Reproduce “Learning and Reusing Policy Decompositions for Hierarchical Generalized Planning with LLM Agents” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shirin Sohrabi, Haritha Ananthakrishnan, Harsha Kokel, Kavitha Srinivas, Michael Katz

IBM researchers developed Hierarchical Component Learning for Generalized Policies (HCL-GP), an architecture that allows LLM agents to dynamically learn, generalize, and reuse policy components from successful executions. This approach improved scenario goal completion for Claude Sonnet 4.6 to 98.2% on normal tasks and enabled less capable models like GPT-OSS 120B to achieve 62.5% on normal tasks, demonstrating enhanced efficiency and cross-domain generalization.

Key insights

  • Dynamically learning and reusing policy components can dramatically improve LLM agent efficiency and accuracy, reducing debugging iterations and inference costs.
  • The approach enables less capable LLMs to effectively solve complex, multi-domain tasks by providing essential structural guidance through reusable components.
  • The HCL-GP framework successfully integrates classical planning principles like generalized policies and hierarchical decomposition into model-free LLM agents.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.