Refactoring working software is risky: changes can break behavior even when they improve code. The new benchmark FrontierRefactor, described in a recent arXiv abstract, targets this problem for autonomous coding agents. It frames refactoring as a task that must strictly preserve externally observable behavior while prioritizing execution performance and compute efficiency over stylistic improvements.
According to the abstract, FrontierRefactor evaluates agents across multiple scales of codebase refactoring. The emphasis is on performance-oriented changes rather than cosmetic ones, which makes the benchmark distinct from earlier refactoring benchmarks that focus on code quality or correctness alone.
Because the source is a single arXiv abstract, the full methodology and results are not yet available. The benchmark's name suggests it tests frontier-level agents, but the abstract does not specify which models were evaluated or how they scored. Further details will require reading the full paper.