Proving your program is always correct is far harder than testing it, but testing can miss bugs. Writing these formal proofs by hand is extremely difficult and slow. Current tests for automated provers are too easy or too small.
The paper introduces EasyProver, a system where an AI model repeatedly searches and retries to craft proofs in Lean, a math and code verifier. It also creates Aegis, a pipeline that turns code into hard Lean tasks automatically.
EasyProver solves almost all existing . On new harder Aegis tasks it succeeds more than half the time. This shows scalable proof-based evaluation is feasible, giving AI agents a much stronger, cheaper, and more honest correctness signal than test cases.