Verified, then logged.
Sun 22 Nov 2026 · 00:20 KST / 15:20 UTC한국어 | ENSearch /

01 / AI · Analysis

AI-written code improves unsolved scheduling benchmark

AI-written code achieved a smaller gap to the theoretical lower bound than commercial tools in a comparison using the same hardware and execution budget.

AI-generated editorial illustration for this article

Researchers reported that an AI-written scheduling solver reached 525,515 using only open-source tools. A solver is a program that searches for solutions to an optimisation problem.

The paper examines whether AI acting as a coding agent can compete with commercial optimisation tools. Its comparison uses the same hardware and execution budget, linking the reported performance to shared testing conditions.

The researchers also reported checking the generated schedules with independent verification programs. The study therefore pairs its performance results with a procedure for checking whether the schedules are valid.

Performance under equal execution budgets

The paper identifies 525,515 as the result for the tai_100_1000_1 scheduling problem. The researchers reported that this value was lower than those achieved by OptalCP, Hexaly and CP Optimizer. These are study-published benchmark results.

The researchers also reported a comparison using equal execution budgets on the same workstation. The reference point is a lower bound: a theoretical limit below which a solution cannot fall. In the study-published results, the AI-written solver's gap above that bound falls within a lower range than those of the comparison tools.

The paper states that the researchers reported every run and used independent checkers to verify every schedule. Its abstract does not disclose the specific time limit used for the equal-budget comparison.

Wider application without code changes

The researchers reported that the solver improved the results across the entire problem family without code changes. They also reported establishing optimal solutions for some problems, extending the findings beyond the initial benchmark.

The paper separately reports improvements from AI-written code on other scheduling problems. These cover flexible job assignment and scheduling involving transport robots. The paper presents them as separate results, rather than attributing all improvements to the solver used in the earlier comparison.

The researchers also reported new lower bounds for the MKT problems, showing that some previously published values cannot be achieved using the distributed problem files. This finding makes checking whether a result is feasible on the same files part of assessing progress, alongside comparing lower scores.