Context-Mediated Domain Adaptation in Multi-Agent Sensemaking Systems

A team of researchers including Anton Wolter and Niklas Elmqvist has published a paper on arXiv introducing a novel paradigm called 'context-mediated domain adaptation.' The core insight addresses a fundamental limitation in human-AI collaboration: domain experts possess tacit knowledge they cannot easily articulate through explicit prompts or specifications. When these experts modify AI-generated outputs—by correcting terminology, restructuring arguments, or adjusting emphasis—current systems typically treat these edits as simple endpoint corrections. The researchers argue these modifications are actually rich, implicit specifications that should reshape the AI's subsequent reasoning processes.

To implement this, they built Seedentia, a web-based multi-agent sensemaking framework. Seedentia establishes bidirectional semantic links between the user-edited artifacts and the underlying reasoning of its LLM-powered agents. This allows the system to perform 'specification bootstrapping,' where vague initial prompts evolve into precise domain models through iterative collaboration. It also enables 'implicit knowledge transfer' by reverse-engineering the intent behind user edits and 'in-context learning' where agent behavior adapts based on observed correction patterns.

In an evaluation with domain experts who generated and modified research questions from academic papers, the Seedentia system successfully extracted 46 distinct domain knowledge entries solely from analyzing user modification patterns. This demonstrates the technical feasibility of capturing latent expertise through edit analysis. The paper notes that while promising, the limited sample size constrains broader conclusions about systematic quality improvements. The work represents a significant shift from viewing human edits as corrections to treating them as a continuous, rich source of training data for adaptive multi-agent systems.