A novel bio-inspired paradigm harnesses evolutionary processes and world leakage to create self-evolving future prediction agents that overcome limitations in current LLM-based forecasting systems.

Wei et al. (arXiv:2604.15719) introduce Milkyway, which keeps the base LLM fixed and instead maintains an evolving "future prediction harness" for factor tracking, evidence gathering, interpretation, and uncertainty handling. The system extracts internal feedback from temporal contrasts when the same unresolved question is revisited, writing reusable guidance back into the harness before any final outcome; a retrospective check occurs only after resolution (Wei et al., 2026). On FutureX this raised scores from 44.07 to 60.90 and on FutureWorld from 62.22 to 77.96.

Coverage of forecasting agents has largely overlooked how temporal internal feedback supplies dense supervision signals years before outcomes, a mechanism directly analogous to Salimans et al. (arXiv:1703.03864) evolution strategies that optimize via environmental fitness without gradients, and to Shinn et al. (arXiv:2303.11366) Reflexion agents that use self-generated feedback; Tetlock & Gardner (2005) forecasting data similarly show incremental belief updating is key yet absent from outcome-only RLHF. Original abstracts miss that "world leakage"—public information drift—functions as a continuous evolutionary pressure enabling adaptation across questions rather than isolated episodes.

Viewed through the bio-inspired lens, Milkyway's persistent harness encodes transferable evolutionary memory, addressing autonomous AI scalability limits by allowing self-improvement via environmental signals instead of scarce labeled resolutions and mitigating catastrophic forgetting seen in continual fine-tuning approaches.