Test a math agent on hard open questions

DeepMind’s Aletheia, built on Gemini Deep Think, found proofs for four Erdős problems in a large benchmark run.

Done with Gemini· Advanced

What they did
DeepMind made Aletheia as a research agent for mathematics on Gemini Deep Think. The system is split into three roles: one drafts candidate arguments, one checks the reasoning in everyday language, and one improves the attempt after problems are found. It can also decide that no solution is in reach, and it looks up prior work with Google Search and web pages before moving on. Researchers then let it operate in a semi-autonomous test using 700 unsolved Erdős questions taken from Bloom's database.
How it went
Aletheia independently settled Erdős-652, 654, 1040, and 1051. The 1051 case also appeared as a broader statement in another paper. The report says the work did not amount to a major breakthrough.
Tips
A system that can stop itself when the trail goes cold may save a lot of manual checking. Even when an agent gives a complete proof, a person should still inspect the final steps by hand.

Try it yourself with Gemini

I want to investigate this open math problem: [problem statement]. First, look for earlier papers or references to see whether it has already been resolved. Then try to prove it or find a counterexample, checking your own reasoning as you go and correcting mistakes when they appear. If your current line of attack fails, try a different one. If you cannot make real progress, say so plainly. Before you do anything irreversible, such as submitting, sending, deleting, or paying, stop and ask me to confirm. Finish with one of these outcomes: solved, partial progress, or stuck, and name the steps I should review myself.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on deepmind.google ↗

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.