Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2609–2634 July 9-14, 2023 ©2023 Association for Computational Linguistics Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models Lei Wang1 Wanyu Xu2 Yihuai Lan Zhiqiang Hu3 Yunshi Lan4 Roy Ka-Wei Lee3 Ee-Peng Lim1∗ 1Singapore Management University 2Southwest Jiaotong University 3Singapore University of Technology and Design 4East China Normal University Abstract Large language models (LLMs) have recently been shown to deliver impressive performance in various NLP tasks. To tackle multi-step rea- soning tasks, few-shot chain-of-thought (CoT) prompting includes a few manually crafted step-by-step reasoning demonstrations which enable LLMs to explicitly generate reasoning steps and improve their reasoning task accu- racy. To eliminate the manual effort, Zero- shot-CoT concatenates the target problem state- ment with “Let’s think step by step” as an in- put prompt to LLMs. Despite the success of Zero-shot-CoT, it still suffers from three pit- falls: calculation errors, missing-step errors, and semantic misunderstanding errors. To ad- dress the missing-step errors, we propose Plan- and-Solve (PS) Prompting. It consists of two components: first, devising a plan to divide the entire task into smaller subtasks, and then car- rying out the subtasks according to the plan. To address the calculation errors and improve the quality of generated reasoning steps, we extend PS prompting with more detailed in- structions and derive PS+ prompting. We eval- uate our proposed prompting strategy on ten datasets across three reasoning problems. The experimental results over GPT-3 show that our proposed zero-shot prompting consistently out- performs Zero-shot-CoT across all datasets by a large margin, is comparable to or exceeds Zero-shot-Program-of-Thought Prompting, and has comparable performance with 8-shot CoT prompting on the math reasoning problem. The code can be found at https://github.com/AGI- Edgerunners/Plan-and-Solve-Prompting. 1 Introduction Large language models (LLMs) (Brown et al., 2020; Thoppilan et al., 2022; Chowdhery et al., 2022) have recently proven highly effective in var- ious NLP tasks. Unlike the previous pre-trained language models (PTMs) (Devlin et al., 2019; Liu ∗Corresponding author. 5 10 15 20 25 30 35 Ratio (%) 7 12 27 Calculation Error Step Missing Error Semantic Misunderstanding Figure 1: Error analysis of 46 GSM8K problems with in- correct answers returned by Zero-shot-CoT using GPT- 3 LLM. Following Wei et al. (2022b) and Wang et al. (2022a), we assign “Calculation Error” (7%), “Step Missing Error” (12%), or “Semantic misunderstanding Error” (27%) to each incorrect answer. et al., 2019), these LLMs are typically provided as a service, with no access to model parameters due to commercial considerations and potential risks of misuse (Sun et al., 2022). Thus, it is challenging to fine-tune LLMs for downstream tasks (He et al., 2021; Houlsby et al., 2019; Devlin et al., 2019). Instead, we leverage LLMs to solve complex rea- soning problems by eliciting their strong reasoning abilities over their embedded knowledge using in- structions (or trigger sentences). So far, LLMs have shown impressive abilities to solve new reasoning problems by simply conditioning them on a few illustrative examples …

Content truncated to 500 words