Teacher-Aware Evolution of Heuristic Programs from Learned Optimization Policies

Researchers propose a teacher-aware evolutionary framework that leverages pre-trained learned optimization policies to guide the automatic design of heuristic programs for combinatorial optimization problems. The method uses behavioral feedback from teacher policies during evolution rather than relying solely on endpoint performance, achieving better results than baseline LLM-driven approaches without requiring neural inference at deployment.

This research addresses a fundamental challenge in automated algorithm design: how to efficiently discover effective heuristics for hard combinatorial problems. Traditional LLM-based approaches for heuristic generation depend heavily on end-to-end performance metrics, which provide sparse feedback during the search process. The proposed framework reframes this problem by treating independently trained optimization policies as behavioral teachers that provide dense, localized feedback throughout evolution.

The innovation lies in decoupling the inference-time solution from the training-time guidance system. Rather than deploying neural networks alongside heuristics—creating computational overhead and deployment complexity—the method uses teacher policies exclusively during the discovery phase. This approach bridges two disparate research threads: neural combinatorial optimization and automated algorithm design, suggesting that learned models can serve multiple roles beyond their primary function.

The experimental validation across scheduling, routing, and graph optimization benchmarks demonstrates practical value. The framework discovers static, executable heuristics that achieve superior performance compared to performance-only baselines, while maintaining computational efficiency at deployment. This matters for practitioners in operations research, logistics, and network optimization who face strict computational constraints.

The broader implication is methodological: behavioral signals from learned models can effectively guide symbolic program search, opening new avenues for hybrid AI systems. Future work likely explores scaling this approach to larger problems, using diverse teacher architectures, and applying it to domains beyond combinatorial optimization where similar decomposition strategies might apply.

• →Teacher-aware evolution uses learned optimization policies as behavioral feedback sources rather than direct deployment components
• →The method discovers static executable heuristics that outperform performance-only LLM baselines without neural inference overhead
• →Behavioral signals from teacher policies provide dense local feedback that improves search efficiency compared to sparse endpoint metrics
• →Framework demonstrates effectiveness across scheduling, routing, and graph optimization benchmarks
• →Approach opens new possibilities for repurposing learned models as guidance mechanisms in automated algorithm discovery