Mixed-Initiative Methods for Verifiable, Controllable Creation in Scientific Research

Abstract

Scientific research is inherently creative. Scientists rely on their creativity constantly, whether to identify new means of inspiration, invent novel research ideas, or find innovative framings of their work to communicate to broader audiences. However, scientists face two key challenges when entering a creative process. To start, they must overcome fixation, a common phenomenon that leads people to think about a problem in familiar terms rather than seek out new solutions. For scientists, this means exploring concepts and ideas outside of their immediate research area and scholarly circles. Furthermore, scientists must work within the context of an exponentially growing body of scientific literature, which can be unwieldy to organize and utilize to its full potential. Artificial intelligence (AI) and large-language-model (LLM) systems are increasingly being used for creative purposes and can provide suggestions to help scientists overcome fixation. In combination with information retrieval, AI systems can also provide suggestions that make creative use of the expansive scientific literature. Nonetheless, many modern AI systems are black boxes that do not present explanations to understand their actions or controls to steer them. In this dissertation, I demonstrate that scientists working on creative steps of the scientific process benefit more from AI suggestions when mixed-initiative methods are implemented to help them (1) verify the suggestions for relevance to their creative vision and (2) steer the suggestions to better align with that vision, all while staying rooted in the literature. I begin by showing the complementary benefits of local and global content-based explanations for understanding and adjusting an AI research-paper recommendation feed. Next, I present Papers-to-Posts, a human-LLM tool for translating research papers into blog posts through a novel mechanism--interactive reverse source outlines, which allow users to recognize and change what source content was and was not selected by the LLM for inclusion in the summary article. I then introduce Scideator, a system that implements a human-LLM workflow for iterative research idea generation through transparent, controllable recombination of research-paper facets. To conclude, I reflect on opportunities for future work with respect to investigating the generalizability of the presented mixed-initiative methods as well as exploring the spectrum of scientific literature components used in these methods.