Skip to content
Anton Wolter
Research

How can experts teach agents through their edits?

Context-Mediated Domain Adaptation

Treating experts' edits to AI-generated work as a way to adapt multi-agent systems to their domain.

Team
Anton Wolter, Leon Haag, Vaishali Dhanoa, Niklas Elmqvist

Problem

Domain experts know more about their field than they can write down as a prompt or a specification. When they correct an AI-generated draft, by fixing terminology, restructuring an argument or shifting emphasis, those edits reveal what they know.

Approach

Instead of asking experts to describe their domain up front, we treat their modifications to the agents' output as context that reshapes how the agents reason. The specification then grows out of iterative collaboration rather than being fixed at the start.

System

We built Seedentia, a multi-agent sensemaking system in which agent behavior adapts to the patterns in a user's corrections.

Findings

In an evaluation with domain experts, 46 domain knowledge entries were extracted from their modifications.

Publications