We introduce aligned probing, a novel interpretability framework that aligns the behavior of language models (LMs), based on their outputs, and their internal representations (internals). Using this framework, we examine over 20 OLMo, Llama, and Mistral models, bridging behavioral and internal perspectives for toxicity for the first time. Our results show that LMs strongly encode information about the toxicity level of inputs and subsequent outputs, particularly in lower layers. Focusing on how unique LMs differ offers both correlative and causal evidence that they generate less toxic output when strongly encoding information about the input toxicity. We also highlight the heterogeneity of toxicity, as model behavior and internals vary across unique attributes such as Threat. Finally, four case studies analyzing detoxification, multi-prompt evaluations, model quantization, and pre-training dynamics underline the practical impact of aligned probing with further concrete insights. Our findings contribute to a more holistic understanding of LMs, both within and beyond the context of toxicity.

Nevertheless, there is a significant lack of resources to assess the impact of this linguistic shift on classification using language models (LMs), which are probably not trained on such variations. To address this gap, we present Lou, the first dataset featuring high-quality reformulations for German text classification covering seven tasks, like stance detection and toxicity classification. Evaluating 16 mono-and multi-lingual LMs on Lou shows that genderfair language substantially impacts predictions by flipping labels, reducing certainty, and altering attention patterns. However, existing evaluations remain valid, as LM rankings of Figure 1: A German stance detection instance from the original and reformulated instances do not significantly Lou dataset. We reformulate the masculine formulation differ. While we offer initial insights Konsumenten (consumers) regarding six inclusive or on the effect on German text classification, the neutral strategies, highlighted in yellow. Translation: findings likely apply to other languages, as consistent Consumers must be well supported.

We introduce Holmes, a benchmark to assess the linguistic competence of language models (LMs) - their ability to grasp linguistic phenomena. Unlike prior prompting-based evaluations, Holmes assesses the linguistic competence of LMs via their internal representations using classifier-based probing. In doing so, we disentangle specific phenomena (e.g., part-of-speech of words) from other cognitive abilities, like following textual instructions, and meet recent calls to assess LMs' linguistic competence in isolation. Composing Holmes, we review over 250 probing studies and feature more than 200 datasets to assess syntax, morphology, semantics, reasoning, and discourse phenomena. Analyzing over 50 LMs reveals that, aligned with known trends, their linguistic competence correlates with model size. However, surprisingly, model architecture and instruction tuning also significantly influence performance, particularly in morphology and syntax. Finally, we propose FlashHolmes, a streamlined version of Holmes designed to lower the high computation load while maintaining high-ranking precision.

The long-standing problem of spam and fraudulent messages in the comment sections of Instagram pages in the financial sector claims new victims every day. Instagram's current spam filter proves inadequate, and existing research approaches are primarily confined to theoretical concepts. Practical implementations with evaluated results are missing. To solve this problem, we propose ScamSpot, a comprehensive system that includes a browser extension, a fine-tuned BERT model and a REST API. This approach ensures public accessibility of our results for Instagram users using the Chrome browser. Furthermore, we conduct a data annotation study, shedding light on the reasons and causes of the problem and evaluate the system through user feedback and comparison with existing models. ScamSpot is an open-source project and is publicly available at https://scamspot.github.io/.

Pre-trained language models (LMs) perform well in In-Topic setups, where training and testing data come from the same topics. However, they face challenges in Cross-Topic scenarios where testing data is derived from distinct topics -- such as Gun Control. This study analyzes various LMs with three probing-based experiments to shed light on the reasons behind the In- vs. Cross-Topic generalization gap. Thereby, we demonstrate, for the first time, that generalization gaps and the robustness of the embedding space vary significantly across LMs. Additionally, we assess larger LMs and underscore the relevance of our analysis for recent models. Overall, diverse pre-training objectives, architectural regularization, or data deduplication contribute to more robust LMs and diminish generalization gaps. Our research contributes to a deeper understanding and comparison of language models across different generalization scenarios.

Language models (LMs) excel in in-distribution (ID) scenarios where train and test data are independent and identically distributed. However, their performance often degrades in real-world applications like argument mining. Such degradation happens when new topics emerge, or other text domains and languages become relevant. To assess LMs' generalization abilities in such out-of-distribution (OOD) scenarios, we simulate such distribution shifts by deliberately withholding specific instances for testing, as from the social media domain or the topic Solar Energy. Unlike prior studies focusing on specific shifts and metrics in isolation, we comprehensively analyze OOD generalization. We define three metrics to pinpoint generalization flaws and propose eleven classification tasks covering topic, domain, and language shifts. Overall, we find superior performance of prompt-based fine-tuning, notably when train and test splits primarily differ semantically. Simultaneously, in-context learning is more effective than prompt-based or vanilla fine-tuning for tasks when training data embodies heavy discrepancies in label distribution compared to testing data. This reveals a crucial drawback of gradient-based learning: it biases LMs regarding such structural obstacles.

Stance detection deals with identifying an author's stance towards a target. Most existing stance detection models are limited because they do not consider relevant contextual information which allows for inferring the stance correctly. Complementary context can be found in knowledge bases but integrating the context into pretrained language models is non-trivial due to the graph structure of standard knowledge bases. To overcome this, we explore an approach to integrate contextual information as text which allows for integrating contextual information from heterogeneous sources, such as structured knowledge sources and by prompting large language models. Our approach can outperform competitive baselines on a large and diverse stance detection benchmark in a cross-target setup, i.e. for targets unseen during training. We demonstrate that it is more robust to noisy context and can regularize for unwanted correlations between labels and target-specific vocabulary. Finally, it is independent of the pretrained language model in use.