GP-evolved heuristics guide LLMs to faster, stabler project scheduling

In dynamic multi-mode project scheduling, activities can be executed in multiple modes, with durations uncertain and resources limited. Traditional priority rules provide fast online decisions, but designing them requires deep domain expertise. Genetic programming (GP) hyper-heuristics can automatically evolve these rules, while large language models (LLMs) offer a flexible interface for interpreting scheduling information. However, zero-shot LLMs often lack domain knowledge, consume excessive tokens, and produce inconsistent outputs across repeated queries. A new paper from researchers Yuan Tian, Yi Mei, and Mengjie Zhang flips the typical hybrid approach: instead of using LLMs to guide GP evolution, they extract knowledge from high-quality GP-evolved rules and inject it into an online LLM decision maker. The paper, posted on arXiv (2607.27698), introduces four injection mechanisms: Feature Selection, Feature Hint, Rule Reference, and Rule Follow, each representing a different way of simplifying context or providing explicit decision logic.

The evaluation covers scheduling performance, token consumption, decision stability, and the feature focus in generated rationales. Results show GP-derived guidance generally improves the unguided LLM, but representation matters. Simplifying the decision context (Feature Selection) or supplying explicit decision logic (Rule Follow) proved more effective than merely highlighting important features (Feature Hint). Feature Selection offered the best token efficiency, while Rule Follow achieved strong performance at a greater token cost. Guidance also stabilized decisions and shifted which features appeared in the LLM's explanations. This work suggests that for constrained optimization tasks, combining the search power of GP with the interpretability of LLMs—in the right direction—can make AI schedulers both cheaper and more reliable.