Markus Strohmaier’s research while affiliated with University of Mannheim and other places

Stereotypical bias encoded in language models (LMs) poses a threat to safe language technology, yet our understanding of how bias manifests in the parameters of LMs remains incomplete. We introduce local contrastive editing that enables the localization and editing of a subset of weights in a target model in relation to a reference model. We deploy this approach to identify and modify subsets of weights that are associated with gender stereotypes in LMs. Through a series of experiments, we demonstrate that local contrastive editing can precisely localize and control a small subset (< 0.5%) of weights that encode gender bias. Our work (i) advances our understanding of how stereotypical biases can manifest in the parameter space of LMs and (ii) opens up new avenues for developing parameter-efficient strategies for controlling model properties in a contrastive manner.

This paper proposes temporally aligned Large Language Models (LLMs) as a tool for longitudinal analysis of social media data. We fine-tune Temporal Adapters for Llama 3 8B on full timelines from a panel of British Twitter users, and extract longitudinal aggregates of emotions and attitudes with established questionnaires. We validate our estimates against representative British survey data and find strong positive, significant correlations for several collective emotions. The obtained estimates are robust across multiple training seeds and prompt formulations, and in line with collective emotions extracted using a traditional classification model trained on labeled data. To the best of our knowledge, this is the first work to extend the analysis of affect in LLMs to a longitudinal setting through Temporal Adapters. Our work enables new approaches towards the longitudinal analysis of social media data.

In recent years, several measures have been developed for evaluating group fairness of rankings. Given that these measures were developed with different application contexts and ranking algorithms in mind, it is not straightforward which measure to choose for a given scenario. Previous work has already laid out some categorizations of measures and explored relationships between these, however, there has not yet been a thorough mathematical analysis of practically grounded properties of individual measures. In this paper, we therefore apply an axiomatic approach to perform a comprehensive analysis of existing group fairness measures that have been developed in the context of fair ranking. To this end, we propose a set of fourteen properties for group fairness measures that consider different ranking settings. These properties specifically provide information about how fairness scores of ranked outputs can be interpreted and contextualized. For a given use case, one can then identify which properties are of interest, and select a measure based on whether it satisfies these properties. We further apply our properties on twelve existing group fairness measures, and through both empirical and theoretical results demonstrate that most of these measures only satisfy a small subset of the proposed properties. These findings highlight limitations of existing measures, and provide insights into how to evaluate and interpret different fairness measures in practical deployment. Overall, our work can assist practitioners in selecting appropriate group fairness measures for a specific application, and also aid researchers in designing and evaluating such measures.

Measuring the similarity of different representations of neural architectures is a fundamental task and an open research challenge for the machine learning community. This paper presents the first comprehensive benchmark for evaluating representational similarity measures based on well-defined groundings of similarity. The representational similarity (ReSi) benchmark consists of (i) six carefully designed tests for similarity measures, (ii) 23 similarity measures, (iii) eleven neural network architectures, and (iv) six datasets, spanning over the graph, language, and vision domains. The benchmark opens up several important avenues of research on representational similarity that enable novel explorations and applications of neural architectures. We demonstrate the utility of the ReSi benchmark by conducting experiments on various neural network architectures, real world datasets and similarity measures. All components of the benchmark are publicly available and thereby facilitate systematic reproduction and production of research results. The benchmark is extensible, future research can build on and further expand it. We believe that the ReSi benchmark can serve as a sound platform catalyzing future research that aims to systematically evaluate existing and explore novel ways of comparing representations of neural architectures.

We illustrate how standard psychometric inventories originally designed for assessing noncognitive human traits can be repurposed as diagnostic tools to evaluate analogous traits in large language models (LLMs). We start from the assumption that LLMs, inadvertently yet inevitably, acquire psychological traits (metaphorically speaking) from the vast text corpora on which they are trained. Such corpora contain sediments of the personalities, values, beliefs, and biases of the countless human authors of these texts, which LLMs learn through a complex training process. The traits that LLMs acquire in such a way can potentially influence their behavior, that is, their outputs in downstream tasks and applications in which they are employed, which in turn may have real-world consequences for individuals and social groups. By eliciting LLMs’ responses to language-based psychometric inventories, we can bring their traits to light. Psychometric profiling enables researchers to study and compare LLMs in terms of noncognitive characteristics, thereby providing a window into the personalities, values, beliefs, and biases these models exhibit (or mimic). We discuss the history of similar ideas and outline possible psychometric approaches for LLMs. We demonstrate one promising approach, zero-shot classification, for several LLMs and psychometric inventories. We conclude by highlighting open challenges and future avenues of research for AI Psychometrics.

Wikipedia is one of the most successful collaborative projects in history. It is the largest encyclopedia ever created, with millions of users worldwide relying on it as the first source of information as well as for fact-checking and in-depth research. As Wikipedia relies solely on the efforts of its volunteer editors, its success might be particularly affected by toxic speech. In this paper, we analyze all 57 million comments made on user talk pages of 8.5 million editors across the six most active language editions of Wikipedia to study the potential impact of toxicity on editors’ behavior. We find that toxic comments are consistently associated with reduced activity of editors, equivalent to 0.5–2 active days per user in the short term. This translates to multiple human-years of lost productivity, considering the number of active contributors to Wikipedia. The effects of toxic comments are potentially even greater in the long term, as they are associated with a significantly increased risk of editors leaving the project altogether. Using an agent-based model, we demonstrate that toxicity attacks on Wikipedia have the potential to impede the progress of the entire project. Our results underscore the importance of mitigating toxic speech on collaborative platforms such as Wikipedia to ensure their continued success.