Reproduce “Learning Abstractions for Hierarchical Planning in Program-Synthesis 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
Zergham Ahmed, Kazuki Irie, Joshua B. Tenenbaum, Christopher J. Bates, Samuel J. Gershman
TheoryCoder-2, a neuro-symbolic agent, enables AI systems to autonomously learn and reuse abstract concepts for hierarchical planning in program synthesis. This approach significantly improves sample efficiency and generalization across complex tasks, outperforming existing LLM-based methods by integrating large language models with classical symbolic planners.
Key insights
- Autonomous learning and reuse of abstract PDDL operators and grounded Python predicate classifiers dramatically enhance an agent's sample efficiency and generalization capabilities.
- The integration of LLMs for symbolic program synthesis with classical symbolic planners forms an effective neuro-symbolic architecture that leverages both flexible generation and verifiable reasoning.
- Hierarchical planning with learned abstractions facilitates the composition of skills, allowing agents to solve complex, multi-step problems that are challenging for end-to-end or single-level planning methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.