UCLA

Batch Content Generation (BCoG) tasks necessitate the simultaneous generation of a substantial volume of similarly formatted content, such as batches of scientific articles, storybooks, or font sets. Currently, this process is typically accomplished through user-initiated multi-session interactions with generative models. This approach requires users to re-prompt the generative AI multiple times and manually apply changes to each item, thereby increasing the cognitive load associated with fine-tuning across multiple targets. To address these limitations, we propose StubCog, a novel solution mechanism for BCoG tasks, which leverages pipeline management and a task scheduler. An evaluation study conducted with 18 users possessing varying levels of domain expertise demonstrated that (i) in comparison to multi-session operations with dialog-based chatbots, StubCog reduces the time required for batch content generation and fine-tuning by 50%; (ii) StubCog significantly reduces the cognitive load associated with the fine-tuning process. Our findings suggest that StubCog isa promising solution for BCoG tasks that require the generation of large volumes of content.