Answers are grounded in this paper and cite their sources.
The Mentorship and Teamwork problem, introduced in the Google Hash Code 2022 competition, is a large-scale combinatorial optimization problem in which contributors with different skill levels are assigned to project roles with the aim of maximizing the score of completed projects. This work investigates whether a large language model (LLM) can improve solvers that are already strong, rather than generate a solver from scratch. We introduce MicroEvolve, an automated LLM-in-the-loop code-tuning process in which the model proposes localized modifications to an existing solver, including heuristic, structural, and parameter changes. Each candidate is compiled, executed on a fixed problem instance, and evaluated by an external validator; a modification is retained only when it preserves or improves the current best score. The loop requires no model training or fine-tuning and proceeds automatically after the initial solver is selected. We evaluate MicroEvolve on the five non-trivial official competition instances, starting from an adapted version of the strongest publicly available open-source solver we identified and driven by a single model, Claude Opus 4.8. Five independent 100-iteration runs are performed for each instance, for 25 runs and 2,500 proposed modifications in total. Every run improves its corresponding baseline. Considering the best run for each instance, gains range from 0.46% to 53.69%, and MicroEvolve establishes new best-known scores on four of the five instances, exceeding previous best-known results by up to approximately 31%. These results provide evidence that LLM-guided code tuning can serve as an effective automated optimization mechanism, analogous to parameter tuning, for improving mature combinatorial optimization solvers.
No takes yet. Share an insight, caveat, or question.
Lajçi et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: