Weird generalization is a phenomenon in which models fine-tuned on data from a narrow domain (e.g. insecure code) develop surprising traits that manifest even outside that domain (e.g. broad misalignment)-a phenomenon that prior work has highlighted as a critical safety concern. Here, we present an extended replication study of key weird generalization results across an expanded suite of models and datasets. We confirm that surprising (and dangerous) traits can emerge under certain circumstances, but we find that weird generalization is exceptionally brittle: it emerges only for specific models on specific datasets, and it vanishes under simple training-time, prompt-based interventions. We find that the most effective interventions provide prompt context that makes the generalized behavior the expected behavior. However, we show that even very generic interventions that do not anticipate specific generalized traits can still be effective in mitigating weird generalization's effects. Our findings thus help clarify the nature of the safety threat that weird generalization poses and point toward an easily implemented set of solutions.
Large language models (LLMs) are often trained on data that reflect WEIRD values: Western, Educated, Industrialized, Rich, and Democratic. This raises concerns about cultural bias and fairness. Using responses to the World Values Survey, we evaluated five widely used LLMs: GPT-3.5, GPT-4, Llama-3, BLOOM, and Qwen. We measured how closely these responses aligned with the values of the WEIRD countries and whether they conflicted with human rights principles. To reflect global diversity, we compared the results with the Universal Declaration of Human Rights and three regional charters from Asia, the Middle East, and Africa. Models with lower alignment to WEIRD values, such as BLOOM and Qwen, produced more culturally varied responses but were 2% to 4% more likely to generate outputs that violated human rights, especially regarding gender and equality. For example, some models agreed with the statements ``a man who cannot father children is not a real man'' and ``a husband should always know where his wife is'', reflecting harmful gender norms. These findings suggest that as cultural representation in LLMs increases, so does the risk of reproducing discriminatory beliefs. Approaches suc
We recently described a specific type of attractors of two-dimensional discontinuous piecewise linear maps, characterized by two discontinuity lines dividing the phase plane into three partitions, related to economic applications. To our knowledge, this type of attractor, which we call a weird quasiperiodic attractor, has not yet been studied in detail. They have a rather complex geometric structure and other interesting properties that are worth understanding better. To this end, we consider a simpler map that can also possess weird quasiperiodic attractors, namely, a 2D discontinuous piecewise linear map $F$ with a single discontinuity line dividing the phase plane into two partitions, where two different homogeneous linear maps are defined. Map $F$ depends on four parameters -- the traces and determinants of the two Jacobian matrices. In the parameter space of map $F$, we obtain specific regions associated with the existence of weird quasiperiodic attractors; describe some characteristic properties of these attractors; and explain one of the possible mechanisms of their appearance.
Despite its importance, studying economic behavior across diverse, non-WEIRD (Western, Educated, Industrialized, Rich, and Democratic) populations presents significant challenges. We address this issue by introducing a novel methodology that uses Large Language Models (LLMs) to create synthetic cultural agents (SCAs) representing these populations. We subject these SCAs to classic behavioral experiments, including the dictator and ultimatum games. Our results demonstrate substantial cross-cultural variability in experimental behavior. Notably, for populations with available data, SCAs' behaviors qualitatively resemble those of real human subjects. For unstudied populations, our method can generate novel, testable hypotheses about economic behavior. By integrating AI into experimental economics, this approach offers an effective and ethical method to pilot experiments and refine protocols for hard-to-reach populations. Our study provides a new tool for cross-cultural economic studies and demonstrates how LLMs can help experimental behavioral research.
Measuring how real images look is a complex task in artificial intelligence research. For example, an image of a boy with a vacuum cleaner in a desert violates common sense. We introduce a novel method, which we call Through the Looking Glass (TLG), to assess image common sense consistency using Large Vision-Language Models (LVLMs) and Transformer-based encoder. By leveraging LVLMs to extract atomic facts from these images, we obtain a mix of accurate facts. We proceed by fine-tuning a compact attention-pooling classifier over encoded atomic facts. Our TLG has achieved a new state-of-the-art performance on the WHOOPS! and WEIRD datasets while leveraging a compact fine-tuning component.
Much of the research in social computing analyzes data from social media platforms, which may inherently carry biases. An overlooked source of such bias is the over-representation of WEIRD (Western, Educated, Industrialized, Rich, and Democratic) populations, which might not accurately mirror the global demographic diversity. We evaluated the dependence on WEIRD populations in research presented at the AAAI ICWSM conference; the only venue whose proceedings are fully dedicated to social computing research. We did so by analyzing 494 papers published from 2018 to 2022, which included full research papers, dataset papers and posters. After filtering out papers that analyze synthetic datasets or those lacking clear country of origin, we were left with 420 papers from which 188 participants in a crowdsourcing study with full manual validation extracted data for the WEIRD scores computation. This data was then used to adapt existing WEIRD metrics to be applicable for social media data. We found that 37% of these papers focused solely on data from Western countries. This percentage is significantly less than the percentages observed in research from CHI (76%) and FAccT (84%) conferences,
In this work, we consider a class of $n$-dimensional, $n\geq2$, piecewise linear discontinuous maps that can exhibit a new type of attractor, called a weird quasiperiodic attractor. While the dynamics associated with these attractors may appear chaotic, we prove that chaos cannot occur. The considered class of $n$-dimensional maps allows for any finite number of partitions, separated by various types of discontinuity sets. The key characteristic, beyond discontinuity, is that all functions defining the map have the same real fixed point. These maps cannot have hyperbolic cycles other than the fixed point itself. We consider the two-dimensional case in detail. We prove that in nongeneric cases, the restriction, or the first return, of the map to a segment of straight line is reducible to a piecewise linear circle map. The generic attractor, different from the fixed point, is a weird quasiperiodic attractor, which may coexist with other attractors or attracting sets. We illustrate the existence of these attractors through numerous examples, using functions with different types of Jacobian matrices, as well as with different types of discontinuity sets. In some cases, we describe poss
Data annotation remains the sine qua non of machine learning and AI. Recent empirical work on data annotation has begun to highlight the importance of rater diversity for fairness, model performance, and new lines of research have begun to examine the working conditions for data annotation workers, the impacts and role of annotator subjectivity on labels, and the potential psychological harms from aspects of annotation work. This paper outlines a critical genealogy of data annotation; starting with its psychological and perceptual aspects. We draw on similarities with critiques of the rise of computerized lab-based psychological experiments in the 1970's which question whether these experiments permit the generalization of results beyond the laboratory settings within which these results are typically obtained. Do data annotations permit the generalization of results beyond the settings, or locations, in which they were obtained? Psychology is overly reliant on participants from Western, Educated, Industrialized, Rich, and Democratic societies (WEIRD). Many of the people who work as data annotation platform workers, however, are not from WEIRD countries; most data annotation worker
In this work, we explore Landin's Knot, which is understood as a pattern for encoding general recursion, including non-termination, that is possible after adding higher-order references to an otherwise terminating language. We observe that this isn't always true -- higher-order references, by themselves, don't lead to non-termination. The key insight is that Landin's Knot relies not primarily on references storing functions, but on unrestricted quantification over a function's environment. We show this through a closure converted language, in which the function's environment is made explicit and hides the type of the environment through impredicative quantification. Once references are added, this impredicative quantification can be exploited to encode recursion. We conjecture that by restricting the quantification over the environment, higher-order references can be safely added to terminating languages, without resorting to more complex type systems such as linearity, and without restricting references from storing functions.
The problem of the existence of non-pseudo-$\aleph_1$-compact $\mathbb R$-factorizable groups is studied. It is proved that any such group is submetrizable and has weight larger than $ω_1$. Closely related results concerning the $\mathbb R$-factorizability of products of topological groups and spaces are also obtained (a product $X\times Y$ of topological spaces is said to be $\mathbb R$-factorizable if any continuous function $X\times Y\to \mathbb R$ factors through a product of maps from $X$ and $Y$ to second-countable spaces). In particular, it is proved that the square $G\times G$ of a topological groups $G$ is $\mathbb R$-factorizable as a group if and only if it is $\mathbb R$-factorizable as a product of spaces, in which case $G$ is pseudo-$\aleph_1$-compact. It is also proved that if the product of a space $X$ and an uncountable discrete space is $\mathbb R$-factorizable, then $X^ω$ is heredirarily separable and heredirarily Lindelöf.
The Doer Effect states that completing more active learning activities, like practice questions, is more strongly related to positive learning outcomes than passive learning activities, like reading, watching, or listening to course materials. Although broad, most evidence has emerged from practice with tutoring systems in Western, Industrialized, Rich, Educated, and Democratic (WEIRD) populations in North America and Europe. Does the Doer Effect generalize beyond WEIRD populations, where learners may practice in remote locales through different technologies? Through learning analytics, we provide evidence from N = 234 Ugandan students answering multiple-choice questions via phones and listening to lectures via community radio. Our findings support the hypothesis that active learning is more associated with learning outcomes than passive learning. We find this relationship is weaker for learners with higher prior educational attainment. Our findings motivate further study of the Doer Effect in diverse populations. We offer considerations for future research in designing and evaluating contextually relevant active and passive learning opportunities including leveraging familiar tech
This paper presents a vision for creating AI systems that are inclusive at every stage of development, from data collection to model design and evaluation. We address key limitations in the current AI pipeline and its WEIRD representation, such as lack of data diversity, biases in model performance, and narrow evaluation metrics. We also focus on the need for diverse representation among the developers of these systems, as well as incentives that are not skewed toward certain groups. We highlight opportunities to develop AI systems that are for everyone (with diverse stakeholders in mind), with everyone (inclusive of diverse data and annotators), and by everyone (designed and developed by a globally diverse workforce).
In human factor fields such as human-computer interaction (HCI) and psychology, researchers have been concerned that participants mostly come from WEIRD (Western, Educated, Industrialized, Rich, and Democratic) countries. This WEIRD skew may hinder understanding of diverse populations and their cultural differences. The usable privacy and security (UPS) field has inherited many research methodologies from research on human factor fields. We conducted a literature review to understand the extent to which participant samples in UPS papers were from WEIRD countries and the characteristics of the methodologies and research topics in each user study recruiting Western or non-Western participants. We found that the skew toward WEIRD countries in UPS is greater than that in HCI. Geographic and linguistic barriers in the study methods and recruitment methods may cause researchers to conduct user studies locally. In addition, many papers did not report participant demographics, which could hinder the replication of the reported studies, leading to low reproducibility. To improve geographic diversity, we provide the suggestions including facilitate replication studies, address geographic and
Critical voices within and beyond the scientific community have pointed to a grave matter of concern regarding who is included in research and who is not. Subsequent investigations have revealed an extensive form of sampling bias across a broad range of disciplines that conduct human subjects research called "WEIRD": Western, Educated, Industrial, Rich, and Democratic. Recent work has indicated that this pattern exists within human-computer interaction (HCI) research, as well. How then does human-robot interaction (HRI) fare? And could there be other patterns of sampling bias at play, perhaps those especially relevant to this field of study? We conducted a systematic review of the premier ACM/IEEE International Conference on Human-Robot Interaction (2006-2022) to discover whether and how WEIRD HRI research is. Importantly, we expanded our purview to other factors of representation highlighted by critical work on inclusion and intersectionality as potentially underreported, overlooked, and even marginalized factors of human diversity. Findings from 827 studies across 749 papers confirm that participants in HRI research also tend to be drawn from WEIRD populations. Moreover, we find
