In an intriguing piece of research, Google DeepMind researchers recently discovered that they could improve the AI’s math skills by inserting phrases like ‘let's think step by step’ in their prompts. According to a report in Ars Technica, they found that the most effective phrase for improving performance was, ‘take a deep breath and work on this problem step by step.’
When used with Google's PaLM 2 large language model (LLM), the phrase helped it to record accuracy levels of 80.2 percent in tests, compared to only 34 percent accuracy when there was no special prompting. Using the ‘let’s think step by step’ prompt pushed accuracy up to 71.8 percent.
DeepMind researchers came to the discovery through a new methodology known as Optimization by PROmpting (OPRO), which they outlined in a recent paper. The method aims to improve the performance of large language models (LLMs) such as OpenAI’s ChatGPT and Google’s PaLM 2 using natural language (everyday speech).
It is not clear why the phrases make such a big difference in the LLMs’ performance. LLMs have no ability to reason in strictly human terms and are merely pulling from a massive data set of language phrases found on the web and in books.
Researchers hope the discovery will help in the overall accuracy levels of LLMs.