Shulou(Shulou.com)11/24 Report--

CTOnews.com, April 4, OpenAI's latest language model, GPT-4, can not only generate all kinds of text like humans, but also design and execute tests to evaluate and improve its performance. This "reflective" technology has enabled GPT-4 to make significant progress in a number of more difficult tests, with a 30 per cent increase in test performance.

GPT-4 is the most advanced system introduced by OpenAI after GPT, GPT-2 and GPT-3, and it is also the largest multimodal model (can accept image and text input, output text). It uses deep learning technology and artificial neural network to imitate human writing.

Researchers Noah Shinn and Ashwin Gopinath wrote in the paper: "We have developed a novel technology that allows AI agents to simulate human self-reflection and evaluate their performance. When GPT-4 completes various tests, it adds additional steps to allow it to design its own tests to check its own answers, identify errors and deficiencies, and then modify its solution based on its findings. "

In the HumanEval coding test, GPT-4 uses the self-reflection loop, and the accuracy increases from 67% to 88%.

GPT-4 can criticize its own performance by designing and executing tests, as shown by the AlfWorld test results, which can greatly improve its performance. The research team used this technique to conduct several different performance tests on GPT-4. In the HumanEval test, GPT-4 needed to solve 164unprecedented Python programming problems, with an accuracy of 67%. After using reflection technology, the accuracy was improved to 88%. In Alfworld testing, AI needs to make decisions and solve multi-step tasks by performing some permissible actions in a variety of interactive environments. After using reflective technology, the accuracy of GPT-4 increased from 73% to 97%, and only 4 tasks failed. In the HotPotQA test, GPT-4 had access to Wikipedia and answered 100 questions that required parsing content and reasoning from multiple supporting documents, with an accuracy of 34 per cent, but 54 per cent using reflection technology.

This study shows that the solution to the AI problem sometimes depends on AI itself. CTOnews.com found that this is a bit like building an adversarial network, which is a way for two AI to improve each other's skills, such as one AI trying to generate pictures that look like real pictures, and the other AI trying to tell which is fake and which is real. But in this case, GPT is both a writer and an editor, improving the quality of his output through self-reflection.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.