EasyProver clears all 189 VERINA tasks with zero fine-tuning

EasyProver arrives as formal software verification has typically demanded either specialized tooling or models trained on the target proof system. The preprint's approach instead layers recursive agents, isolated contexts, asynchronous search, parallel tactic probing and an elastic elaboration service on top of an off-the-shelf frontier model, with no additional training. That combination is what lets it clear established benchmarks like VERINA and NTP4VC without the usual fine-tuning overhead.

The results extend beyond those benchmarks: 669 of 672 PutnamBench problems and all six IMO 2026 formalizations were also proved. Alongside the system, the team released Aegis, built by converting competitive programming problems into Rust verification tasks; its pipeline generated 11,533 candidate specifications validated differentially against Lean, targeting implementation-level obligations such as fixed-width arithmetic, explicit failure, generated loops, casts and functionalized mutation.

On a stratified 2,000-statement sample from Aegis, EasyProver proved 1,054 statements (52.7%) within a 30-minute budget per statement. Because those tasks are drawn from real contest programs rather than constructed examples, the figure is presented as a first baseline that is harder to game than synthetic benchmarks — giving practitioners a zero-training reference system and a stratified Rust benchmark to measure against.