ARLArena introduces SAMPO for stable agentic RL

This is the kind of paper agent builders should pay attention to: less about flashy demos, more about making long-horizon agent training actually reproducible. If SAMPO holds up, it could help move agentic RL from brittle lab curiosity toward a usable systems recipe.

• –The core contribution is not just another optimizer tweak; ARLArena standardizes the testbed so stability claims are easier to compare across tasks
• –SAMPO targets the biggest practical pain point in agentic RL: runs that collapse before agents learn useful multi-step behavior
• –Coverage across web, search, embodied, and game-style environments matters because many current RL-for-agents results only look good in narrow settings
• –The open GitHub release gives researchers a concrete baseline for extending to software engineering and tool-using agents, which the repo lists as an upcoming direction