Microsoft Research has developed an advanced prompting method called Medprompt+ to enhance the performance of OpenAI’s GPT-4 in language understanding benchmarks, surpassing Google’s Gemini Ultra model. Led by Chief Science Officer Eric Horvitz, Research Engineering Director Harsha Nori, and Principal Researcher Yin Tat Lee, the Microsoft Research team has successfully propelled GPT-4 to new heights in the Measuring Massive Multitasking Language Understanding (MMLU) evaluations.
By adopting Medprompt+, Microsoft has tipped the scales back in favor of GPT-4, which was initially outperformed by Google’s Gemini Ultra on various benchmarks. The new technique involves steering GPT-4 with prompts that elicit higher quality responses. Medprompt+ improves on the original approach by integrating simple and complex prompts, formulating a policy for integrating outputs, and establishing a superior evaluation method. This tailored approach combines base and inferred confidences of candidate answers to provide the most accurate response.
As of now, these advanced prompting techniques are limited to internal applications and are not yet available to the general public. Microsoft acknowledges the importance of improving the user experience for individuals with varying levels of tech proficiency.
Committing to further research, Microsoft aims to explore zero- or few-shot prompting strategies essential for intuitive use of language models. In an effort to foster transparency and collaboration, Microsoft Research has made additional details about the new prompting techniques and their applications accessible through their GitHub repository. This allows the wider AI research community to gain insights into ongoing developments and innovations in the field.
Microsoft’s breakthrough in optimizing prompt-based methods marks a significant milestone. While the company continues its research endeavors, it prioritizes creating an out-of-the-box experience that caters to users’ diverse needs and ensures seamless adoption.
By empowering GPT-4 with Medprompt+, Microsoft Research has pushed the boundaries of AI prompting techniques, exceeding Google’s Gemini Ultra model in language understanding benchmarks. Although these advancements are currently limited to internal use, they pave the way for further progress in the field of AI language models, promising more refined and accurate responses to user queries. The Microsoft Research team’s dedication and commitment to innovation continue to drive the evolution of AI technology, benefiting users worldwide.
