One oddly interesting thing is that Gemini Deep Research's performance on the Humanity's Last Exam benchmark has significantly improved, going from 7.95% in December 2024 to a SoTA score of 26.9% and 32.4% with higher compute in June 2025[1].
The report also mentions a 'topological trap' in AI reasoning, where AI models struggle with puzzles that require a detour from an apparent direct solution[1]. Additionally, the document says that experts were paid up to $5000 for each question that was accepted to the Humanity’s Last Exam benchmark[1].
Sign Up To Try Advanced Features
Get more accurate answers with Super Search, upload files, personalized discovery feed, save searches and contribute to the PandiPedia.
Sorry, Pandi could not find an answer.
Let's look at alternatives:
- Modify the query.
- Start a new thread.
- Remove sources (if manually added).
- Request a manual search from our human research team.