New AI system called Milkyway learns to make better future predictions by analyzing how evidence and predictions change over time rather than waiting for final outcomes.

Milkyway is a self-evolving agent system that keeps the base LLM fixed while updating a persistent memory component to improve predictions

The system uses 'internal feedback' - temporal contrasts between earlier and later predictions on the same unresolved question - to guide learning

Traditional approaches rely on final outcomes for supervision, which is too coarse-grained to improve factor tracking, evidence gathering, and uncertainty handling

By revisiting the same unresolved questions over time, the agent can identify omissions in its earlier prediction process and continuously refine its approach