PRISM: PRIvileged Sibling-guided Multi-Agent Topology Design

PRISM (PRIvileged Sibling-guided Multi-Agent Topology Design) is a post-training method that recovers per-query signal from existing corpora by using a teacher with privileged information from a sibling graph to guide a student model in generating collaboration topologies for multi-agent systems. The approach addresses the limitation of current methods that lack per-query preference ranking, causing drift toward frequently occurring topologies rather than query-specific ones.