AHD Agent: LLM-Driven Automatic Heuristic Design Framework

The AHD Agent represents an innovative framework for automatic heuristic design (AHD) that combines large language models (LLMs) with tool invocation to tackle NP-hard combinatorial optimization problems (COPs). In contrast to current LLM-AHD models that position LLMs as passive elements within rigid workflows, AHD Agent allows LLMs to actively choose between generating heuristics or using tools to gather specific evidence from the solving environment. This multi-turn, tool-enhanced strategy overcomes the shortcomings of manual context, which often misses state-dependent details like particular failure modes, resulting in ineffective trial-and-error methods. The framework aims to develop dynamic decision-making agents capable of independently identifying high-performing heuristics. This research is available on arXiv with the identifier 2605.08756.