Maestro
A Lean formal proof framework based on a hierarchical coordinator-prover architecture. Sub-provers report the reasons for failure while a global coordinator performs unified scheduling, decoupling global planning from local search.
Timeline (last 90 days)
Tools such as Expo and Maestro are used to let the AI click, take screenshots, and compare functionality on desktop, web, and mobile by itself.
The developer uses tools such as Maestro to have the AI click on the screen and verify actual behavior on desktop, Web and mobile via screenshots.
The research team evaluated Maestro with four frontier LLMs on seven mathematics and three software verification benchmarks.
Using Claude Opus 5, Maestro autonomously discovered proofs for 34 of 48 theorems in the June 2026 version of arXivLean, whereas the native agent managed only 19.
Maestro is a Lean prover framework whose core is a coordinator with a full proof-tree view that can respond based on failure causes with actions such as supplementary guidance, lemma correction, or withdrawing a subtree and replanning.
Maestro's design principle is to make failure causes structurally expressible, shifting the system from blind search to evidence-based scheduling, an idea that can be borrowed for formal verification workflows.
On benchmarks shared with published prover systems, Maestro solved 669/672 problems on PutnamBench, 59/60 on Lean-IMO-Bench, and achieved 6/6 on IMO 2026.
Maestro's coordination mechanism is also applied to verified code synthesis scenarios, where code is sent back for repair when the checker provides a counterexample.
All Facts (8)
Tools such as Expo and Maestro are used to let the AI click, take screenshots, and compare functionality on desktop, web, and mobile by itself.
50%UnverifiedThe developer uses tools such as Maestro to have the AI click on the screen and verify actual behavior on desktop, Web and mobile via screenshots.
50%UnverifiedThe research team evaluated Maestro with four frontier LLMs on seven mathematics and three software verification benchmarks.
50%UnverifiedUsing Claude Opus 5, Maestro autonomously discovered proofs for 34 of 48 theorems in the June 2026 version of arXivLean, whereas the native agent managed only 19.
50%UnverifiedMaestro is a Lean prover framework whose core is a coordinator with a full proof-tree view that can respond based on failure causes with actions such as supplementary guidance, lemma correction, or withdrawing a subtree and replanning.
50%UnverifiedMaestro's design principle is to make failure causes structurally expressible, shifting the system from blind search to evidence-based scheduling, an idea that can be borrowed for formal verification workflows.
50%UnverifiedOn benchmarks shared with published prover systems, Maestro solved 669/672 problems on PutnamBench, 59/60 on Lean-IMO-Bench, and achieved 6/6 on IMO 2026.
50%UnverifiedMaestro's coordination mechanism is also applied to verified code synthesis scenarios, where code is sent back for repair when the checker provides a counterexample.
50%