Enhancing SLMs for Sustainable Code Optimization in Radio Astronomy

This paper studies small language models (SLMs) for generating, optimizing, and porting radio-astronomy software in the context of the LOFAR telescope upgrade. The upgrade is expected to increase computational requirements by 40 times while keeping energy use from growing. The authors enhance SLM-based code generation with multi-sampling and compiler feedback. They report that multi-sampled SLMs can match or outperform larger models using fewer computational resources, while compiler feedback consistently improves all tested models. The proposed pipeline can also incorporate retrieval-augmented generation, static analysis, and dynamic analysis tools.