Milkyway System Introduces Internal Feedback for LLM Future Prediction Agents

A new research paper introduces Milkyway, a self-evolving agent system designed for future prediction tasks where decisions must be made before outcomes are known. The system addresses the challenge of predicting unresolved questions using only publicly available information at the time of prediction. Unlike existing approaches that primarily improve from final outcomes, Milkyway leverages temporal contrasts between earlier and later predictions to expose omissions in the earlier prediction process. This signal, termed internal feedback, helps guide earlier factor tracking, evidence gathering, interpretation, and uncertainty handling. The system keeps the base model fixed while updating a persistent component. The research was announced on arXiv with the identifier 2604.15719v1, highlighting the difficulty of public evidence evolving while useful supervision arrives only after question resolution. The paper frames such problems as future prediction, where LLM agents must form predictions for unresolved questions. Most existing methods still improve mainly from final outcomes, which are too coarse for earlier guidance. The Milkyway approach represents a significant advancement in AI agent systems for prediction tasks.