GPT-6 Astra resolved reasoning loops in Qwen-350-8B within 50 forward passes without calculating gradients or conducting traditional training