“In a previous DeepMind paper, researchers improved a task's performance from 5% to 95% on Gemini 1.5 Flash by adding manually-built reasoning strategies on top of optimized prompts.”