Large language models (LLMs) are the foundation of the current successes of artificial intelligence (AI), however, they are unavoidably biased. To effectively communicate the risks and encourage mitigation efforts these models need adequate and intuitive descriptions of their discriminatory properties, appropriate for all audiences of AI. We suggest bias profiles with respect to stereotype dimensions based on dictionaries from social psychology research. Along these dimensions we investigate gender bias in contextual embeddings, across contexts and layers, and generate stereotype profiles for twelve different LLMs, demonstrating their intuition and use case for exposing and visualizing bias.

LLMs can perform unseen tasks by learning from a few in-context examples. How in-context learning works is still uncertain. We investigate the mechanisms of in-context learning on a challenging non-language task. The task requires the LLM to generate pixel matrices representing images of basic shapes. We introduce a framework to analyze if this task is solved by recognizing similar formats from the training data (task recognition) or by understanding the instructions and learning the skill de novo during inference (task learning). Our experiments demonstrate that LLMs generate meaningful pixel matrices with task recognition and fail to learn such tasks when encountering unfamiliar formats. Our findings offer insights into LLMs’ learning mechanisms and their generalization ability to guide future research on their seemingly human-like behavior.

Though successfully applied in research and industry large pretrained language models of the BERT family are not yet fully understood. While much research in the field of BERTology has tested whether specific knowledge can be extracted from layer activations, we invert the popular probing design to analyze the prevailing differences and clusters in BERT’s high dimensional space. By extracting coarse features from masked token representations and predicting them by probing models with access to only partial information we can apprehend the variation from ‘BERT’s point of view’. By applying our new methodology to different datasets we show how much the differences can be described by syntax but further how they are to a great extent shaped by the most simple positional information.