Solve open math problems autonomously
DeepMind's Aletheia agent on Gemini Deep Think was evaluated on 700 open Erdős problems and autonomously solved four.
What they did
DeepMind built Aletheia, a math research agent on Gemini Deep Think, with three parts: a generator, a natural-language verifier and a reviser. The verifier flags flaws, and the agent either revises the solution or restarts. It can also admit failure. It uses Google Search and web browsing to check published literature. In a semi-autonomous evaluation it was run against 700 open problems from Bloom's Erdős database.
How it went
It autonomously solved four open Erdős questions: 652, 654, 1040 and 1051. Erdős-1051 also led to a generalization in a separate paper. The post claims no major-advance or breakthrough-level results.
Worth knowing
The agent's ability to say it can't solve a problem saved researchers time. The prompts and outputs are on GitHub, and the full details are in the arXiv paper.
Labels
DeepMind built Aletheia, a math research agent on Gemini Deep Think, with three parts: a generator, a natural-language verifier and a reviser.
DeepMind built Aletheia, a math research agent on Gemini Deep Think, with three parts: a generator, a natural-language verifier and a reviser.
It uses Google Search and web browsing to check published literature.
DeepMind's Aletheia agent on Gemini Deep Think was evaluated on 700 open Erdős problems and autonomously solved four.
In a semi-autonomous evaluation it was run against 700 open problems from Bloom's Erdős database.
The “Material” label failed its quality check and is not used for selection.
Details
- Agents
- Gemini
- Source
- other
- Source date
- undated
- Accounts
- 1
- In registry since
- Oct 12, 2026