AHD Agent: Agentic Reinforcement Learning for Automatic Heuristic Design

AHD Agent introduces a multi-turn reinforcement learning framework that enables LLMs to proactively query solving environments for state-dependent evidence. Unlike passive LLM-based heuristic generators, this agentic approach uses tool integration to retrieve specific failure modes, reducing trial-and-error exploration when designing heuristics for NP-hard combinatorial optimization problems.