RLVE: Scaling Up RL with Adaptive Verifiable Environments

The environment axis of RL scaling, from a UW-led team (Zhiyuan Zeng; senior authors Natasha Jaques, Simon Du, Yulia Tsvetkov) with Ai2 (Hamish Ivison), UIUC and LMSYS. RLVE replaces static RL datasets — whose learning signal vanishes when problems are too easy or too hard for the current policy — with verifiable environments that procedurally generate problems and adapt their difficulty distribution to the policy as training progresses, each with algorithmically checkable rewards.

Ships RLVE-Gym, a 400-environment suite built through manual environment engineering. Accepted to ICML 2026; IsoCompute cites it as the environment-scaling dimension of the LLM RL-scaling literature.