Definition
Plain language
DeepMind's industrial-scale AI system for research mathematics.
As stated in the literature
A DeepMind research-math system used as a frontier baseline on the First Proof benchmark; declines to output when it cannot verify a proof, contrasted with RMA and GPT-5.2R.
Why it matters: In math, an unverifiable confident answer is worse than no answer, so a system that knows when to abstain is more trustworthy than one that always responds.
For example, when faced with a research-math problem it can't verify, Aletheia refuses to output a guess rather than producing a confident-looking but unsound proof.
Heard on the show
“DeepMind's heavy industrial math system, Aletheia, declined to answer at all.”Episode 076 — Same Model, Organized Differently: How an Agent Architecture Beat Frontier Systems at Research Math