Studies conducted on Western, Educated, Industrialized, Rich, and Democratic (WEIRD) samples are considered atypical of the world's population and may not accurately represent human behavior. In this study, we aim to quantify the extent to which the ACM FAccT conference, the leading venue in exploring Artificial Intelligence (AI) systems' fairness, accountability, and transparency, relies on WEIRD samples. We collected and analyzed 128 papers published between 2018 and 2022, accounting for 30.8% of the overall proceedings published at FAccT in those years (excluding abstracts, tutorials, and papers without human-subject studies or clear country attribution for the participants). We found that 84% of the analyzed papers were exclusively based on participants from Western countries, particularly exclusively from the U.S. (63%). Only researchers who undertook the effort to collect data about local participants through interviews or surveys added diversity to an otherwise U.S.-centric view of science. Therefore, we suggest that researchers collect data from under-represented populations to obtain an inclusive worldview. To achieve this goal, scientific communities should champion data
In this paper we study some structure properties of primitive weird numbers in terms of their factorization. We give sufficient conditions to ensure that a positive integer is weird. Two algorithms for generating weird numbers having a given number of distinct prime factors are presented. These algorithms yield primitive weird numbers of the form $mp_1\dots p_k$ for a suitable deficient positive integer $m$ and primes $p_1,\dots,p_k$ and generalize a recent technique developed for generating primitive weird numbers of the form $2^np_1p_2$. The same techniques can be used to search for odd weird numbers, whose existence is still an open question.
In this brief essay, I reflect on how Mark Fisher's definitions of the weird and the eerie could be applied in communicative data visualization. I ask how visualization designers might elicit these two impressions when a viewer is engaging with multimodal representations of data. I argue that there are situations in which viewers should feel uncertain or suspicious of unseen forces that account for the presence or absence of audiovisual patterns. Finally, I conclude that the ability to appreciate the weird and the eerie in data is particularly important at this moment in history, one marked by significant ecological and economic disruption.
Scoring function (SF) measures the plausibility of triplets in knowledge graphs. Different scoring functions can lead to huge differences in link prediction performances on different knowledge graphs. In this report, we describe a weird scoring function found by random search on the open graph benchmark (OGB). This scoring function, called AutoWeird, only uses tail entity and relation in a triplet to compute its plausibility score. Experimental results show that AutoWeird achieves top-1 performance on ogbl-wikikg2 data set, but has much worse performance than other methods on ogbl-biokg data set. By analyzing the tail entity distribution and evaluation protocol of these two data sets, we attribute the unexpected success of AutoWeird on ogbl-wikikg2 to inappropriate evaluation and concentrated tail entity distribution. Such results may motivate further research on how to accurately evaluate the performance of different link prediction methods for knowledge graphs.
Weird machines---the computational models accessible by exploiting security vulnerabilities---arise from the difference between the model a programmer has in her head of how her program should run and the implementation that actually executes. Previous attempts to reason about or identify weird machines have viewed these models through the lens of formal computational structures such as state machines and Turing machines. But because programmers rarely think about programs in this way, it is difficult to effectively apply insights about weird machines to improve security. We present a new view of weird machines based on techniques from programming languages theory and secure compilation. Instead of an underspecified model drawn from a programmers' head, we start with a program written in a high-level source language that enforces security properties by design. Instead of state machines to describe computation, we use the well-defined semantics of this source language and a target language, into which the source program will be compiled. Weird machines are the sets of behaviors that can be achieved by a compiled source program in the target language that cannot be achieved in the so
Weird numbers are abundant numbers that are not pseudoperfect. Since their introduction, the existence of odd weird numbers has been an open problem. In this work, we describe our computational effort to search for odd weird numbers, which shows their non-existence up to $10^{21}$. We also searched up to $10^{28}$ for numbers with an abundance below $10^{14}$, to no avail. Our approach to speed up the search can be viewed as an application of reverse search in the domain of combinatorial optimization, and may be useful for other similar quest for natural numbers with special properties that depend crucially on their factorization.