StackingNet: Collective Inference Across Independent AI Foundation Models

A new meta-ensemble framework called StackingNet can aggregate the output predictions of independent, black-box AI foundation models to improve accuracy and reduce group-wise disparities, according to research published on arXiv. [1][2] The framework, introduced by Siyang Li and colleagues, operates at the inference stage without requiring access to any model's internal parameters or training data. [1][2] It functions by coordinating the complementary strengths of independently developed systems, which the authors note currently "remain isolated and cannot readily share their capabilities." [2] Foundation models are large machine-learning models trained on vast datasets for broad applicability, with prominent examples including OpenAI's GPT series and Google's BERT. [7] StackingNet was tested across three distinct domains: language comprehension, visual attribute estimation, and academic paper rating. [1][2] In each case, the meta-ensemble consistently outperformed both individual models and classic ensemble methods. [1][2] The performance gains persisted even when the underlying base models were uniformly strong, and the benefits widened as the pool of contributing models grew more diverse. [2] The mechanism behind the improvement is attributed to variance reduction and consensus alignment among independent models, rather than any emergent group cognition. [2] Beyond boosting accuracy, StackingNet also ranks model reliability and can identify or prune models that degrade overall performance. [1][2] This approach treats model diversity—often a source of inconsistency—as a resource for cooperation. [2] The work arrives amid a period of rapid scaling in artificial intelligence. The field of AI research was formally founded at a 1956 Dartmouth workshop, and has since cycled through periods of optimism and so-called "AI winters" of reduced funding. [4] Investment surged after 2012 when graphics processing units began accelerating neural networks, and accelerated further after 2017 with the introduction of the transformer architecture. [9] The subsequent AI boom of the 2020s has been fueled by large language models and generative AI applications. [8][9] Foundation models now span modalities including text, images, music, and robotic control. [7] StackingNet offers a practical method for coordinating these increasingly powerful but siloed systems. By turning a collection of black-box models into a cooperative inference network, the framework provides a path toward what the researchers call "coordinated artificial intelligence." [2]