Local reproduction of Dream-RSI: model weights stay frozen while an executable search policy evolves, via persistent discovery trees and an offline replay engine that scores candidate policies at zero extra inference cost.
python tree-search research-reproduction llm-agents recursive-self-improvement offline-evaluation self-improving-agents replay-simulation policy-evolution
-
Updated
Oct 4, 2026 - Python