ARLArena: Demystifying Policy Gradient Stability in Agentic Reinforcement Learning

Submitted to ICML, 2026