched 16 music video clips in dyads, in four attendance modalities: Alone, Voice, Avatar, and Avatar+Voice. We measured physiological responses (electrodermal activity, heart rate, and respiration rate) and subjective measures of presence, social presence, co-presence, social connectedness, user experience, and emotional response. Attending the experience with access to social cues, particularly in the Avatar+Voice condition, significantly strengthened physiological synchrony, presence dimensions, social connectedness and user experiences compared to the Alone condition, where no social indicators were available. Avatars had an overall greater impact than Voice in intensifying shared experiences, indicating that virtual bodies amplify co-experiences both subjectively and physiologically. We demonstrate that physiological synchrony can serve as a measure of co-experience in social VR music settings. Our findings show its potential for evaluating and designing embodied interaction in extended reality (XR), contributing to the optimization of presence dimensions, social connectedness and user experience. We inform XR researchers and practitioners about the role of avatars and voices in shaping shared VR experiences and discuss how synchrony-based measures can support future research and design efforts to understand co-experiences in social XR
Exploring Socio-Visual Dimensions of Input Visualization
No full textWe introduce the concept of socio-visual dimensions in data visualization, referring to the nuanced interplay between social and visual dimensions in people’s interactions with visualizations. From this, we discuss how input visualizations include implicit social dimensions that can be further explored and leveraged as part of a socially motivated research agenda for data visualization. Visualization researchers and designers create visual representations of data, often with the intent to inspire people to act for the greater good. Yet, research has only engaged in a limited sense with how to design with such intent and how people respond to these socially motivated visualizations. Understanding the socio-visual dimensions is crucial for addressing global challenges like climate change and public health crises, where collective action depends on fostering togetherness, a sense of community, and pro-social behaviour
Gaming Grandma and the Riverside Lodge: Space and Place at the end of a Grand Skyrim Adventure
Get PDFPrompts First, Precision Later: Reviving the Vision of Natural Language Programming for Computing Education
No full textGenerative AI (GenAI) is disrupting Computer Science Education, proving to be increasingly capable at more and more challenges. Although some educators consider this a threat to computing education, we argue that the move towards natural language was inevitable. Since the beginning, our discipline has continually increased the level of abstraction in each new representation. We have progressed from hardware dip switches, through special purpose languages and visual representations like flow charts, all the way now to "natural language"expressions of problems that serve as "prompts"for generating solutions.Now students can finally engage in problem solving via the abstraction level of the "language"with which they have been "problem solving"all their lives. And the arbitrary association of "computer programming"with "English"has finally been broken, thanks to the natural language translation capabilities of large language models. We must adjust our pedagogy to this new era. Similar to how we don't start teaching math by teaching decontextualised math notation, we shouldn't start teaching programming by requiring precise programming language syntax. New learners should start from the highest level of abstraction we have: prompts first