DGA2D: LLM framework automates algorithm design, beating baselines by 10.96 points

A team of researchers (Jiale Zhao, Zimu Chen, et al.) from an academic collaboration has released a preprint on arXiv introducing DGA2D (Directed Graph-Guided Automated Algorithm Design), a new framework that leverages Large Language Models (LLMs) to automatically construct heuristic algorithms for NP-hard combinatorial optimization problems (COPs). The core innovation is representing the open-ended program space as a directed graph: each node is a functional operator that can be instantiated by one of several candidate code implementations, and directed walks through the graph correspond to complete algorithmic pipelines.

This structure overcomes three key limitations of existing LLM-driven automated heuristic design (AHD) methods: rigidity to solver templates, low reliability of generated operators, and ineffective credit assignment. DGA2D introduces a first-order path-dependent credit assignment mechanism that evaluates code variations strictly based on their topological context, enabling more effective search. In extensive experiments across 12 diverse COPs—spanning complex scheduling and routing problems—DGA2D demonstrated consistent empirical advantages, reducing the average normalized gap by up to 10.96 percentage points compared to state-of-the-art LLM baselines. The work suggests a path toward fully autonomous, system-level algorithm design with LLMs, though it remains an academic paper with 46 pages and no released code yet.