AI Forecasting Improves with Diverse Model Ensembles

New research into AI forecasting systems reveals that simply combining predictions from the most accurate large language models (LLMs) is not the optimal strategy for maximizing accuracy. Instead, the study emphasizes the critical role of diversity among the models within an ensemble. On binary questions from the Metaculus AI Benchmark, it was observed that many frontier LLMs tend to produce highly correlated predictions. This redundancy limits the value of adding more forecasts from similar models. The most effective ensembles were those that combined accurate models with diverse predictive patterns, meaning their errors were complementary rather than overlapping. For instance, models such as Grok 4 were found to contribute significantly to ensemble performance because their predictions were less correlated with other leading LLMs. This suggests that the true power of an "AI crowd" lies in strategically selecting models that offer distinct perspectives and error profiles, rather than indiscriminately sampling from a pool of high-performing but similar models.