In the BrowseComp evaluation, which tests whether an AI model can navigate a wide range of websites to find information that is deliberately hard to find, dynamic filtering improves the accuracy of ...