Decoded
Feed

Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment

arXiv:2608.23691AIDMMA24 aug 20261 min readHN ×3PDF

Six AI agents run their own math research community and find new results on open problems.

604

points

Kissing configuration size, 11 dimensions

Beats AlphaEvolve's 593-point record on a classical geometry problem

How it works

  1. Pipelines limit discovery

    Systems like AlphaEvolve follow fixed search loops that chase one score. Theorems and explanations need researchers to step in.

  2. Agents run a lab

    GPT, Claude and Gemini agents pick directions, run code, and publish reviewed papers that later agents read and cite.

  3. New math results

    Five of 12 AlphaEvolve problems got new results, all checked by exact constructions or Lean proofs.

submit codewrite upacceptedread, citeLean checkTask andevaluatorSix mixedagentsResearch CenterexperimentsReviewedpapersArchiveliteratureVerifieddiscoveriessubmit codewrite upacceptedread, citeLean checkTask andevaluatorSix mixedagentsResearch CenterexperimentsReviewedpapersArchiveliteratureVerifieddiscoveries
From the paper