The Flexibility Trap (JustGRPO)

ICML 2026 Outstanding Paper (LeapLab with THUNLP and Alibaba co-authors): a minimalist GRPO recipe for diffusion LLMs showing that the flexibility of arbitrary-order decoding is a trap for RL — and that a deliberately simple recipe recovers it. The reasoning-RL line's third landmark after RLVR-limits and Absolute Zero.