Large Language Models Are Multitask Chain-of-Thought Prompting Optimizers

Area of Science:

• Artificial Intelligence
• Natural Language Processing

Background:

• Large language models (LLMs) show impressive reasoning abilities but are highly sensitive to prompt design.
• Current prompt engineering methods like manual chain-of-thought (CoT) and automated generation are labor-intensive and lack generalizability.

Purpose of the Study:

• To develop a more efficient and generalizable method for enhancing LLM reasoning performance.
• To leverage LLMs for iterative prompt self-refinement and ensemble top-performing prompts.

Main Methods:

• Treated the LLM as a multitask optimizer for iterative prompt self-refinement using natural language task descriptions and few-shot in-context learning (ICL).
• Ensembled top-performing prompts at inference time to account for task-dependent prompt sensitivity.

Main Results:

• Achieved substantial performance gains over prior baselines across various LLMs.
• Demonstrated improvements of up to 6.0% on mathematical reasoning and 10.2% on commonsense reasoning benchmarks.

Conclusions:

• The proposed multitask optimization and prompt ensembling approach offers a powerful and efficient alternative to traditional prompt engineering techniques.
• This method significantly enhances LLM reasoning capabilities, particularly in complex domains like mathematics and commonsense understanding.