Design and Evaluation of Multi-Agent AI Oracle Systems for Prediction Market Resolution
Researchers have developed and evaluated multi-agent AI oracle systems designed to improve the accuracy of prediction market resolutions. By comparing independent aggregation and deliberative consensus approaches against single-LLM baselines, they found that confidence-weighted voting achieved the highest accuracy at 83.43%. The study also highlighted limitations due to error correlations and proposed hybrid AI-human systems that auto-resolve unanimous, high-confidence questions, flagging the rest for human review. AI
IMPACT Multi-agent systems show promise for improving AI-driven decision-making in complex prediction tasks.