According to @_avichawla, RULER lets LLMs score trajectories from plain English criteria, easing brittle reward design for agents, as reported on X.