The Google DeepMind team found that using the new "mantra" (Take a deep breath) combined with the familiar "Let's think step by step" (Let's think step by step) large model scores improved from 71.8 to 80.2 on the GSM8K data set. (Let's think step by step) on the GSM8K dataset from 71.8 to 80.2 points.
The big model "deep breathing" math up another 8 points! Google's DeepMind finds AI designs its own cue words to outperform humans
Previous: 英伟达等以1亿美元投资AI生物创企Inceptive
Next: 老黄参投，MIT华人女科学家融资2亿美元