Google DeepMind published a technical report in February 2026, now receiving renewed attention, describing a system built to attempt research-level mathematics. The paper, titled “Towards Autonomous Mathematics Research” (arXiv:2602.10177), was submitted February 10 and last revised March 6. Authors include Demis Hassabis, Quoc V. Le, and more than 25 co-authors affiliated with Google DeepMind.
The system is referred to in coverage as “Aletheia” and described as built on “Gemini 3 Deep Think,” though neither designation was confirmed in the available abstract text. The research distinguishes between competition-level mathematics, solving structured problems with known solutions, like the International Mathematical Olympiad, and research-level mathematics, which involves generating and verifying novel proofs in open territory. That distinction is the paper’s central claim: the system is designed for the harder, less-defined problem.
Google DeepMind’s technical report states the system scored 95.07% on the IMO-ProofBench Advanced benchmark and solved 6 out of 10 novel mathematical lemmas in what is described as a “FirstProof” challenge. These are the vendor’s own reported figures from the vendor’s own paper, no independent third-party evaluation of these claims was confirmed in available materials. Secondary coverage in medium.com and infoq.com reports that the system also addressed 4 open questions from a conjectures database associated with mathematician Paul Erdős, though this could not be independently confirmed from the source materials.
The self-filtering behavior described in secondary coverage is what practitioners should pay most attention to. Rather than generating plausible-sounding but incorrect proofs, the failure mode that makes AI systems dangerous in high-stakes reasoning tasks, coverage describes the system as explicitly outputting “no solution found” when it cannot establish a valid proof. If accurate, this is a design choice with implications far beyond mathematics. AI systems that know what they don’t know and say so are structurally different from systems that hallucinate with confidence.
A note on timing: this paper is from February 2026, not a new April release. If Google DeepMind made a separate public announcement on April 19 that triggered this coverage cycle, that source was not confirmed in available materials. The research itself is two months old. The attention may reflect growing interest in the “research agent” category rather than a new development.
What to watch: whether independent evaluators replicate the benchmark results, and whether other labs (OpenAI, Anthropic, Meta) publish comparable research agent systems. Aletheia, if its self-filtering behavior holds under independent testing, represents a meaningful design direction rather than just a benchmark entry. The category, AI systems attempting novel research rather than executing defined tasks, is new enough that one vendor’s technical report does not establish it. The next 60 days should produce responses from competing labs.