AI scientific discovery: OpenAI math & autonomous research
Key Questions
What math problems did OpenAI's reasoning model solve?
OpenAI's reasoning model disproved a 1946 conjecture and solved the Erdős planar unit distance problem using Lean formal proofs and evolutionary search. It also solved 7 out of 10 novel hard math problems, noted as remarkable progress by @emollick.
What is the MaxProof framework used for?
MaxProof scales math proofs with RL and is part of advances in verifiable reasoning. It supports autonomous scientific discovery efforts including DiscoPER's recovery of 8/9 known patterns on an ecological benchmark.
How does the grounded autonomous research pipeline ensure reliability?
The pipeline for computational physics uses redundancy, fresh-context isolation, and adversarial review in a fault-tolerant LLM setup. It provides practical guidance for building reliable autonomous research agents.
OpenAI reasoning model disproves 1946 conjecture and solves Erdős planar unit distance problem using Lean formal proofs and evolutionary search. MaxProof framework scales math proofs with RL. Tweet by @emollick highlights solving 7/10 novel hard math problems is remarkable progress. DiscoPER (Autonomous Scientific Discovery via Iterative Meta-Reflection) recovers 8/9 known patterns on ecological benchmark using second-order meta-reflection. Grounded autonomous research pipeline (fault-tolerant LLM) for computational physics — uses redundancy, fresh-context isolation, adversarial review. Practical for building autonomous research agents. Coding agents can replicate scientific ML papers (dair_ai); concrete signal of agentic research reproduction capability.