Carlos
- Updated: March 11, 2026
- 1 min read
HiMAC: Hierarchical Macro-Micro Learning for Long-Horizon LLM Agents
We present an in‑depth overview of the HiMAC framework, a hierarchical macro‑micro learning approach that enables large language model (LLM) agents to tackle long‑horizon tasks with structured planning and reliable execution. HiMAC decomposes decision‑making into a high‑level planner that generates a blueprint and a low‑level executor that carries out goal‑conditioned actions. The paper introduces a critic‑free hierarchical policy optimisation method and an iterative co‑evolution training strategy, achieving state‑of‑the‑art results on ALFWorld, WebShop and Sokoban. For a full technical description, see the original arXiv paper. Learn more about our research and related projects at ubos.tech.