A team of researchers from Google DeepMind and academic institutions has unveiled Aletheia, an AI system capable of autonomous mathematics research. The system, detailed in a paper titled Towards Autonomous Mathematics Research, leverages an advanced version of Gemini Deep Think and novel inference-time scaling to tackle problems beyond International Mathematical Olympiad level.

Aletheia demonstrates capabilities ranging from solving PhD-level exercises to generating original research papers without human intervention. In one milestone, the system calculated structure constants in arithmetic geometry called eigenweights for a paper (Feng26) entirely through AI computation. Another achievement involved human-AI collaboration on proving bounds for systems of interacting particles called independent sets (LeeSeo26).

The researchers also conducted an extensive evaluation of 700 open problems from Bloom's Erdos Conjectures database, with Aletheia autonomously solving four open questions. To promote transparency in AI-assisted mathematics, the team proposes quantifying autonomy levels and introducing human-AI interaction cards. The work represents a significant step toward AI systems that can contribute meaningfully to mathematical research rather than just solving competition problems.