Aletheia

Aletheia is a multi-agent mathematics research system developed by Google DeepMind. It is not publicly available — access is not offered via API, early access programme, or institutional licence. This note is created for situational awareness, to track developments and inform any future access pathway.

What it does

Aletheia uses a three-agent loop (Generator, Verifier, Reviser) powered by Gemini Deep Think, operating end-to-end in natural language. Unlike formal proof systems such as AlphaProof or harmonic-aristotle, it does not require problems to be formalised in Lean 4. It uses Google Search, web browsing, and Python execution to reduce hallucinations and verify computations.

Key results (March 2026)

• Feng26: First fully AI-generated publication-grade mathematics paper, in arithmetic geometry. Classified Level A2 (Essentially Autonomous, Publication Grade) under the paper’s own Autonomy Levels taxonomy.
• Erdős problems: Deployed on 700 open problems; resolved four genuine open conjectures (Erdős-652, 654, 1040, 1051).
• FirstProof benchmark: 6 of 10 research-level problems solved (best-of-2); one solution at publication grade.

Full analysis: aletheia-autonomous-maths-2026

Autonomy Levels taxonomy

The paper introduces a proposed standard for documenting AI contribution to mathematical results:

• Autonomy axis: H (Primarily Human) / C (Human-AI Collaboration) / A (Essentially Autonomous)
• Significance axis: 0 (Negligible) through 4 (Landmark Breakthrough), with 2 = Publication Grade

This taxonomy is potentially applicable beyond Aletheia — it provides vocabulary for Sheffield’s GenAI Principles for Research when describing AI contribution in publications.

Limitations

• Not publicly accessible; no realistic pathway for Sheffield adoption in the near term.
• Success rate on open problems is low (6.5% meaningfully correct on Erdős database).
• Prone to specification gaming and hallucination even with tool use.
• Results are currently elementary relative to frontier research.
• Data contamination from pretraining cannot be fully excluded.

Governance notes

Any future equivalent system would involve sending mathematical problem statements to US-hosted infrastructure. A DPA review would be required before use with unpublished research. Researchers should not share unpublished problem statements with externally hosted AI systems.

Sources

• arXiv:2602.10177v3 — primary paper
• GitHub: google-deepmind/superhuman/tree/main/aletheia