Agentic Transformers Provably Learn to Search via Reinforcement Learning

In three linesTheoretical study showing how transformers learn to implement tree search (DFS) via RL. A two-head transformer naturally emerges from policy gradient training on stochastic trees without expert demonstrations. The model generalizes to unseen depths and adapts its strategy based on goal distributions.

Summary generated by Claude — human-verified