RIVER Framework Evaluates Claim-Level Dynamics in Multi-Agent Systems

RIVER decomposes multi-agent outputs into atomic claims to measure convergence, exploration, and retention during interactions. Testing on 40 news-generation cases shows that pooled interaction increases claim convergence but narrows the semantic information space compared to distributed relay methods.