Openai has just iroduced its advanced models of artificial ielligence, O3 and O4-Mini, which performs stunning in areas such as coding and solving mathematical problems. But these models are facing a serious problem: hallucinations, that is, producing inaccurate or fabricated information.
Corary to expectation, these models are more illusions than previous OpenAI models, such as O1 and Gpt-4O. This, which even Openai itself does not know the exact reason, has raised concerns about the accuracy of these new technologies.
According to Openai reports, the O3 model produces inaccurate information in 5 % of personal information questions (PERSONQA benchmarks), compared to the previous O1 and O3-MINI models, respectively.
The O4-MINI model has been worse and has an illusory illusions in 5 % of cases. Independe experimes by the Translux laboratory have also shown that O3 sometimes makes unrealistic claims about its accouability process, such as executing the code on an imaginary device. Experts believe that the learning methods used in these models may exacerbate this problem.
This illusory can limit the use of new models in areas such as law or medicine, which is critical accuracy. However, Openai is considering solutions such as adding web search capabilities that can improve accuracy. For example, the Gpt-4O model is up to 5 % with web search. While Openai and the artificial ielligence industry are moving towards advanced reasoning models, solving hallucinations has become a key challenge that requires further research.
