While recent work has found that vision-language models trained under the Contrastive Language Image Pre-training (CLIP) framework contain intrinsic social biases, the extent to which different upstream pre-training features of the framework relate to these biases, and hence how intrinsic bias and downstream performance are connected has been unclear. In this work, we present the largest comprehensive analysis to-date of how the upstream pre-training factors and downstream performance of CLIP models relate to their intrinsic biases. Studying 131 unique CLIP models, trained on 26 datasets, using 55 architectures, and in a variety of sizes, we evaluate bias in each model using 26 well-established unimodal and cross-modal principled Embedding Association Tests. We find that the choice of pre-training dataset is the most significant upstream predictor of bias, whereas architectural variations have minimal impact. Additionally, datasets curated using sophisticated filtering techniques aimed at enhancing downstream model performance tend to be associated with higher levels of intrinsic bias. Finally, we observe that intrinsic bias is often significantly correlated with downstream performance (0.3 ≤ r ≤ 0.8), suggesting that models optimized for performance inadvertently learn to amplify representational biases. Comparisons between unimodal and cross-modal association tests reveal that social group bias depends heavily on the modality. Our findings imply that more sophisticated strategies are needed to address intrinsic model bias for vision-language models across the entire model development pipeline.

Garden path sentences (i.e. “the horse raced past the barn fell”) are sentences that readers initially incorrectly parse, requiring partial or total re-analysis of the sentence structure. Given human difficulty in parsing garden paths, we aim to compare transformer language models’ performance on these sentences. We assess a selection of models from the BERT family which have been fine-tuned on the question-answering task, and evaluate each model’s performance on comprehension questions based on garden path and control sentences. We then further investigate the semantic roles assigned to arguments of verbs in garden path and control sentences by utilizing a probe task to directly assess which semantic role(s) the model assigns. We find that the models have relatively low performance in certain instances of question answering based on garden path contexts, and the model incorrectly assigns semantic roles, aligning for the most part with human performance.