IneqAgent, a novel AI agent, marries large‑language‑model reasoning with symbolic computation to tackle formal inequality proofs in Lean4, achieving a 94% success rate on benchmark problems.
Core Mechanics and Performance
- Employs lemma‑level Monte Carlo Tree Search to orchestrate lemma construction, goal reformulation, and backend selection.
- Integrates multiple symbolic backends—sum‑of‑squares, Bernstein, and successive difference‑substitution—to generate exact certificates for Lean proofs.
- Combines backend outputs with LLM‑generated proof steps, using Lean’s feedback loop to assemble fully verified proofs.
- On four standard benchmarks, DeepSeek‑V4‑Flash powered IneqAgent solves 175 of 186 problems (94.09%), surpassing all evaluated methods.
- Outperforms Direct Agent across all backbones at low inference‑cost thresholds, with lower mean inference costs on successful runs.
- Produces Lean‑verified strengthenings of Dittert’s conjecture in dimensions four and five, and certifies the Holens–Djokovic (4,4) inequality, confirming Merris’s conjecture in dimension four.
From Lean Theory to AI‑Driven Proofs
The project builds on the Lean theorem prover’s growing ecosystem, where formal verification increasingly relies on automated reasoning. By embedding a Monte Carlo tree search within an LLM framework, IneqAgent bridges symbolic mathematics and natural‑language reasoning, a first for inequality proving. Prior attempts largely depended on hand‑crafted tactics or single‑backend approaches, limiting scalability.
Implications for Formal Methods and AI Research
With its high success rate and cost efficiency, IneqAgent sets a new standard for automated theorem proving, potentially accelerating verification in mathematics and software engineering. The framework’s modular design invites integration with other proof assistants and could inspire hybrid AI‑symbolic systems for broader formal reasoning tasks. Future work may extend the agent to higher‑dimensional conjectures and explore real‑time proof synthesis in industrial settings.