Vision-language models (VLMs) are widely used in multimodal systems, but they can pick up spurious and harmful associations from their training data. Most existing efforts concentrate on removing biases that are already known.
The SPELUNKA framework, described in a new arXiv preprint, takes a different approach: it aims to find unanticipated associations that have not yet been identified. This could help surface hidden problems before they cause harm.
The abstract notes that prior work focuses on removing known biases, implying SPELUNKA fills a gap by discovering unknown ones. However, details on how the framework works or its results are not provided in the abstract.