Tag
This paper studies how making inter-agent relation semantics explicit in LLM-based multi-agent systems acts as convergence pressure, increasing agreement but not reliably improving accuracy. The authors argue relational priors should be used diagnostically and task-specifically, not as a default add-on.