DIVERGE: Closed-Loop Semantic Routing for Creative Agent Populations

Large Language Models (LLMs) can be used to generate candidate ideas cheaply. However, parallel sampling can waste effort on overlapping ideas, whereas it is difficult to prevent models from getting trapped in narrow lines of enquiry when prompting sequentially. We present DIVERGE, a Meta-Agent designed to improve the coverage of ideas achieved by a population of sub-agents that it manages. On the Alternative Uses Task, DIVERGE’s closed-loop control (using an offline-calibrated, frozen semantic scaffold) improves population coverage relative to a population-aware one-shot manager using the same scaffold. To mitigate some shortcomings of the OCSAI creativity/novelty metric, we include a relative originality ranking algorithm that evaluates novelty between populations produced by different methods. Exploratory work on enhancing DIVERGE suggest that interface construction and online allocation are distinct Meta-Agent capabilities.