ScholaWrite2.0: A Dataset of End-to-End Scholarly Process
Current AI research assistants can observe what a researcher is doing in the moment, but lack understanding of the intention and context behind each action, making their suggestions shallow and often unhelpful. To build cognitively-aligned AI research assistants, we need to capture the full research workflow and understand the intention behind each activity and how each activity influences the ones that follow. We argue that capturing the full research workflow, along with the intentions behind each activity, can yield principles and insights that inform the design and training of more cognitively-aligned AI research assistants.
We conducted a pilot study in which three researchers, with different experience levels, each recorded two sessions of 1-2 hours from the same ongoing research project, supplemented by post-study reflections on their goals, challenges, and attitudes toward AI tools. Beyond the dataset, we also contribute a hierarchical annotation taxonomy with three layers — Action, Strategy, and Goal — capturing what the researcher is doing, why they are doing it, and how it fits into the broader research process.
We will annotate and analyze these sessions to examine where human expectations and AI responses diverge, and how context and memory accumulate across a research session. These findings will inform the design of AI research assistants that are better aligned with how researchers actually think and work.