In early 2026, Google DeepMind unveiled Aletheia, a specialized AI research agent powered by the Gemini 3 Deep Think model. Named after the Greek goddess of truth and disclosure, Aletheia represents a major leap from competition-level mathematics (like the IMO) to fully autonomous, professional-level scientific discovery.
Aletheia was introduced to bridge the "evaluation gap" between solving textbook problems and conducting original research. Unlike standard LLMs that often hallucinate citations or logic, Aletheia utilizes a specialized "agentic loop" consisting of three parts:
Generator: Proposes candidate solutions for complex research problems; Verifier: A natural language mechanism that rigorously checks for logical flaws or calculation errors; Reviser: Corrects identified errors iteratively until a final output is approved.
By applying inference-time scaling - essentially giving the model more "thinking time" rather than just more training data - Aletheia achieved a record 95.1% accuracy on the IMO-Proof Bench Advanced. More significantly, it demonstrated the ability to navigate vast scientific literature using Google Search to ensure factual and citation accuracy.
Autonomous Mathematics Research Levels
As of February 2026, Aletheia has already moved beyond simulations into real-world contributions: Autonomous Research Paper: Aletheia generated a complete research paper (internally cited as Feng26) in arithmetic geometry without any human intervention, calculating complex structural constants known as "eigenweights."; The Erdős Conjectures: Deployed against 700 open problems from Bloom’s Erdős database, Aletheia successfully resolved four open questions autonomously. One solution (Erdős-1051) was so robust it led to a generalized follow-up paper co-authored with humans; The FirstProof Challenge: In late-February 2026 news, Aletheia reportedly solved 6 out of 10 problems in the "FirstProof Challenge" - a benchmark designed to be "contamination-proof" - outperforming competitors like OpenAI’s private reasoning models.
The mathematical community has reacted with a mix of awe and cautious pragmatism. DeepMind has proposed a taxonomy of “Autonomous Mathematics Research Levels,” similar to self-driving car levels, to standardize how AI-assisted discoveries are credited in the future.