Researchers at the Allen Institute for AI and the University of Pennsylvania have developed LitPivot, a system designed to guide research ideation by organizing academic literature around specific parts of a proposal draft. Hugging Face published details of the project on October 23, 2026, ahead of its scheduled presentation at the UIST 2026 conference in Detroit.
LitPivot works by segmenting an idea into three facets: the problem, the proposed contribution, and the evaluation plan. Using papers retrieved through Asta Paper Finder, the system groups relevant literature around the facet a researcher chooses to edit. For example, selecting the evaluation facet prompts LitPivot to cluster papers by their evaluation methods. The system then generates critiques supported by direct excerpts from selected papers and suggests revisions that authors can accept, modify, or reject.
When changes to one facet leave another outdated, LitPivot flags the discrepancy. An altered contribution, for instance, triggers an alert if the existing evaluation plan no longer tests the new direction adequately. The tool then reranks surrounding literature based on the revised text.
System design followed an analysis of meeting notes from four completed computer science projects, which uncovered 21 literature-initiated pivots. These pivots divided into attempts to differentiate an idea from existing studies and efforts to align with them. A formative study of five researchers using text editors and standard search tools further demonstrated that authors struggled to critically evaluate drafts without concrete, literature-backed suggestions.
In a controlled trial, 17 researchers each developed one assigned proposal using LitPivot and another using a baseline text editor paired with an AI question-answering tool. Two blinded domain experts evaluated the drafts and rated proposals developed with LitPivot higher for how well they grounded and supported ideas with relevant literature. A follow-up study examined five researchers using LitPivot on active projects during 30-minute sessions, where two tested early concepts and three adjusted mature framings.
LitPivot joins a broader collection of scientific tooling developed at Ai2, which includes the Theorizer and AutoDiscovery components in Asta, the hypothesis generator HARPA, the recombination tool Scideator, and the literature-grounded feedback tool IdeaSynth.
