GroundWork turns argumentation from chat text into editable shared human-LLM state
The GroundWork preprint, submitted on November 6, operationalizes the Toulmin argumentation model as a typed graph and closed operation vocabulary. Coordinated corpus views and a workbench expose claims, evidence, assumptions, qualifiers, and rebuttals so people and models can retrieve, propose, compare, test, and revise arguments over the same state. Domain experts decide which model proposals enter the agreed argument. The formative evaluation covered three scenarios in intelligence, law, and medicine and feedback from four experts in consulting, product, law, and immunology.
Why it matters
The work points to a more consequential direction for generative UI than drawing an answer as cards. It turns the intermediate structure that a model reasons over into persistent, traceable interface objects that users can directly edit. The interface becomes both an output layer and a state layer for jointly maintaining evidence relationships, assumptions, and the boundary of professional judgment.
A closer look
GroundWork encodes grounds, claims, warrants, backing, qualifiers, and rebuttals as typed relations with explicit operations for adding, assessing, connecting, and revising them. The LLM may retrieve evidence from the corpus and propose claims or rebuttals, while experts control entry into shared state. The study found that domains recognized similar argument objects but differed in construction order and division of labor. An item's argumentative role also did not establish its credibility, recency, or contextual fit. The authors therefore separate a stable underlying representation from the flexible path through which a person reasons.
Who this is for
Teams building generative UI, knowledge workbenches, research assistance, legal technology, medical evidence tools, intelligence analysis, RAG, and human-AI collaboration systems.
Original sources
- GroundWork arXiv v1 paper and full text ↗arxiv.org
- GroundWork full text — system, expert feedback and limitations ↗arxiv.org
Revisions & corrections (1)
— Evidence-complete primary HCI research with an original bilingual synthesis about structured, persistent, editable human-LLM interface state; limitations are explicit.