LeanMarathon introduces a multi-agent system designed to solve the challenges of long-horizon mathematical autoformalization by creating reliable, scalable proof blueprints. It demonstrates how durable software harnesses can enable reliable AI co-mathematics.
The paper presents LeanMarathon, a novel multi-agent framework aimed at resolving the failures encountered in long-horizon Lean autoformalization, where statements drift, dependencies tangle, and context decays during large mathematical developments. The system abstracts a single, evolving blueprint (a Lean file) that simultaneously serves as a proof skeleton, a natural-language proof graph, and a shared system record.
Four contract-scoped agents collaboratively construct, audit, prove, and repair this blueprint. They are guided by a two-stage orchestrator that ensures target fidelity through adversarial review before executing parallel, CI-gated proof discharge. This methodology transforms brittle, multi-hour runs into many local, recoverable transactions.
Evaluated across four significant Erdős problems, LeanMarathon successfully formalized all seven target theorems with no 'sorry', proving 258 lemmas and theorems. The research concludes that reliable AI co-mathematics necessitates durable harnesses capable of preserving target fidelity across extended mathematical processes.