Budget-Efficient Automatic Algorithm Design via Code Graph
Researchers from arXiv have introduced a novel framework for budget-efficient automatic algorithm design (AAD) using Large Language Models (LLMs). Addressing the inefficiencies of existing pipelines that redundantly rewrite full algorithms, this study proposes a directed acyclic graph representation. Instead of generating complete code, the LLM provides compact corrections—such as adding, replacing, or removing code blocks—which augment the graph to create new algorithms. This approach enables correction-level credit assignment, optimizing the search policy to maximize fitness within limited computational costs. Theoretical insights balance search depth and breadth, while empirical validation on three combinatorial optimization problems demonstrates the method's superiority over full-algorithm searches at equal token budgets. Additionally, findings suggest that rich contextual information aids performance only when the LLM's prior knowledge is shallow, potentially hindering results otherwise. This advancement significantly enhances the cost-effectiveness and efficiency of leveraging AI for complex algorithmic problem-solving.
Wire timeline
Budget-Efficient Automatic Algorithm Design via Code Graph
Researchers from arXiv have introduced a novel framework for budget-efficient automatic algorithm design (AAD) using Large Language Models (LLMs). Addressing the inefficiencies of existing pipelines that redundantly rewrite full algorithms, this study proposes a directed acyclic graph representation. Instead of generating complete code, the LLM provides compact corrections—such as adding, replacing, or removing code blocks—which augment the graph to create new algorithms. This approach enables correction-level credit assignment, optimizing the search policy to maximize fitness within limited computational costs. Theoretical insights balance search depth and breadth, while empirical validation on three combinatorial optimization problems demonstrates the method's superiority over full-algorithm searches at equal token budgets. Additionally, findings suggest that rich contextual information aids performance only when the LLM's prior knowledge is shallow, potentially hindering results otherwise. This advancement significantly enhances the cost-effectiveness and efficiency of leveraging AI for complex algorithmic problem-solving.
cs.AI updates on arXiv.org