Article Featured Image Graphic - Microsoft Researchers Extend GPT-4 With New Prompting Method - SEO Gazette

Microsoft’s GPT-4 Prompting Technique Research

Dec. 4, 2023 5:17 am PST | SEO Gazette | By Luke Ross

Revolutionizing AI: Microsoft’s New GPT-4 Prompt Technique

In a groundbreaking study, Microsoft researchers have unveiled a novel prompting method that significantly enhances the performance of GPT-4, a generalist AI model, enabling it to surpass even specialist AI models trained for specific topics. This advancement, detailed in a recent Search Engine Journal article, marks a pivotal shift in the capabilities of generative AI.

Advanced Prompting Techniques: A New Frontier

The study confirms what many advanced generative AI users have already discovered: sophisticated prompting techniques, known as prompt engineering, can produce remarkable results in AI-generated content. A key technique employed by the researchers is Chain of Thought (CoT) reasoning, a method initially outlined by Google in May 2022. CoT enables AI to break down tasks into reasoned steps, enhancing its problem-solving abilities.

Peter Hatherley, founder of Authored Intelligence web app suites, lauds the effectiveness of CoT prompting, stating, “Chain of thought prompting takes your seed ideas and turns them into something extraordinary.” Hatherley incorporates CoT into his custom GPTs to amplify their capabilities.

Medprompt: A Benchmark in AI Performance

The researchers tested their Medprompt technique against four foundation models, including Flan-PaLM 540B, Med-PaLM 2, GPT-4, and GPT-4 MedPrompt, using various medical benchmark datasets. GPT-4, equipped with Medprompt, consistently outperformed its competitors across all datasets.

The Three Pillars of Medprompt

The Medprompt technique is built on three prompting strategies:

Dynamic Few-Shot Selection: This enables the AI model to select relevant examples during training, focusing on a small set of representative examples.

Self-Generated Chain of Thought: This technique automates the creation of chain-of-thought examples, guiding the AI model through reasoning steps in natural language.

Choice Shuffle Ensembling: To combat position bias and greedy decoding, this approach shuffles multiple choice answers, enhancing the diversity and accuracy of the AI’s responses.

Implications for Generative AI

The success of Medprompt extends beyond the medical domain. It demonstrates that general foundation models like GPT-4, when equipped with advanced prompting techniques, can achieve high-quality output in any knowledge area without the need for extensive domain-specific training. This finding has profound implications for the future of generative AI, potentially reducing the need for resource-intensive specialized model training.


Microsoft’s research represents a significant leap in the field of generative AI, showcasing the power of advanced prompting techniques. The Medprompt method not only elevates the capabilities of GPT-4 but also opens new possibilities for AI applications across various knowledge domains.

For more details, read the full research paper: Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine (PDF).

Article written by Luke Ross, The SEO Gazette

Copyright Disclaimer: Under section 107 of the Copyright Act of 1976, allowance is made for “fair use” for purposes such as criticism, comment, news reporting, teaching, scholarship, education and research. Fair use is a use permitted by copyright statute that might otherwise be infringing.

Any references to other articles, information, or images are purely for the sake of discussion and analysis. We do not hold any rights to images or external content referenced, and we disclaim any liability arising from the use or application of the information provided herein.