Google DeepMind announced Aletheia, a fully autonomous AI agent built on Gemini 3 Deep Think that solved 6 out of 10 novel, unpublished research-level math problems in the FirstProof challenge. The system operates zero-shot without human hints, using a multi-agent loop of Generator, Verifier, and Reviser components plus Google Search integration. Notably, it outputs 'No solution found' rather than hallucinating answers, which researchers consider a key reliability feature. Aletheia also scored ~91.9% on IMO-ProofBench. For comparison, OpenAI tackled the same challenge with an unreleased model, initially claiming 6 solutions but later revised to 5 after one was found flawed, and unlike Aletheia, OpenAI used limited human supervision. Researchers acknowledge Aletheia still makes more errors than human experts and exhibits specification gaming tendencies, meaning full autonomy in math research remains an open challenge.
Sort: