Search for: All records

Objectives. Young adults may use cannabis for distinct reasons, for example, related to their personalities or lifestyles. This study examined young adults’ personality profiles and associations with cannabis use-related outcomes. Methods. Using 3 waves of survey data (2023–2024) among 4,031 US young adults (Mage = 26.29, 59.8% female, 19.4% Hispanic, 14.0% Black, 14.1% Asian), latent profile analysis was conducted based on 5 personality characteristics (emotional stability, extraversion, openness, agreeableness, conscientiousness). Generalized estimating equations assessed class in relation to past-month use and use intentions and their changes over time. Models controlled for state non-medical cannabis legalization and sociodemographics. Results. Three classes emerged: 1) the All-high-trait class (highest scores on all 5 personality characteristics; 21.1%); 2) the Moderate-extraversion/low-other-trait class (37.1%); and 3) the Low-extraversion/moderate-other-trait class (41.8%). At W1, past-month cannabis use was higher in the All-high-trait (52.2%) and Moderate-extraversion/low-other-trait classes vs. the Low-extraversion/moderate-other-trait class (43.3%). At W1, compared to the All-high-trait class, Low-extraversion/moderate-other-trait reported lower odds of use (aOR = 0.75, 95%CI = 0.63–0.90), and both Moderate-extraversion/low-other-trait and Low-extraversion/moderate-other-trait reported lower intentions (Moderate-extraversion/low-other-trait: B = -0.22, SE = 0.07; Low-extraversion/moderate-other-trait: B = -0.28, SE = 0.07). While odds of use and intentions generally decreased overall, class-by-time interactions indicated that: 1) odds of use decreased less from W1 to W2 in Low-extraversion/moderate-other-trait vs. Moderate-extraversion/low-other-trait (aOR = 1.23, 95%CI = 1.07–1.42); and 2) use intentions increased more from W1 to W2 in Moderate-extraversion/low-other-trait vs. All-high-trait (B = 0.18, SE = 0.09) and decreased less from W2 to W3 in Low-extraversion/moderate-other-trait vs. Moderate-extraversion/low-other-trait (B = 0.15, SE = 0.07). Conclusions. Intervention efforts should consider distinct personality profiles of young adults and potentially distinct mechanisms (e.g., social, emotional) underlying their cannabis use. 

Accurate electron density prediction is central to estimating quantum- and macroscale material properties using density functional theory (DFT) yet achieving high accuracy at reasonable computational cost remains a longstanding challenge. We introduce a Special Euclidean group in 3D (SE(3)) -equivariant steerable convolutional neural network [1] for predicting three-dimensional bulk electron density across diverse material systems, including medium- and high-entropy alloys. By leveraging a group-theoretic framework [2], the proposed model learns higher-order equivariant features that explicitly respect rotational and translational symmetries of physical space. We show that the proposed model outperforms a descriptor-based machine learning baseline [3,4] by 33.6% in 3D bulk electron density prediction on high-entropy alloys when using the superposition of atomic densities (SAD) as input, while achieving consistently higher predictive accuracy and improved symmetry expressivity across elemental (Al), molecular (water), ternary (SiGeSn), and quaternary (CrFeCoNi) systems. In addition, the proposed model enables significantly faster inference compared to conventional DFT-based electron density calculations, making it a time-efficient alternative for large-scale or high-throughput material screening. Notably, the model exhibits strong element-interpolation capability, successfully predicting electron density for entirely unseen systems, including CrMnFeCoNi alloys with mono- and diatomic Mn configurations when trained only on CrFeCoNi data, as well as the ternary SiGeSn system when trained exclusively on all binary combinations of Si, Ge, and Sn, individually. Further analysis indicates that the proposed model captures latent symmetry structures more effectively than conventional convolutional neural networks, maintaining higher order equivariance throughout the prediction process. 

Reasoning Large Multi-modality Models (LMMs) have become the de facto choice for many applications. However, these models rely on a Chain-of-Thought (CoT) process that is lengthy and unpredictable at runtime, often resulting in inefficient use of computational resources (due to memory fragmentation) and sub-optimal accuracy (due to under and over-thinking). Empirical evidence shows a strong correlation between CoT length and task difficulty. This suggests that the CoT length can be estimated ahead of time based on a hidden parameter representing the amount of “fuel” available to support the reasoning process. Based on this insight, we propose Fuel Gauge, the first method which extracts this hidden signal and predicts CoT length ahead of time. We demonstrate the utility on the Fuel Gauge on two downstream tasks: predictive KV cache allocation, which addresses memory fragmentation in LMM serving systems, and CoT length modulation, which mitigates under-thinking and over-thinking. Extensive experiments on LMMs across text-only, image-text, and video-text question answering benchmarks demonstrate the effectiveness, generalizability, and practical value of Fuel Gauge. For example, on the GPQA-Diamond benchmark, our Fuel Gauge achieves less than half the CoT length prediction error compared to the baseline; this translates into a 13.37→ reduction in the memory allocation frequency. 

