Gemini Models Experience Loops

Researchers have noticed that Gemini models, including versions 3.1 Pro and 3.5 Flash, are getting stuck in thinking loops more often. To investigate, a test of 100 tasks was conducted, revealing that the 3.5 Flash model got stuck 23 times, while the 3.1 Pro model got stuck 16 times. This issue suggests a potential problem with the models' ability to process and complete tasks efficiently. The frequency of these loops is concerning and may impact the overall performance of the Gemini models. The cause of this issue is not yet clear, but it highlights the need for further testing and refinement of the models. The study used Google Antigravity and Google AI Studio to analyze the models' behavior.