A research team has built a framework that makes competing AI models cooperate — without any of them revealing how they work inside.
The system, called StackingNet, sits above a pool of independently built foundation models and aggregates their output predictions at inference time. It needs no access to internal parameters or training data, which means it can work with proprietary or closed models. In tests spanning language comprehension, visual attribute estimation, and academic paper rating, StackingNet beat both individual models and classic ensemble methods. It also identifies models that drag the group down and can prune them — useful when you don't fully trust every model in the pool.
The gains come from variance reduction and consensus alignment, not from any mysterious emergent intelligence among the models — the paper is explicit on this point, which is worth noting given how much ensemble hype leans on vague "collective" language. More practically, the accuracy improvements grow as the pool of models becomes more diverse, which flips the usual problem: heterogeneity that normally causes inconsistency becomes an asset.
The broader implication is that scaling laws may not be the only path forward. Instead of pouring resources into one larger model, organizations could pool specialized models and let a coordination layer do the work — a cheaper, more modular bet that also sidesteps the need to share proprietary weights with anyone.