What's this about?
A replacement for PPO-Clip's flat, Euclidean clipping in LLM reinforcement learning: clip on the curved (Riemannian) geometry of the policy space instead, preventing the exploration collapse where models converge on a narrow set of answers (9↑).
What exactly is "exploration collapse"?
RL training rewarding what already works until the model stops trying anything else — the diversity of attempts shrinks, and with it the chance of finding better solutions.
Why is it worth noting?
RLVR is the engine of the current reasoning boom; optimizer-level fixes to its failure modes propagate to every lab's training runs.