SPADE Has One Model Write Executable Gym Environments and Rewards It by How Much a Hint Helps the Solver
Import AI 470 covers SPADE, from a nine-university group including UW, Stanford, CMU, MIT, and NUS, which alternates an LLM between Environment Designer writing long-horizon training environments as Python programs with a Gym-style interface and Reasoning Agent learning to act in them. The designer's reward is hint-based regret: attach a privileged hint such as a partial solution sketch, and score the gap in the solver's return with versus without it. Training three Qwen3 backbones with GRPO for 400 rollouts of 25 environments each, Qwen3-30B-A3B reaches a 58.3 game-suite average, 8.1 over base and 5.3 over the strongest fixed-environment baseline, with the same recipe improving tool-use environments on every backbone. Code and checkpoints are public.
↳ Follow the thread