Picking the right CoT-Recipe lifts held-out accuracy substantially, even with no in-context reasoning chains at test time: 📈 up to +130% on symbolic tasks using pretrained Qwen-2.5 models.