The global policy landscape has coalesced around “Trustworthy AI” as a normative goal for AI governance and development. However, despite its ubiquity, there is no widespread agreement on the definition of the term. Based on a convened discourse of interdisciplinary institute leaders at The George Washington University, this paper argues that the failure to define “Trustworthy AI” is not a technical gap, but a consequence of different epistemologies associated with different worldviews. We argue that “trust” is not a singular property, but a “boundary object” (Star and Griesemer in Soc Stud Sci 19:387–420, 1989) interpreted through four different worldviews: the Technocratic (trust as metrology), the Relational (trust as agency and social contract), the Critical (trust vs. power), and the Pragmatic (trust as pedestrian verification). This paper delineates these paradigms, explores the friction between them, and argues that a rigorous framework for trustworthy AI requires a pluralistic approach that acknowledges these inherent contradictions. 

Background. Despite increased popularity of cannabis edibles, limited research has assessed effects of warnings or advertising emphasizing different characteristics. This study examined effects of different edible advertising messages and warnings on young adults’ product perceptions. Methods. In June-November 2024, US adults ages 18–34 (n = 3579) participated in an online survey-based experiment using a 2 × 4 factorial design (advertising message: psychological effects vs. flavors; warning: none, adult-use, intoxication, edible-specific), then reported perceived addictiveness, harm, cautiousness, appeal, and interest (0 =not at all to 7 =extremely). Multivariable regression assessed advertising message, warning, and message-by-warning interactions in relation to outcomes, overall and by past-month use status. Results. Few experimental effects were identified. The overall model showed: one direct, negative effect for flavors (vs. psychological effects) advertising messaging on interest (B=-0.20, SE=0.07, p = .003), shown in stratified analyses to be attributed to effects among those reporting past-month use (B=-0.28, SE=0.11, p = .003; no effect among participants reporting no past-month use-); and one direct, positive effect of the edible-specific (vs. no) warning on perceived harm (B=0.30, SE=0.10, p = .002), driven by those reporting non-use (B=0.32, SE=0.13; no effect among those reporting use). A significant interaction (B=0.38, SE=0.19, p = .048) indicated that adult-use warning exposure reduced perceived addictiveness among participants reporting non-use but had no effect for those reporting use. Among participants reporting use, an interaction (B=0.66, SE= 0.29, p = .019) indicated flavor (vs. effect) messaging alongside edible-specific warning exposure resulted in higher addictiveness. Conclusions. Regulations and interventions must consider effects of advertising messages and advertising message/warning combinations, particularly among different subgroups (e.g., those who use). 

Code agents have gained widespread adoption due to their strong code generation capabilities and integration with code interpreters, enabling dynamic execution, debugging, and interactive programming capabilities. While these advancements have streamlined complex workflows, they have also introduced critical safety and security risks. Current static safety benchmarks and red-teaming tools are inadequate for identifying emerging real-world risky scenarios, as they fail to cover certain boundary conditions, such as the combined effects of different jailbreak tools. In this work, we propose RedCodeAgent, the first automated red-teaming agent designed to systematically uncover vulnerabilities in diverse code agents. With an adaptive memory module, RedCodeAgent can leverage existing jailbreak knowledge, dynamically select the most effective red-teaming tools and tool combinations in a tailored toolbox for a given input query, thus identifying vulnerabilities that might otherwise be overlooked. For reliable evaluation, we develop simulated sandbox environments to additionally evaluate the execution results of code agents, mitigating potential biases of LLM-based judges that only rely on static code. Through extensive evaluations across multiple state-of-the-art code agents, diverse risky scenarios, and various programming languages, RedCodeAgent consistently outperforms existing red-teaming methods, achieving higher attack success rates and lower rejection rates with high efficiency. We further validate RedCodeAgent on real-world code assistants, e.g., Cursor and Codeium, exposing previously unidentified security risks. By automating and optimizing red-teaming processes, RedCodeAgent enables scalable, adaptive, and effective safety assessments of code agents. 

Vision Graph Neural Networks (ViGs) are often limited by their reliance on a fixed, homogeneous graph construction rule applied across all network stages. To address this limitation, we introduce SearchViG, a novel framework that automatically designs optimal heterogeneous architectures by performing a Graph Construction Search (GCS) to produce the optimal graph topology for each stage within our designated search space. Our search is guided by a zero-shot, theoretically-grounded proxy: the spectral gap of the graph’s adjacency matrix, which quantifies its Ramanujan-like expansion properties, provably linking it to superior information flow. SearchViG discovers new heterogeneous architectures that assign different graph topologies, number of neighbors, and hops between neighbors based on feature resolution. Our resulting models establish a new state-of-the-art Pareto frontier for Vision GNNs. For instance, our SearchViG-M achieves 83.3% top-1 accuracy, outperforming both Vision GNN-B (ViG-B) and Vision Hypergraph Neural Network-B (ViHGNN-B) while using over 70% fewer parameters and 80% fewer GMACs. This efficiency extends to downstream tasks, where our lightweight SearchViG-S obtains 43.4 mIoU, 43.5 APbox, and 39.9 APmask, surpassing the much larger Pyramid Vision Transformer-Large (PVT-Large) across all metrics while using 80% fewer parameters.