SPADE: Self-Play in Adaptive Synthetic Executable Environments

Benchmark Model Rank Results
mathematical-reasoning-on-aime-2026SPADE (Games, Qwen3-30B-A3B-Instruct-2507)#35Accuracy: 74.4
mathematical-reasoning-on-aime-2025SPADE (Games, Qwen3-8B)#59Accuracy: 68.8
multiple-choice-question-answering-mcqa-on-gpqa-diamondSPADE (Games, Qwen3-30B-A3B-Instruct-2507)#68Accuracy: 75.8
code-generation-on-livecodebench-v6SPADE (Games, Qwen3-8B)#69Pass@1: 0.494