ARTFEED — Contemporary Art Intelligence

Lean Refactor: Agentic Framework for Multi-Objective Proof Optimization

other · 2026-05-22

Researchers have introduced Lean Refactor, a retrieval-augmented agentic framework designed to refactor Lean proofs with multi-objective control. The framework addresses three key challenges: multi-objective optimization (proof length, compilation cost, version compatibility), fragile version compatibility across Lean/Mathlib releases, and scalability issues with training-based pipelines. Lean Refactor uses a frozen agentic LLM guided by retrievals from a curated database of refactoring strategies annotated with metadata such as supported versions and expected cost reduction. The system is plug-and-play and version-robust, eliminating the need for repeated fine-tuning with new LLM releases. The work is published on arXiv under identifier 2605.20244.

Key facts

  • Lean Refactor is a plug-and-play retrieval-augmented agentic framework
  • It targets multi-objective controllable refactoring of Lean proofs
  • Addresses three challenges: multi-objective optimization, version compatibility, and scalability
  • Uses a frozen agentic LLM with retrievals from a curated strategy database
  • Database contains multi-objective refactoring strategies with metadata
  • Metadata includes supported Lean/Mathlib versions and expected compilation-cost reduction
  • Eliminates need for repeated fine-tuning with new LLM releases
  • Published on arXiv with identifier 2605.20244

Entities

Institutions

  • arXiv

Sources