(Link to paper)[https://arxiv.org/pdf/2201.11903.pdf]
This work explores how the reasoning ability of large language models can be unlocked by a simple method motivated by two ideas.
Guide the model to solve a arithmatic/reasoning problem the same way as a human does. We humans break these problems into smaller problems and then solve these smaller problems step by step.
DIRECT QUOTE FROM THE PAPER
Consider one’s own thought process when solving a complicated reasoning task such as a multi-step math word problem. It is typical to decompose the problem into intermediate steps and solve each before giving the final answer: “After Jane gives 2 flowers to her mom she has 10 . . . then after she gives 3 to her dad she will have 7 . . . so the answer is 7.” The goal of this paper is to endow language models with the ability to generate a similar chain of thought—a coherent series of intermediate reasoning steps that lead to the final answer for a problem.
LLM have exhibited two behaviors that can be leveraged for making them better at arithmatic and commonsense tasks.
Task results can be improved by generating natural language rationales that lead to final answer.
Few shot technique yields desired result