Teacher-Aware Evolution Improves LLM Heuristic Design
A new arXiv preprint (2605.10634) introduces a teacher-aware evolutionary framework for automatic heuristic design using large language models. The method employs independently trained learned optimization policies as behavioral teachers, querying them on states visited by candidate heuristic programs to obtain local feedback. This approach combines task performance with teacher-derived behavioral signals to guide evolution. Experiments on scheduling, routing, and graph optimization benchmarks demonstrate improvements over performance-driven LLM heuristic evolution baselines, with no neural inference required at deployment.
Key facts
- arXiv preprint 2605.10634 proposes teacher-aware evolutionary framework for LLM-based heuristic design.
- Method uses learned optimization policies as behavioral teachers.
- Teachers are queried on states visited by candidate heuristic programs.
- Local feedback from teachers guides evolution alongside task performance.
- Experiments on scheduling, routing, and graph optimization benchmarks.
- Outperforms performance-driven LLM heuristic evolution baselines.
- Requires no neural inference at deployment.
- Learned optimization policies can be repurposed as behavioral feedback.
Entities
Institutions
- arXiv