Diverse Evidence Boosts Multi-Agent AI Forecasting Accuracy

This research highlights a critical flaw in current multi-agent forecasting systems: when all agents receive identical information, their deliberation often leads to "herding" rather than genuine improvement. To address this, the study proposes introducing designed information asymmetry, where agents are given both shared public evidence and unique private subsets of information. This ensures each agent possesses exclusive knowledge that can only be shared through deliberation. The theoretical basis for this approach demonstrates that partitioning evidence in this manner reduces the correlation of errors between agents. The researchers instantiated this concept in InfoDelphi, a framework that combines relevance-aware evidence routing, iterative deliberation based on rationales, and confidence-weighted aggregation of forecasts. Evaluations on PolyGym, a benchmark of 375 binary forecasting questions, show that InfoDelphi significantly outperforms leading single-agent and multi-agent baselines, improving Brier scores by 12-18% and accuracy by 4-8 percentage points. The findings confirm that diverse input information is the key enabler for effective multi-agent reasoning and superior forecasting.