Google DeepMind Aletheia Autonomously Solves 4 Open Erdős Problems and Publishes Peer-Reviewed Paper Without Human Intervention

MarkTechPost·high signal

Aletheia, powered by Gemini 3 Deep Think, autonomously solved 4 open problems from the Erdős Conjecture database and generated a peer-reviewed mathematics paper (Feng26) without any human intervention, announced February 2026. The agent uses a Generator-Verifier-Reviser harness with Google Search grounding to prevent hallucinated citations, scoring 91.9% on IMO-ProofBench Advanced — with 100x compute reduction vs the 2025 version for IMO-level problems. This is the clearest documented case of an AI agent completing a full scientific research cycle from hypothesis to publishable result in a rigorous formal domain.