Envs-FORGE Synthesizes Optimized Training Environments for Agent Reinforcement Learning

Envs-FORGE uses a mixed-integer linear program to select environment-synthesis actions that target a specific learning frontier. It automatically rewrites instructions, tests, and Docker environments based on verifier rewards to provide optimal difficulty for RL training, demonstrated on Qwen 3.5 35B.