Belief models crack hidden‑info games

Researchers trained a second neural network to predict opponents' hidden pieces and then sampled plausible game states to run a search, enabling an AI (Ataraxos) to beat the world Stratego champion while using modest compute. The approach reduced required self‑play games and compute by orders of magnitude compared with prior work and was published in Nature. — If cheap belief‑modeling lets agents plan under uncertainty, it lowers the technical barrier to deploying autonomous systems in real‑world domains where information is incomplete, raising policy, security, and labor questions.